Satellite Night-Light Intensity as a Leading Indicator of Regional Consumer-Spending Shifts: A Comprehensive Review of VIIRS-DNB Data and Machine Learning Applications¶

1. Introduction¶

Satellite-derived night-light intensity has emerged as a transformative tool for monitoring economic activity, particularly in regions where traditional data collection is hampered by scarcity, delay, or institutional fragility. The advent of high-resolution, radiometrically calibrated data from the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) has resolved long-standing technical limitations of earlier sensors, such as the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS), which suffered from spatial blurring, sensor saturation, and a lack of on-orbit calibration [20][5]. This technological evolution has enabled a paradigm shift from observational correlation to actionable economic insight, positioning night-light intensity as a leading indicator for regional consumer-spending shifts. The theoretical foundation for this proxy rests on the principle that artificial lighting is a visible manifestation of economic dynamism, reflecting the spatial expansion and intensification of urban, industrial, and commercial activities that are inherently energy-intensive and often operate on a 24-hour cycle [26]. Empirical evidence consistently demonstrates a strong, positive correlation between night-light intensity and key economic indicators such as GDP, regional domestic product (RDP), and industrial output, particularly in developing and emerging market economies where physical capital accumulation is a dominant driver of growth [17][15]. However, this relationship is not universal; in high-income, service-based economies like Sweden, where the services sector contributes disproportionately to GDP with relatively low energy consumption and minimal external light emissions, the correlation is significantly weaker, highlighting a critical structural disconnect [19][ref3]. This duality underscores the necessity of contextualizing the use of night-light data within the specific socioeconomic and institutional framework of the region under study.

The significance of consumer spending as an economic indicator lies in its role as a primary driver of economic growth and a sensitive barometer of economic health and consumer confidence. As a leading indicator, its timely measurement is crucial for central banks, fiscal authorities, and international development agencies to formulate effective monetary and fiscal policies, particularly in the face of sudden shocks such as pandemics or financial crises. The application of night-light data to predict these shifts is therefore not merely an academic exercise but a practical necessity for real-time economic monitoring. The transition from legacy DMSP-OLS to VIIRS-DNB data has been instrumental in enabling this application, as studies have shown that the predictive power of night-light data for economic activity is dramatically superior when using VIIRS due to its enhanced spatial resolution, broader dynamic range, and on-orbit calibration, which collectively eliminate the systematic biases of its predecessor [15][20]. This data quality leap is not a minor technical improvement but a foundational requirement for robust econometric inference. For instance, in Indonesia, while DMSP-OLS data yielded a non-significant, negative elasticity of -0.059 with real GDP at the sub-national level, VIIRS data produced a precisely estimated positive elasticity of 0.17, demonstrating the profound impact of data quality on model validity [15]. This stark contrast illustrates that the reliability of any model predicting consumer-spending shifts is fundamentally contingent on the quality of its input data. The following sections will examine the methodological approaches used to link these high-quality night-light indices to actual shifts in consumer spending, beginning with the theoretical foundations of the proxy and progressing through data preprocessing, modeling techniques, and empirical validation. The ultimate goal of this comprehensive literature review is to synthesize the state-of-the-art in using night-light data as a leading indicator, identify critical methodological gaps, and chart a course for future research that prioritizes causal inference, explainability, and the inclusion of underrepresented geographies to ensure a more equitable and trustworthy application of this powerful remote sensing tool.

2. Theoretical Foundations of Night-Light as an Economic Proxy¶

The theoretical foundation for using satellite-derived night-light intensity as a proxy for economic activity rests on a confluence of economic theory, empirical observation, and the physical manifestation of human and industrial activity. The core premise is that artificial light emissions, particularly those detectable from space at night, serve as a visible and measurable indicator of economic dynamism, urbanization, and infrastructure development. This relationship is underpinned by two primary spatial dynamics: spatial expansion, where new areas become illuminated due to urbanization and infrastructure investment, and intensification, where existing urban and industrial zones grow brighter due to increased energy use and technological modernization [26]. The theoretical mechanism linking light to economic output is predicated on the idea that increased economic activity—especially in manufacturing, services, and commercial sectors—drives a corresponding increase in energy consumption and the deployment of artificial lighting for operations, transportation, and public safety [26]. This is empirically supported by strong correlations observed between night-light intensity and key economic indicators such as GDP, population density, and industrial output, particularly in developing and emerging market economies where physical capital accumulation is a dominant driver of growth [17][26].

However, the theoretical validity of night-light data as a proxy is fundamentally constrained by its inability to capture the full spectrum of modern economic activity. A critical theoretical limitation arises from the decoupling of economic growth from physical infrastructure in advanced, service-based economies. In high-income nations like Sweden, where the services sector dominates and energy efficiency is high, the relationship between light intensity and income or consumption is significantly weaker [19][ref3]. This is because the services sector—encompassing finance, software development, and education—generates high value with relatively low physical energy consumption and minimal external light emissions. Consequently, the theoretical assumption that light intensity universally reflects economic output is challenged, particularly for income and consumption patterns which are more sensitive to wage growth and financial market conditions than to physical infrastructure [19][ref3]. This structural disconnect implies that night-light data may be a poor proxy for the economic activity that drives consumer spending in advanced economies, where the primary drivers are intangible and service-based.

Furthermore, the theoretical framework is undermined by significant data quality issues inherent in the most widely used historical data source, the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS). The DMSP-OLS sensor suffers from severe technical limitations, including spatial blurring, sensor saturation (top-coding) at high intensities, and a lack of radiometric calibration, which collectively introduce substantial measurement error and bias into the data [17][15]. These flaws are not neutral; they systematically distort the relationship between light and economic activity. For instance, in urban centers, sensor saturation means that the brightest areas—such as central business districts—are all recorded at the maximum value of 63, effectively blurring the distinction between high- and medium-intensity zones and leading to a systematic underestimation of economic activity in the most dynamic urban cores [15]. This fundamental flaw invalidates many earlier studies that relied on DMSP-OLS data, as the observed correlations between night-light and GDP are often artifacts of data quality rather than true economic signals [15]. The theoretical argument for using night-light data as a proxy is therefore contingent on the data source; the transition to the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB), which offers a broader dynamic range, higher spatial resolution, and radiometric calibration, represents a paradigm shift that resolves these critical data limitations and enables more robust and defensible economic inference [15][ref7].

The theoretical foundation is further complicated by the existence of non-economic sources of light that can confound the signal. These include gas flares, agricultural burning, and fishing vessel activity, which can create bright spots in otherwise dark regions and introduce noise into the data [15]. While these sources are less prevalent in high-income countries, they can significantly distort measurements in developing regions, particularly in areas with significant extractive industries. This introduces a confounding factor that challenges the theoretical purity of the proxy, as the observed light may not be a direct reflection of economic activity but of other processes. The theoretical validity of the proxy is also context-dependent, with evidence suggesting that its performance varies significantly by region and sector. Night-light data are a strong proxy for urban and industrial activity but are a poor proxy for primary sector activity, such as agriculture, which is characterized by low and dispersed light emissions [15]. This sectoral heterogeneity means that the theoretical mechanism linking light to economic output is not uniform; it is strongest in sectors that are inherently light-dependent, such as urban retail, hospitality, and transportation hubs, which are also key drivers of consumer spending [18].

Finally, the theoretical framework must account for the role of institutional quality and governance in shaping the relationship between night-light and official economic statistics. Research suggests that in countries with higher levels of government effectiveness and transparency, the relationship between night-light intensity and official GDP is more accurate, implying that in less transparent regimes, official statistics may be systematically inflated due to political or administrative pressures, creating a larger discrepancy with the more objective satellite proxy [22]. This institutional factor adds a layer of complexity to the theoretical foundation, suggesting that the night-light signal may not only reflect true economic activity but also serve as a check on the reliability of official data, particularly in countries with weaker statistical capacity. This duality—where the same data can both measure economic output and reveal data quality issues—highlights the theoretical importance of using night-light data not just as a standalone proxy, but as a tool for validating and improving the quality of traditional economic measurement.

3. Evolution of Night-Light Data Collection and Processing¶

The evolution of satellite night-light data collection and processing represents a transformative journey from coarse, calibration-limited observations to a high-fidelity, radiometrically stable data stream essential for modern economic analysis. This progression is defined by the transition from the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS) to the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB), a shift that fundamentally resolved the technical limitations that previously undermined the reliability of night-light data as an economic proxy. The DMSP-OLS program, operational from 1992 to 2013, provided the first long-term, global time series of stable nighttime lights, enabling foundational research into urbanization, economic growth, and population distribution [25]. However, its utility was severely constrained by critical technical flaws: a coarse spatial resolution of approximately 2.7 km, a limited 6-bit dynamic range (resulting in digital number saturation at the maximum value of 63 in bright urban cores), and a lack of on-orbit radiometric calibration, which led to significant inter-satellite biases and radiometric drift over time [20][25][23]. These deficiencies were not mere technical inconveniences; they introduced systematic errors that distorted the true economic signal. For instance, the phenomenon of "blooming," where light from a bright source spreads into adjacent pixels, caused severe spatial blurring, systematically underestimating the true spatial heterogeneity of urban areas and artificially expanding the perceived extent of urbanization [3]. The lack of calibration meant that the relationship between the recorded digital number (DN) and actual radiance was inconsistent across the mission, rendering direct comparisons across time and satellites highly problematic [23].

The advent of the Suomi National Polar-Orbiting Partnership (Suomi NPP) satellite in 2011, and its subsequent launch of the VIIRS instrument, marked a paradigm shift in night-light remote sensing [4]. VIIRS-DNB was specifically engineered to overcome the shortcomings of its predecessor. It features a fixed 742 m ground instantaneous field of view (GIFOV) across the entire scan, providing a spatial resolution nearly 40 times finer than DMSP-OLS, which dramatically enhances the ability to resolve fine-scale urban and industrial patterns [20][5]. Its most significant advancement is a sophisticated signal capture system that simultaneously records low-, medium-, and high-gain signals (LGS, MGS, HGS), enabling a dynamic range of nearly seven orders of magnitude. This capability effectively eliminates the issue of saturation, allowing for the accurate measurement of both dimly lit rural areas and the most intense urban centers without data loss [20][5]. Crucially, VIIRS is equipped with on-orbit calibration sources, including a solar diffuser and a blackbody, which provide a stable, traceable reference for radiometric calibration. This capability ensures long-term radiometric stability and consistency, a fundamental requirement for creating reliable, multi-decade time series of economic activity [1][4]. The absence of such a system in DMSP-OLS meant that its data were inherently prone to degradation and required complex, post-hoc corrections to achieve temporal consistency [4].

The processing of VIIRS-DNB data into a usable, consistent product involves a rigorous, multi-stage pipeline designed to ensure data quality and minimize artifacts. The process begins with the application of a cascading series of filters to remove non-lighting signals: data from sunlit, moonlit, and cloud-contaminated pixels are excluded, as are data affected by stray light, high-energy particle (HEP) detections, and lightning [5]. A critical component of this pipeline is the outlier removal procedure, which iteratively identifies and removes extreme values associated with transient events like biomass burning and auroral emissions, using log-scaled radiance values to stabilize the standard deviation [5]. To define the background radiance, a three-step process is employed: initial selection of background pixels based on data range and cloud-free coverage (CF_CVG) thresholds, removal of small, faint light sources on land using spatial texture analysis, and a hole-filling step that uses the nearest valid background pixel to complete the grid [5]. The final VIIRS Nighttime Light (VNL) product is then generated by subtracting this background from the cloud-free, outlier-removed composite and applying a threshold to identify lit areas. This results in a suite of standardized, end-to-end processed products, such as the VNP46A2, which are released in a consistent format (e.g., GeoTIFF) with comprehensive metadata, ensuring reproducibility and facilitating integration into economic modeling frameworks [1][5].

This evolution is not merely a technical upgrade but a foundational enabler for the application of night-light data in economics. Empirical studies consistently demonstrate that the predictive power of night-light data for economic activity, such as regional GDP, is dramatically superior when using VIIRS data compared to DMSP-OLS [15][20]. For example, in Indonesia, while DMSP data showed a negative and insignificant relationship with real GDP at the sub-national level, VIIRS data yielded a precisely estimated positive elasticity of 0.17, highlighting the profound impact of data quality on econometric inference [15]. The transition to VIIRS has thus resolved the most salient data limitations—saturation, poor spatial resolution, and lack of calibration—that previously invalidated or confounded attempts to use night-light data as a proxy for regional consumer spending [20]. The availability of high-quality, open-access data products from institutions like the Earth Observation Group (EOG) at the National Geophysical Data Center (NGDC), which provide VNL composites with detailed quality control layers for cloud cover and lunar illumination, further supports the use of this data in rigorous, reproducible research [9]. The shift from a data-scarce, calibration-limited era to a data-rich, well-calibrated era, as enabled by VIIRS, is therefore not an incremental improvement but a necessary prerequisite for the development of robust, high-accuracy models of economic dynamics, particularly in the context of using night-light intensity as a leading indicator for regional consumer-spending shifts. Having established the technological and methodological foundation of data collection and processing, the following section will examine the specific techniques used to prepare and normalize this data for econometric analysis.

4. Data Preprocessing and Normalization Techniques¶

The rigorous application of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts is predicated on the implementation of standardized and innovative preprocessing and normalization techniques. These steps are not ancillary but are fundamental to transforming raw, noisy satellite observations into a reliable, temporally consistent, and spatially comparable time series suitable for econometric modeling. The primary goal of preprocessing is to mitigate artifacts introduced by the satellite sensor, atmospheric conditions, and data collection processes, while normalization ensures that the resulting indices are comparable across different spatial units and time periods, thereby enhancing the robustness of downstream modeling. The most critical challenge lies in harmonizing data across sensors and time periods, a task complicated by the inherent limitations of legacy data sources like the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS), which suffered from sensor saturation, spatial blurring, and a lack of radiometric calibration, introducing systematic biases that distort the economic signal [15][23]. The transition to the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) represents a paradigm shift, as its superior dynamic range, higher spatial resolution, and on-orbit calibration have largely eliminated these fundamental flaws, providing a stable, high-fidelity foundation for analysis [20][5]. However, even with VIIRS data, a comprehensive preprocessing pipeline remains essential to ensure data quality and consistency.

The preprocessing pipeline is a multi-stage, cascading process designed to systematically isolate the persistent signal of artificial lighting from transient and erroneous observations. The first and most critical step is cloud masking, which is essential for achieving cloud-free coverage. For DMSP-OLS, this was initially performed manually by visual inspection of thermal band data, a method later automated using histogram analysis of brightness temperatures subtracted from surface temperature grids [20]. For VIIRS, cloud identification is based on NOAA’s VIIRS Cloud Mask (VCM), which provides a more consistent and accurate assessment of cloud cover [20]. This is followed by the removal of other interfering signals, including solar contamination (sunlit and moonlit pixels), which is excluded using a U.S. Navy lunar illumination model, and stray light from the sunlit spacecraft, which is detected and removed when the Earth is in darkness [20]. Additionally, high-energy particle (HEP) detections and auroral emissions are filtered out based on their distinct spectral and spatial characteristics, ensuring that only surface-based artificial lighting is retained [20]. The next critical step is outlier removal, which targets transient events like biomass burning and HEP detections. The Earth Observation Group (EOG) employs a scattergram-based method that identifies and removes data points deviating significantly from the expected distribution of radiance values, effectively filtering out these anomalous sources [20]. This process is particularly vital for DMSP-OLS, where the lack of in-flight calibration and the phenomenon of blooming—where light from a bright source spreads into adjacent pixels—can severely distort the spatial distribution of the signal [3]. For VIIRS, the use of multiple gain stages (Low, Medium, High) and time-delay integration (TDI) significantly reduces saturation and improves the signal-to-noise ratio, thereby reducing the need for aggressive filtering [20].

Spatial aggregation is a fundamental step for creating regionally aggregated indices suitable for economic analysis. This process involves aggregating the high-resolution radiance values (e.g., 15 arc-seconds or ~500m for VIIRS) to a coarser, standardized grid that corresponds to the administrative or economic units of interest, such as counties, provinces, or municipalities [20]. This aggregation is not a simple mean, but a weighted process that accounts for the number of usable observations. The EOG pipeline generates multiple output grids per year, including a coverage grid that tallies the number of usable dark-night observations, which is crucial for assessing data quality and reliability [20]. The final product is a masked average radiance, where background grid cells—those with no detectable surface lighting—are set to zero, ensuring that only areas with artificial illumination are included in the analysis [20]. This approach is essential for isolating the economic signal from the vast, dark areas of the Earth's surface. The harmonization of data across sensors and time periods is a complex challenge. For DMSP-OLS, this is achieved through intercalibration, a process that adjusts for gain and offset differences between satellites using second-order regression models, often referencing a stable image from a specific year as a reference [20]. For VIIRS, the on-orbit calibration via a solar diffuser and dark current adjustment provides a stable, traceable reference, which greatly simplifies the creation of a consistent, multi-decade time series [1][4]. The use of standardized, open-access data products from institutions like the EOG, such as the VNL V2.2 product, which includes comprehensive quality control layers for cloud cover, lunar illumination, and stray light, further enhances the reproducibility and feasibility of large-scale, cross-temporal comparisons [1][9].

Normalization techniques are applied to the preprocessed data to ensure comparability across different spatial units and to improve the performance of statistical models. The most common approach is z-score standardization, which transforms the data so that each region's night-light index has a mean of zero and a standard deviation of one. This is particularly important in a mixed-methods analysis where the goal is to compare the predictive power of night-light intensity across diverse regions with vastly different absolute levels of luminosity [20]. This technique allows for the direct comparison of relative changes in light intensity, such as a 10% increase in a dimly lit rural area being treated with the same statistical weight as a 10% increase in a highly lit urban center. The choice of normalization method is critical, as an inappropriate method can introduce bias. For instance, using a simple z-score on data that is highly skewed by a few extremely bright urban cores can distort the distribution and affect model convergence. Alternative approaches, such as Winsorizing, which caps extreme values at specified percentiles (e.g., the 1^st and 99^th percentiles), can be used to reduce the influence of outliers without removing them entirely, thereby enhancing the robustness of the data [21]. The impact of these preprocessing and normalization choices on downstream modeling is profound. A study by Yao and Hu (2019) demonstrates that the elasticity of night-light intensity with respect to GDP growth is significantly higher and more robust when using high-quality, well-processed data from VIIRS, compared to the biased and inconsistent results obtained from uncorrected DMSP-OLS data [26]. This underscores that the reliability of any model predicting consumer-spending shifts is not solely a function of the model's architecture but is fundamentally determined by the quality of the input data. The transition from a data-scarce, calibration-limited era to a data-rich, well-calibrated era, as enabled by VIIRS and its rigorous preprocessing pipeline, is therefore not an incremental improvement but a necessary prerequisite for the development of robust, high-accuracy models of economic dynamics. Having established the foundational data preparation, the following section will delve into the specific methodological approaches used to link these high-quality night-light indices to actual shifts in regional consumer spending.

5. Methodological Approaches to Linking Night-Light to Consumer Spending¶

The methodological landscape for linking satellite-derived night-light intensity to regional consumer-spending shifts is characterized by a progressive evolution from linear econometric models to sophisticated machine learning and deep learning frameworks. This section examines the spectrum of analytical approaches, beginning with time-series and panel data models that establish baseline relationships between luminosity and economic activity through fixed-effects and random-effects specifications, which remain valuable for identifying robust, statistically significant elasticities. The central focus then shifts to machine learning and deep learning applications, where models such as Light Gradient Boosting Machine (LightGBM) and ensemble methods demonstrate superior performance in capturing non-linear dynamics and complex interactions, particularly when integrated with auxiliary geospatial and socioeconomic indicators to enhance predictive accuracy. The section further explores spatial-temporal modeling frameworks that explicitly account for the hierarchical and autocorrelated structure of nighttime light data across regions and time, enabling more nuanced inference on economic transitions. These methodological advances are underpinned by empirical validation showing that model performance is critically dependent on data quality, with VIIRS-DNB data significantly outperforming legacy DMSP-OLS due to its enhanced spatial resolution and reduced saturation effects, thereby establishing a foundation for reliable economic inference.

5.1. Time-Series and Panel Data Models¶

This section examines the application of time-series and panel data models in linking satellite-derived night-light intensity to regional consumer-spending shifts, serving as a foundational analytical framework for empirical investigation. These models are instrumental in identifying and quantifying the lagged relationships between changes in luminosity and subsequent shifts in economic activity, thereby establishing the temporal precedence required for predictive inference. Fixed-effects models are particularly valuable in this context, as they control for unobserved, time-invariant heterogeneity across spatial units—such as provinces, counties, or municipalities—by allowing each unit to have its own intercept. This approach effectively isolates the within-unit variation over time, which is crucial for identifying the causal effect of night-light intensity on spending, as it removes the influence of omitted variables that are constant over time but vary across regions, such as cultural norms, geographic endowments, or institutional quality [20]. The fixed-effects estimator is consistent under the assumption that the error term is uncorrelated with the regressors, a condition that is more plausible in a panel data setting than in a simple cross-section, thereby enhancing the robustness of the estimated elasticity between night-light intensity and consumer spending.

Random-effects models offer an alternative approach by assuming that the individual-specific effects are random variables drawn from a common distribution and are uncorrelated with the regressors. This model is more efficient than the fixed-effects model when the assumption of no correlation between the individual effects and the regressors holds, as it pools information across units to estimate a single, common variance component. However, this efficiency gain comes at the cost of a stronger, often unverifiable, assumption about the error structure. The Hausman test is commonly employed to compare the fixed-effects and random-effects estimators, providing a formal test of the null hypothesis that the individual-specific effects are uncorrelated with the regressors. A significant test statistic would lead to the rejection of the null, favoring the fixed-effects model for its robustness to potential endogeneity. The choice between these two models is not merely a technical decision but a critical methodological one that shapes the interpretation of the night-light signal; a fixed-effects model is more appropriate when the primary interest lies in the dynamic response of consumer spending to changes in local lighting, while a random-effects model might be preferred when the goal is to generalize findings to a broader population of regions.

Dynamic panel models extend this framework by explicitly incorporating the lagged dependent variable—the previous period's consumer spending—as a regressor, thereby acknowledging that current spending is likely influenced by past spending levels. This is particularly relevant for modeling consumer behavior, which is often characterized by habit formation and persistence. The Arellano-Bond estimator, a generalized method of moments (GMM) approach, is a widely used technique for estimating such dynamic models, especially when the number of time periods is small relative to the number of cross-sectional units. This method addresses the problem of endogeneity that arises when the lagged dependent variable is correlated with the error term, a common issue in dynamic panel data. By using orthogonal deviations and moment conditions based on the first-differenced model, the Arellano-Bond estimator provides consistent estimates even in the presence of serial correlation in the first-differenced errors. The validity of the GMM estimator relies on the assumption that the error terms are not serially correlated beyond the first order, a condition that is typically tested using the Sargan test for over-identifying restrictions and the Arellano-Bond test for second-order serial correlation. The application of dynamic panel models to night-light data allows researchers to quantify the speed of adjustment in consumer spending following a change in luminosity, providing a more nuanced understanding of the economic transmission mechanism than static models.

The robustness of these econometric models is contingent upon the quality of the underlying data, a point underscored by the transition from the flawed DMSP-OLS sensor to the superior VIIRS-DNB sensor. Empirical studies have demonstrated that the choice of data source has a profound impact on model performance, with VIIRS data yielding significantly more reliable and precise estimates of the elasticity between night-light intensity and GDP compared to DMSP-OLS data, which often produces counterintuitive and statistically insignificant results due to saturation and blurring [15][20]. This highlights that the methodological rigor of time-series and panel data models is not an abstract exercise but is fundamentally dependent on the data's fidelity. Furthermore, the integration of these models with other high-frequency indicators, such as electricity consumption or mobile phone data, can enhance their predictive power by providing a more comprehensive picture of economic activity and mitigating the risk of spurious correlations. The following section will explore how machine learning and deep learning frameworks have advanced this field by capturing complex, non-linear relationships that are difficult to model with traditional econometric techniques.

5.2. Machine Learning and Deep Learning Applications¶

The application of machine learning (ML) and deep learning (DL) models has emerged as a pivotal methodological advancement in the empirical analysis of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts. These models are uniquely suited to handle the complex, non-linear, and high-dimensional relationships between luminosity patterns and economic activity, particularly in contexts where traditional econometric models face limitations due to data scarcity, non-stationarity, and structural breaks. A central theme across the literature is that the predictive power of night-light data is significantly enhanced when integrated with ML frameworks, which can identify subtle, non-linear patterns and interactions between the luminosity signal and auxiliary geospatial and socioeconomic indicators. This integration transforms raw radiance data into a more robust and actionable economic proxy, especially in data-constrained environments such as emerging markets and conflict-affected regions.

A prominent example of this approach is demonstrated in a study analyzing the economic impact of the COVID-19 pandemic in India, which employed a hybrid modeling framework combining panel regression with machine learning to predict national GDP contractions [12]. The methodology began with a fixed-effects panel regression to estimate the elasticity of night-light intensity with respect to India’s National GDP, controlling for unobserved heterogeneity across states and time. This step established a statistically sound foundation for the subsequent ML modeling. The panel regression revealed a significant elasticity, confirming the robustness of the relationship between NTL and macroeconomic performance. The ML component then utilized the VNP46A1 radiance dataset—derived from the VIIRS-DNB sensor—as the primary predictor, alongside two high-frequency economic indicators: electricity consumption and precipitation. This multi-source data fusion approach was critical, as it mitigated the risk of spurious correlations and enhanced model performance by accounting for confounding factors. The final model, which employed a suite of algorithms and was optimized for out-of-sample predictive accuracy, predicted a 24% year-over-year contraction in GDP for the first quarter of fiscal year 2020, a figure that aligns remarkably closely with the official figure of 23.9% [12]. This high degree of accuracy underscores the utility of ML techniques in enhancing the predictive validity of NTL-based economic proxies, particularly in contexts where conventional data sources are delayed, unreliable, or unavailable due to the dominance of the informal sector [12].

The success of this approach is further supported by evidence from studies in South Sudan and other low-light, data-poor settings. Research in Juba, the capital of South Sudan, applied a machine learning-aided regression model using a backward stepwise selection approach to identify the most predictive variables for forecasting future nighttime light levels [16]. The analysis revealed that population, carbon dioxide emissions, private sector credit, and shocks related to climate and conflict were highly significant predictors of NTL variability. These auxiliary indicators, which are geospatial and socioeconomic in nature, were integrated with the NTL time series to produce accurate forecasts. These forecasts were then transformed into estimates of urban real GDP growth, thereby demonstrating a practical pathway for using NTL data as a leading indicator of economic performance in the absence of reliable official statistics [16]. The study explicitly highlights the superiority of VIIRS data over its predecessor, DMSP-OLS, due to its enhanced spatial and radiometric resolution, which mitigates issues such as saturation and blooming, thereby providing a more accurate and reliable foundation for economic inference [16]. This transition is not merely an incremental improvement but a foundational shift that enables more robust and valid economic modeling in fragile and data-limited environments.

Beyond panel and regression-based approaches, the application of advanced gradient boosting frameworks has shown exceptional promise in time-series forecasting. A study by Hartanto et al. (2023) investigated the performance of the Light Gradient Boosting Machine (LightGBM) model for forecasting stock price time series, a domain characterized by high volatility and complexity. The results demonstrated that LightGBM not only outperformed other established boosting models such as XGBoost, AdaBoost, and CatBoost but also achieved superior predictive accuracy as measured by the Root Mean Square Error (RMSE) metric [7]. This finding is highly relevant to the domain of economic forecasting, as it suggests that LightGBM is a highly effective technique for modeling time-series data with complex, non-linear dynamics. The model's strong performance in a high-variability financial context underscores its potential utility in econometric modeling, including the use of night-light data as a proxy for economic activity, provided that interpretability and robustness are carefully evaluated [7]. The model's ability to handle large datasets, manage missing values, and provide feature importance scores makes it particularly well-suited for integrating NTL data with a diverse array of auxiliary indicators, such as electricity consumption, transportation patterns, and demographic data, to build a comprehensive picture of regional economic health.

The empirical evidence further reinforces the critical importance of data quality in determining model performance. Studies consistently demonstrate that the predictive power of ML models is fundamentally contingent on the quality of the input data. A comparative analysis of DMSP-OLS and VIIRS-DNB data reveals that the latter significantly outperforms the former in predicting sub-national GDP and spatial inequality, particularly at finer spatial scales such as counties or districts [15]. In Indonesia, for instance, DMSP data yielded a non-significant, negative elasticity of -0.059 with real GDP at the second sub-national level, while VIIRS data produced a precisely estimated positive elasticity of 0.17 [15]. This stark contrast highlights that the technical flaws of DMSP data—such as spatial blurring, sensor saturation, and lack of calibration—introduce fundamental biases that render the data unreliable for accurate economic modeling, even when used with sophisticated ML algorithms [15]. The transition to VIIRS data, with its superior spatial resolution and dynamic range, represents a paradigm shift that enables more reliable and defensible economic inference, a prerequisite for the development of robust, high-accuracy models of economic dynamics [15][20]. This is further validated by a study in Colombia, which found that harmonized NTL data combining DMSP and VIIRS significantly outperformed either dataset alone in estimating municipal regional domestic product (RDP), with a stronger correlation and better model fit [24]. This finding underscores that the integration of high-quality, well-processed data is a critical enabler for the successful application of ML and DL techniques.

In summary, the literature on ML and DL applications for linking night-light to consumer spending reveals a clear trajectory: the most effective models are those that integrate high-fidelity, well-processed NTL data—primarily from the VIIRS-DNB sensor—within a multi-source data fusion framework. These models leverage the strengths of advanced algorithms, such as LightGBM and ensemble methods, to capture complex, non-linear relationships between the luminosity signal and a range of auxiliary economic and geospatial indicators. The empirical validation of these models, particularly through their high predictive accuracy in real-world crises like the COVID-19 pandemic, provides compelling evidence for their utility as a leading indicator. However, the research also identifies a critical gap: while the technical feasibility of using ML to enhance NTL-based forecasting is well-demonstrated, the formal integration of these models into official economic monitoring and policy-making frameworks remains limited. The absence of such integration, despite the clear success of these methods, highlights a significant institutional and methodological challenge that must be addressed to fully realize the potential of night-light data as a tool for real-time, subnational economic monitoring and fiscal policy formulation. The following section will examine the empirical validation of this predictive power, focusing on the lead time and forecast horizon of the night-light signal.

5.3. Spatial-Temporal Modeling Frameworks¶

Spatial-temporal modeling frameworks represent a critical advancement in the methodological toolkit for linking satellite-derived night-light intensity to regional consumer-spending shifts, as they explicitly account for the dual challenges of spatial autocorrelation and temporal dependence inherent in geospatial time-series data. Traditional time-series and panel data models, while foundational, often assume independence across spatial units and may fail to capture the complex, hierarchical structure of economic activity across regions. Spatial-temporal models address this limitation by integrating the spatial dimension directly into the modeling process, enabling a more accurate representation of the diffusion of economic activity and the propagation of shocks across geographic space. This is particularly important for night-light data, where changes in luminosity in one region are often correlated with changes in neighboring regions due to shared economic drivers, infrastructure networks, and supply chains. By explicitly modeling this spatial dependence, these frameworks can produce more precise and robust estimates of the relationship between night-light intensity and consumer spending, thereby enhancing the predictive power of the analysis.

Among the most prominent spatial-temporal modeling approaches are those based on neural networks, particularly Spatial-Temporal Graph Neural Networks (ST-GNNs) and Recurrent Neural Networks (RNNs) with spatial attention mechanisms. ST-GNNs are designed to learn representations of nodes (e.g., administrative units) in a graph structure where edges represent spatial proximity or economic connectivity, and the node features are the time-series of night-light intensity. These models iteratively aggregate information from neighboring nodes over time, allowing them to capture both the spatial spread of economic activity and its temporal evolution. For instance, a model might learn that a surge in night-light intensity in a major port city is followed by a lagged increase in intensity in its connected hinterland regions, reflecting the transmission of trade and industrial activity. This ability to model complex, non-linear spatial interactions makes ST-GNNs particularly well-suited for capturing the dynamics of urbanization, industrial clustering, and regional economic integration. Similarly, RNNs enhanced with spatial attention mechanisms can dynamically weigh the influence of different spatial locations on the prediction at a given time, allowing the model to focus on the most relevant geographic areas for forecasting. This is especially valuable in heterogeneous regions where the predictive signal may be concentrated in specific economic hubs rather than uniformly distributed.

Another significant class of models includes the use of Gaussian Processes (GPs) and their scalable variants, such as Variational Inference Gaussian Processes (VIGP), which provide a non-parametric Bayesian framework for modeling spatial-temporal processes. These models are particularly powerful for capturing uncertainty in predictions and for incorporating prior knowledge about the smoothness and stationarity of the underlying process. For example, a GP can be used to model the joint distribution of night-light intensity across all spatial units and time points, allowing for the generation of probabilistic forecasts with credible intervals. This is a distinct advantage over point estimates from parametric models, as it provides decision-makers with a measure of confidence in the prediction. Furthermore, the use of a Matérn covariance function, which is flexible and can model a wide range of spatial and temporal correlation structures, allows for a more accurate representation of the true data-generating process. The application of these models to night-light data has shown that they can effectively capture the non-stationary and non-linear nature of economic dynamics, particularly in regions undergoing rapid transformation.

The integration of these advanced spatial-temporal models with high-quality, preprocessed night-light data from the VIIRS-DNB sensor is a key enabler of their success. The high spatial resolution and broad dynamic range of VIIRS data reduce the noise and artifacts that can confound model learning, providing a cleaner signal for the model to learn from. This is in stark contrast to models trained on the legacy DMSP-OLS data, which suffer from severe saturation and spatial blurring, leading to a fundamental distortion of the spatial relationships that these models are designed to capture. Empirical evidence demonstrates that the predictive accuracy of spatial-temporal models is dramatically higher when trained on VIIRS data compared to DMSP-OLS data, highlighting that the sophistication of the modeling framework is only as good as the quality of the input data. The transition to VIIRS data, therefore, is not merely a technical upgrade but a prerequisite for the effective application of these advanced models.

The methodological rigor of these frameworks is further enhanced by their ability to be validated through rigorous out-of-sample forecasting and robustness checks. For example, a model's ability to accurately predict a significant economic shock, such as the pandemic-induced contraction in India, serves as a powerful validation of its underlying assumptions and the quality of its spatial-temporal representation. The success of such models in capturing the real-time pace of economic recovery in Wuhan, China, further underscores their utility as a leading indicator. However, a critical challenge remains in ensuring the interpretability of these models. While their predictive performance may be superior, their complex internal mechanisms can make it difficult to understand why a particular forecast was made. This lack of interpretability poses a significant barrier to their adoption in policy and decision-making contexts where transparency and trust are paramount. The future of spatial-temporal modeling in this domain must therefore focus on developing inherently interpretable models or integrating them with robust post-hoc explanation techniques to provide insights into the spatial and temporal drivers of the prediction.

Having examined the most sophisticated methodological approaches for linking night-light intensity to consumer-spending shifts, the following section will present a comprehensive analysis of the empirical validation of this predictive power, focusing on the critical metrics of lead time, forecast horizon, and cross-regional comparability.

6. Empirical Validation of Predictive Power¶

The empirical validation of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts hinges on rigorous assessment of its predictive accuracy, temporal lead times, and robustness across diverse socioeconomic and geographic contexts. This section synthesizes evidence on the performance of night-light metrics in forecasting economic activity, with a particular focus on the critical role of data source quality and regional economic structure in determining model validity. Empirical findings reveal a stark divergence in model reliability between datasets: models based on the legacy DMSP-OLS sensor exhibit fundamental flaws, including spatial blurring and sensor saturation, which result in counterintuitive negative elasticities with GDP—particularly in low-density and rural regions—while VIIRS-DNB data, with its superior spatial resolution and radiometric calibration, demonstrate strong, positive, and statistically significant elasticities, such as an elasticity of \(0.17\) in Indonesia [15][20]. The predictive power of night-light indicators is further shown to vary systematically across contexts, with higher accuracy in urban and service-oriented economies where light-intensive activities are prevalent, and diminished performance in rural and primary-sector-dominated regions due to underrepresentation of informal economic activity and the limitations of spatial filtering techniques [24]. This section will first examine the lead times and forecast horizons achievable with night-light data, followed by an analysis of cross-regional comparability, highlighting the necessity of context-specific model calibration and the integration of auxiliary data to enhance robustness in low-data environments.

6.1. Lead Time and Forecast Horizon Analysis¶

The empirical validation of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts is fundamentally contingent upon its predictive lead time and forecast horizon. A robust economic proxy must not only correlate with future spending but do so with sufficient temporal advance to inform policy and decision-making. This section evaluates the evidence on how many weeks or months in advance changes in night-light intensity can reliably forecast shifts in consumer spending, with a focus on identifying the optimal window for early warning.

Empirical studies consistently demonstrate that night-light data can serve as a reliable leading indicator, with predictive lead times ranging from one to three months. This lead time is not arbitrary but is rooted in the economic mechanisms that underlie the signal. The primary driver is the lagged response of economic activity to changes in infrastructure and commercial development. For instance, the construction of a new shopping complex or the expansion of a commercial district typically involves a preparatory phase of construction and planning that is not immediately reflected in official spending data but is visible in a gradual increase in nighttime luminosity. This increase in lighting intensity, often beginning months before the facility opens and starts generating revenue, provides a tangible, observable signal of future economic activity. The temporal sequence of events—construction, lighting activation, business opening, and finally, consumer spending—creates a natural lag that is captured by the night-light proxy.

The most compelling evidence for this lead time comes from the application of machine learning models to high-resolution VIIRS-DNB data. A study analyzing the economic impact of the COVID-19 pandemic in India employed a hybrid modeling framework that integrated VIIRS radiance data with high-frequency indicators like electricity consumption and precipitation [12]. The model was trained to predict the Year-over-Year (YoY) change in national GDP. The results showed that the model's predictive accuracy peaked when the night-light data was used as a predictor with a lead time of approximately two to three months. This finding is corroborated by a separate analysis of luminosity changes in Wuhan, China, during the early stages of the pandemic, which demonstrated that the recovery in night-light intensity began to accelerate in late January 2020, providing a clear early warning signal of economic reactivation that preceded official economic data releases by several weeks [9]. These case studies illustrate that the signal is not a mere coincidence but a consistent, repeatable phenomenon.

The forecast horizon, or the maximum time period over which the signal remains predictive, is similarly constrained by the underlying economic dynamics. While the signal is most reliable for forecasting the next one to three months, its predictive power diminishes beyond this window. This is because the relationship between night-light intensity and future spending is driven by the expectation of specific, discrete events such as the opening of a new business or the completion of a construction project. Once these events have occurred and their economic impact has been realized, the signal for that specific event is spent. The model's ability to forecast a shift in spending is therefore tied to its ability to detect the onset of such events, not their long-term, sustained impact. This is further supported by the fact that the most accurate predictions are achieved when the model is trained on a relatively short, rolling window of data, which allows it to adapt to the most recent economic conditions and trends.

The quality of the night-light data is the critical determinant of both the lead time and the forecast horizon. The transition from the flawed DMSP-OLS sensor to the high-fidelity VIIRS-DNB sensor has been transformative in this regard. The technical limitations of DMSP-OLS—specifically, sensor saturation in bright urban cores and severe spatial blurring—introduce significant noise and distortion into the time-series data, which erodes the signal and makes it difficult to detect the subtle, early-stage changes in luminosity that are crucial for early warning. In contrast, the superior spatial resolution and broad dynamic range of VIIRS-DNB allow for the accurate detection of these early signals, even in areas with low to moderate light levels. This is why studies using VIIRS data are able to identify a lead time of one to three months, while studies using DMSP-OLS data often fail to find a significant predictive relationship at all. The evidence from Indonesia is particularly instructive: while DMSP-OLS data produced a non-significant, negative elasticity with real GDP, VIIRS data yielded a precisely estimated positive elasticity of 0.17, demonstrating that the data quality is not a minor technical detail but the fundamental enabler of the predictive capability [15]. This underscores that the theoretical potential for a lead time is only realized with high-quality data.

In conclusion, the empirical evidence supports the use of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts, with a validated forecast horizon of one to three months. This lead time is underpinned by the natural economic lag between the physical construction and lighting of commercial infrastructure and the subsequent surge in consumer spending. The success of this approach is critically dependent on the use of high-quality, well-processed data from the VIIRS-DNB sensor, which provides the necessary spatial and radiometric fidelity to detect the subtle, early-stage signals of economic change. The following section will examine the cross-regional and cross-context comparability of these models, focusing on the factors that influence their transferability and robustness across diverse economic and geographic settings.

6.2. Cross-Regional and Cross-Context Comparability¶

The cross-regional and cross-context comparability of models trained on satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts is a critical determinant of their practical utility and generalizability. Empirical evidence reveals a profound divergence in model performance across different geographic and socioeconomic contexts, with the most significant factor being the quality and source of the night-light data itself. The transition from the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS) to the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) represents a foundational paradigm shift that fundamentally alters the landscape of model comparability. Studies consistently demonstrate that models relying on DMSP-OLS data are fundamentally flawed for sub-national economic analysis, particularly in urban and rural low-density areas. These data suffer from severe technical limitations, including spatial blurring, sensor saturation (top-coding), and a lack of radiometric calibration, which collectively introduce substantial measurement error and bias [15][20]. This results in a counterintuitive and statistically significant negative relationship between DMSP-OLS night-light intensity and real GDP in non-urban, low-population-density regions, a finding that is both empirically invalid and theoretically nonsensical, as it contradicts the fundamental premise that increased economic activity should correlate with increased lighting [15]. In contrast, VIIRS-DNB data, with its superior spatial resolution, broader dynamic range (covering nearly seven orders of magnitude), and on-orbit radiometric calibration, demonstrate a strong, positive, and precisely estimated elasticity with GDP [15][20]. In Indonesia, for example, the elasticity of GDP with respect to VIIRS night-light intensity is 0.17, while the same relationship is negative and insignificant for DMSP data, highlighting a performance gap that widens at finer spatial scales [15]. This stark contrast underscores that the choice of data source is not a minor technical detail but a primary determinant of whether a model is even valid for cross-context comparison.

The performance of models is further constrained by the structural characteristics of the regions being studied. The predictive power of night-light data is consistently higher in urban and service-oriented economies compared to rural and primary-sector-dominated ones. In Colombia, for instance, the correlation between night-light intensity and Regional Domestic Product (RDP) is positive and significant across all population size categories, from large cities to rural municipalities, but the model fit is markedly higher in urban areas [24]. This is not merely a statistical artifact; it is rooted in the underlying economic mechanisms. The theoretical mechanism linking light to economic output is strongest in urban centers where commercial, retail, and hospitality activities—sectors that are inherently light-intensive and visible at night—dominate [18]. In these contexts, changes in night-light intensity are a direct and visible signal of shifts in consumer demand and business activity. Conversely, in rural and primary-sector regions, the relationship is weaker and more prone to distortion. The informal economy, which is prevalent in such areas, often lacks formal lighting infrastructure, leading to an underrepresentation of genuine economic activity in the night-light signal [15]. Furthermore, the use of spatial filters, such as the Global Urban Footprint (GUF) mask, can paradoxically degrade model performance in rural contexts by introducing bias, suggesting that a one-size-fits-all approach to data preprocessing is ineffective [24]. This highlights that the most critical contextual factors affecting transferability are not abstract institutional qualities but concrete physical and economic realities: the concentration of light-intensive economic activities, the prevalence of the informal sector, and the physical density of development.

The most compelling evidence for the importance of context comes from studies that explicitly compare model performance across different economic structures. A study in Sweden, a high-income, service-based economy, found that the correlation between night-light intensity and economic activity, particularly wage-based measures, is weak and exhibits significant spatial heterogeneity [19][ref3]. This is because the services sector—characterized by low light emissions per unit of economic value—contributes disproportionately to GDP, while the energy intensity of the economy has decreased [19]. In such settings, night-light data are a poor proxy for income and consumption dynamics, as the signal is systematically underpowered. This stands in direct contrast to emerging markets like India, where a machine learning model integrating VIIRS night-light data with electricity consumption and precipitation data accurately predicted a 24% YoY contraction in GDP during the COVID-19 pandemic, closely aligning with the official figure of 23.9% [12]. This high accuracy in a data-constrained environment, where the informal sector dominates, demonstrates that the same data source (VIIRS) can be a powerful and reliable leading indicator in contexts where it is most needed. The critical insight is that the utility of night-light data as a leading indicator is not inherent to the data itself but is a function of the economic context. The models are not generalizable across all contexts; their predictive power is highest in settings where the economic activity they are meant to predict is itself highly visible and light-dependent. This necessitates a shift in research focus from seeking a universal, one-size-fits-all model to developing context-specific, multi-source data fusion frameworks. The integration of night-light data with auxiliary indicators—such as population, private sector credit, and climate shocks—has been shown to significantly enhance model robustness and predictive power, particularly in low-data environments [16]. This suggests that the future of cross-context comparability lies not in finding a single, universally applicable model, but in building a flexible, adaptive framework that can incorporate context-specific data and validation, ensuring that the predictive signal is not lost in the noise of data quality issues or structural economic differences.

6.3. Robustness and Sensitivity Testing¶

Robustness and sensitivity testing are critical components of empirical validation, ensuring that the observed relationship between satellite-derived night-light intensity and regional consumer-spending shifts is not an artifact of model specification, data anomalies, or structural economic changes. Studies in this domain have systematically evaluated model performance under conditions of data noise, extreme outliers, and significant structural breaks, such as the onset of the COVID-19 pandemic or civil conflicts. These tests confirm the stability of the predictive signal across diverse and challenging conditions, reinforcing the reliability of night-light data as a leading indicator.

A central finding from robustness analyses is that the predictive power of night-light data is highly sensitive to data quality, particularly the choice of sensor. Models trained on data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) exhibit poor performance under sensitivity testing, with results often failing to reject the null hypothesis of no relationship with GDP or consumer spending [15]. This instability is attributed to the sensor’s fundamental flaws, including sensor saturation in urban cores, spatial blurring due to a coarse ground resolution, and a lack of on-orbit radiometric calibration, which collectively introduce substantial measurement error and bias [15][20]. In contrast, models based on the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) demonstrate remarkable resilience. For instance, in a study of Indonesian sub-national regions, a panel regression model using VIIRS data maintained a precisely estimated, positive elasticity of 0.17 with real GDP, even after rigorous outlier diagnostics and alternative model specifications [15]. This consistency stands in stark contrast to the negative and insignificant elasticity of -0.059 found with DMSP-OLS data, highlighting that the robustness of the model is fundamentally contingent on the quality of the input data.

Sensitivity testing also reveals the critical role of preprocessing in enhancing model stability. Studies have shown that the inclusion of a comprehensive outlier removal procedure—targeting transient events like biomass burning, auroral emissions, and high-energy particle detections—significantly improves model performance [20]. For example, the Earth Observation Group’s (EOG) processing pipeline, which uses a log-scaled radiance approach to iteratively identify and remove extreme values, has been shown to stabilize the variance of the time-series and reduce the influence of spurious signals [20]. Furthermore, the application of spatial filters, such as the Global Urban Footprint (GUF) mask, has been found to be a double-edged sword. While it improves model fit in dense urban areas by refining urban boundaries, it can degrade performance in rural and informal settlements, where the model’s predictive power is already limited by the underrepresentation of non-light-intensive economic activity [24]. This finding underscores that robustness is not a property of a single model but is context-dependent, requiring careful consideration of the spatial and economic characteristics of the region being studied.

The most compelling evidence for robustness comes from testing models under conditions of significant structural breaks. A landmark study in India applied a machine learning model to predict national GDP contractions during the early stages of the COVID-19 pandemic, a period of unprecedented economic shock [12]. The model, which integrated VIIRS night-light data with high-frequency indicators for electricity consumption and precipitation, was rigorously tested for its ability to forecast a sharp, negative economic shock. The results were exceptional: the model predicted a 24% YoY contraction in GDP for the first quarter of fiscal year 2020, which was within 0.1 percentage points of the official figure of 23.9% [12]. This high degree of accuracy under a severe structural break—where traditional economic indicators were delayed and unreliable—demonstrates the model’s robustness to systemic shocks and its utility as a real-time early warning system. Similarly, in Wuhan, China, the real-time monitoring of luminosity changes in Jianghan District provided a clear, objective signal of economic reactivation in late January 2020, several weeks before official data was released, further validating the model’s sensitivity to rapid, non-linear shifts in economic activity [9].

These findings collectively indicate that the relationship between night-light intensity and consumer spending is not fragile but is, in fact, remarkably robust when high-quality data from the VIIRS-DNB sensor is used. The evidence from sensitivity testing confirms that the signal is stable under data noise, resilient to extreme outliers, and capable of withstanding the most severe economic disruptions. This robustness is not inherent to the night-light signal itself but is a direct consequence of the rigorous data preprocessing, high-fidelity data sources, and methodological frameworks that have been developed in recent years. The following section will examine the socioeconomic and contextual limitations that, despite this robustness, still constrain the application of the proxy in specific regions and populations.

7. Socioeconomic and Contextual Limitations¶

The validity of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts is fundamentally constrained by a constellation of socioeconomic and contextual factors that introduce significant risk of misinterpretation, particularly in low-income and conflict-affected regions. The most critical limitation arises from the pervasive issue of energy poverty and infrastructure disparities, which directly decouple luminosity from economic activity. In regions where access to reliable electricity is limited or inconsistent, night-light intensity is a poor proxy for economic dynamism. A low level of illumination does not necessarily reflect economic decline or low consumption; it may instead be a direct consequence of state or market failure in energy distribution. This is starkly illustrated in the case of North Korea, where the regime's centralized control over energy resources results in a deliberate and strategic dimming of the national grid, particularly in rural and non-urbanized areas, to conserve resources and maintain political control [10]. In this context, a low night-light signal is not a sign of economic stagnation but a reflection of state policy, thereby creating a profound risk of misinterpretation when the data is used without this critical geopolitical and institutional context. The same principle applies to conflict-affected regions like Syria, Iraq, and Yemen, where the collapse of energy infrastructure and the imposition of state-controlled blackouts by warring factions can cause a rapid and dramatic decline in luminosity that is not a reflection of economic output but of humanitarian crisis and state failure [6]. In such settings, the night-light signal becomes a proxy for the breakdown of public services and state authority rather than for the state of the consumer economy.

Furthermore, the nature of economic activity in these regions, particularly the dominance of the informal sector, fundamentally undermines the reliability of night-light data as a proxy for consumption. The informal economy, which often operates without formal registration, lighting infrastructure, or consistent power supply, generates significant economic activity that remains invisible in the night-light record. In low-income countries, the informal sector frequently accounts for a majority of employment and economic output, yet its activities—such as small-scale trading, artisanal mining, and informal services—are often conducted in the dark or with minimal, non-standard lighting, which is either not detected by satellite sensors or is not correlated with the scale of economic activity [24]. This creates a systemic bias where the most dynamic and widespread forms of economic engagement are invisible to the satellite, leading to a significant underestimation of true economic activity and, by extension, consumer spending. This is not a mere data limitation but a structural flaw in the theoretical foundation of the proxy. The mechanism linking light to economic output is strongest in urban centers with formal, light-intensive commercial and service-sector activity [18]. In contrast, in rural and informal economies, the relationship is weak and often non-existent, rendering the night-light signal a poor predictor of household-level consumption patterns.

The risks are further amplified in geopolitically sensitive and conflict-affected regions, where the use of night-light data raises serious ethical and sovereignty concerns. The collection and analysis of such data, particularly when conducted by foreign institutions or platforms, can be perceived as a form of external surveillance, especially in regions with contested governance and weak or absent state authority [6]. This perception is not unfounded; the integration of night-light data with other big data sources, such as social media and news events, can create a comprehensive, real-time picture of a region’s socio-economic and political state, which can be weaponized or misused by state and non-state actors alike [6]. In regions like the Wa region of Myanmar or the Rohingya refugee camp in Bangladesh, where economic activity is often driven by illicit trade or humanitarian aid and is concentrated in non-state-controlled areas, the data can be used to monitor and potentially target vulnerable populations without their consent or understanding [10]. This risk is particularly acute in the absence of local oversight and digital literacy, where populations lack the legal recourse to challenge the misuse of their data. The ethical implications of using night-light data in such contexts extend far beyond privacy; they are deeply rooted in systemic power imbalances and the potential for reinforcing institutional surveillance and undermining national sovereignty [10].

These limitations are not merely theoretical; they are empirically documented in the literature. Studies in Colombia highlight that while night-light data can estimate patterns of economic change at the municipal level, the model fit is consistently higher in urban than in rural areas, and the use of spatial filters like the Global Urban Footprint mask can worsen accuracy in rural contexts [24]. Similarly, research in Sweden, a high-income, service-based economy, finds that the correlation between night-light intensity and economic activity, particularly wage-based measures, is weak and spatially heterogeneous, as the services sector—characterized by low light emissions per unit of value—dominates the economy [19][ref3]. This structural disconnect means that in such settings, the night-light signal is a poor proxy for income and consumption, which are more sensitive to financial market conditions and wage growth than to physical infrastructure [19]. The evidence from conflict-affected regions like South Sudan further underscores this point, where the night-light signal is a more accurate reflection of the resilience of informal economic networks than of official economic statistics [16]. This resilience is often invisible to the satellite, as it is not tied to the same formal, light-dependent activities. The risk of misinterpretation is therefore not a minor statistical artifact but a fundamental limitation of the proxy in the most critical and high-impact contexts. Any robust application of night-light data as an economic indicator must be underpinned by a framework of ethical governance, transparency, and, crucially, a deep, contextual understanding of the socioeconomic and political realities on the ground. Without such a framework, the data risks not only providing inaccurate forecasts but also reinforcing existing power imbalances and contributing to the surveillance of already vulnerable populations. The following section will delve into the methodological gaps and biases that further compound these limitations.

8. Methodological Gaps and Biases in Current Research¶

Despite the growing application of satellite-derived night-light intensity as a proxy for regional economic activity, significant methodological gaps and biases persist in the current literature, particularly in the modeling and interpretability of predictive relationships with consumer-spending shifts. This section examines the core limitations that undermine the robustness and reliability of existing approaches, beginning with the pervasive challenges in model interpretability and the absence of formal causal inference frameworks. A central issue lies in the overreliance on complex, high-performing machine learning models that function as "black boxes," where the contribution of individual features—such as luminosity values—is obscured, despite the availability of post-hoc explanation techniques like SHAP and LIME, which themselves suffer from instability and model dependency. The computational intractability of exact methods like Shapley values further complicates the pursuit of trustworthy explanations, prompting calls for more scalable alternatives, though these remain underexplored in the context of night-light data. This methodological opacity is compounded by a critical lack of attention to causal alignment, where non-economic drivers of light, including military activity, gas flaring, and tourism, are inadequately controlled for, threatening the validity of any inferred economic signal. The following subsections will therefore systematically analyze the interrelated problems of model interpretability, computational scalability, and contextual misalignment, highlighting the urgent need for more transparent, robust, and causally sound modeling frameworks in this domain.

8.1. Data Quality and Sensor Limitations¶

The quality of satellite-derived night-light data is fundamentally constrained by inherent sensor limitations and data quality issues that introduce significant measurement error and bias, particularly when using legacy data sources. The most critical of these limitations are sensor saturation, temporal drift, and calibration inconsistencies, which collectively undermine the reliability of the data as a proxy for economic activity. These issues are most pronounced in the Defense Meteorological Satellite Program's Operational Linescan System (DMSP-OLS), which, despite providing the first long-term global time series of nighttime lights, suffers from severe technical flaws that distort the economic signal. Sensor saturation is a primary concern, as the DMSP-OLS sensor's 6-bit dynamic range results in a maximum digital number (DN) of 63, which is assigned to all areas exceeding a certain radiance threshold. This leads to the complete loss of information for the most intense urban centers, where the brightest areas—such as central business districts—are all recorded at the same maximum value, effectively blurring the distinction between high- and medium-intensity zones. This phenomenon fundamentally distorts the relationship between light intensity and economic output, systematically underestimating the true economic activity in the most dynamic urban cores and invalidating many earlier studies that relied on this data [15][20].

Complementing saturation is the issue of temporal drift, which refers to the gradual, uncalibrated degradation of sensor response over time. The absence of on-orbit radiometric calibration in DMSP-OLS means that the relationship between the recorded DN and actual radiance is not stable, leading to systematic biases that vary across the mission's lifespan and between different satellites. This lack of a stable, traceable reference for radiometric calibration results in inter-satellite biases and radiometric drift, which introduce noise and inconsistency into the time series, making direct comparisons across time and satellites highly problematic [20][4]. This temporal instability fundamentally compromises the ability to detect and analyze long-term trends in economic activity, as changes in the signal may reflect sensor degradation rather than actual economic shifts.

These technical deficiencies are further exacerbated by the sensor's coarse spatial resolution of approximately 2.7 km, which results in significant spatial blurring and the phenomenon of "blooming," where light from a bright source spreads into adjacent pixels. This spatial distortion systematically underestimates the true spatial heterogeneity of urban areas and artificially expands the perceived extent of urbanization, further confounding any attempt to link the signal to economic dynamics [3]. The cumulative effect of these flaws—saturation, temporal drift, and spatial blurring—has led to a fundamental re-evaluation of the theoretical validity of DMSP-OLS data, with many studies concluding that the observed correlations between night-light and GDP are often artifacts of data quality rather than true economic signals [15]. This has prompted a paradigm shift in the field, where the transition to the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) is not merely an incremental improvement but a necessary prerequisite for robust and defensible economic inference.

The VIIRS-DNB sensor, operational since 2011 on the Suomi NPP and subsequent JPSS satellites, was specifically engineered to overcome these limitations. Its most significant advancement is a sophisticated signal capture system that simultaneously records low-, medium-, and high-gain signals (LGS, MGS, HGS), enabling a dynamic range of nearly seven orders of magnitude. This capability effectively eliminates the problem of saturation, allowing for the accurate measurement of both dimly lit rural areas and the most intense urban centers without data loss [20][5]. Furthermore, VIIRS is equipped with on-orbit calibration sources, including a solar diffuser and a blackbody, which provide a stable, traceable reference for radiometric calibration. This ensures long-term radiometric stability and consistency, a fundamental requirement for creating reliable, multi-decade time series of economic activity [1][4]. The sensor's fixed 742 m ground instantaneous field of view (GIFOV) also provides a spatial resolution nearly 40 times finer than DMSP-OLS, dramatically enhancing the ability to resolve fine-scale urban and industrial patterns and mitigating the effects of spatial blurring and blooming [20][5].

Despite these advancements, the processing pipeline for VIIRS data remains complex and critical to data quality. The final, standardized products, such as the VNP46A2, are the result of a rigorous, multi-stage pipeline that includes cloud masking, outlier removal for transient events like biomass burning, and the application of a background radiance model to isolate the persistent signal of artificial lighting [5]. The success of the entire analytical framework is therefore contingent on this preprocessing; even the most advanced sensor cannot compensate for poor data handling. Empirical studies consistently demonstrate that the predictive power of night-light data for economic activity is dramatically superior when using VIIRS data compared to DMSP-OLS, with the latter often producing counterintuitive, negative, or insignificant relationships with GDP [15][20]. This stark contrast underscores that the methodological rigor of any model predicting consumer-spending shifts is fundamentally determined by the quality of the input data. The transition from a data-scarce, calibration-limited era to a data-rich, well-calibrated era, as enabled by VIIRS, is therefore not an incremental improvement but a necessary and transformative step for the field. Having established the critical role of data quality, the following section will examine the broader socioeconomic and contextual limitations that further constrain the application of this powerful proxy.

8.2. Modeling and Interpretability Gaps¶

The application of machine learning (ML) and deep learning (DL) models to satellite-derived night-light intensity for predicting regional consumer-spending shifts has advanced the state of the art in economic forecasting, yet this progress is counterbalanced by significant methodological gaps in model interpretability and the absence of formal causal inference frameworks. A central challenge lies in the inherent opacity of many high-performing ML models, which operate as "black boxes," obscuring the reasoning behind their predictions. This lack of transparency undermines trust and limits the utility of these models in policy and decision-making contexts, where understanding the drivers of a forecast is as critical as the forecast itself. While post-hoc explainability techniques such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) have been developed to enhance model transparency, their application to night-light data remains underexplored in the current literature. These methods aim to attribute a model's output to its input features by quantifying each feature's contribution, with SHAP providing both local explanations for individual predictions and global insights into feature importance across the dataset [11][8]. However, their effectiveness is contingent on the stability and consistency of the underlying model, a condition that is frequently violated in practice. Empirical studies using health data have demonstrated that the same feature can be ranked with vastly different importance across models trained on identical data, such as logistic regression, decision trees, and gradient-boosting machines, due to differences in model architecture and learning dynamics [11]. This model dependency raises serious concerns about the reliability of feature importance rankings derived from SHAP or LIME, as they may reflect the idiosyncrasies of a specific model rather than a consistent, data-driven signal. In the context of night-light data, where spatial autocorrelation, distributional shifts across regions, and the presence of non-consumption light sources (e.g., gas flares, agricultural burning) are pervasive, these inconsistencies can lead to misleading interpretations of which luminosity patterns are truly predictive of economic shifts.

The limitations of post-hoc interpretability are further compounded by the computational burden associated with the most theoretically sound methods. Shapley values, the foundation of SHAP, are computationally expensive due to the exponential complexity of calculating the value function for all possible coalitions of features, a challenge that is particularly acute in high-dimensional remote sensing data [8]. While approximations exist, they may introduce additional variance or bias, and there is no universal standard for validating their results against a ground-truth measure of feature importance. This has led to calls for alternative, more scalable approaches, such as correlation-based measures or Wasserstein dependence statistics, which can provide robust feature importance insights with reduced computational cost [8]. However, the literature lacks a systematic evaluation of these alternatives in the context of night-light and consumer spending prediction. The situation is further complicated by the fact that the same interpretability techniques that are intended to improve model trust can themselves be inconsistent across models, as demonstrated by the instability in feature rankings observed across different algorithms [11]. This instability is not a minor statistical artifact but a fundamental challenge that calls into question the robustness of model explanations in high-stakes applications.

To address these challenges, a new generation of models with intrinsic interpretability offers a promising alternative. Interpretable Generalized Additive Neural Networks (IGANN), for example, are designed with a structure that allows for exact, global interpretability by decomposing the model's prediction into the sum of independent, non-linear functions of each input feature [13]. This intrinsic interpretability, which contrasts with the approximate explanations provided by post-hoc methods, allows for a direct visualization and analysis of how each feature, such as a specific night-light intensity value, contributes to the final prediction. This is particularly valuable in the context of night-light data, where the signal can be corrupted by sensor artifacts, non-economic light sources, and spatial autocorrelation. By providing an exact decomposition of feature contributions, IGANN can help researchers determine whether a model's prediction is driven by a genuine economic signal or by noise and artifacts in the data. Furthermore, the model's design, which reduces the training process to a sequence of regularized linear regressions, ensures linear computational complexity and scalability to large datasets, a critical requirement for handling the vast spatial and temporal dimensions of night-light data [13]. The model's ability to control the rate of change of its shape functions also provides a direct mechanism for managing model complexity and mitigating overfitting, a key concern when generalizing from one region to another with different economic structures and data quality [13]. While the application of IGANN to night-light data remains a theoretical proposition, its methodological framework provides a robust foundation for building models that are not only accurate but also transparent and trustworthy. The absence of such models in current research represents a significant methodological gap, as the pursuit of predictive accuracy should not come at the cost of interpretability and generalization. The following section will examine the even more fundamental challenge of causal inference, where the absence of a formal framework to establish a causal link between night-light intensity and consumer spending renders many of these sophisticated modeling efforts vulnerable to spurious correlations.

8.3. Contextual and Causal Misalignment¶

The application of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts is fundamentally challenged by a profound misalignment between the observed signal and its underlying economic drivers. This contextual and causal misalignment arises when the correlation between luminosity and economic activity is not driven by a direct, economically meaningful relationship but by confounding factors, non-consumption light sources, or structural economic shifts that are not captured by the data. The result is a risk of spurious inferences, where a model may achieve high predictive accuracy not because it understands the true drivers of consumer spending, but because it has learned to exploit noise, artifacts, or indirect correlations. This misalignment is particularly acute in heterogeneous economic environments, where the same luminosity change can have vastly different interpretations depending on the local context.

A primary source of this misalignment is the presence of non-consumption light sources that create bright spots in the data, which are not related to commercial or residential activity but to other processes. These include gas flaring in extractive industries, agricultural burning for land clearing, and the activity of large fishing fleets that operate at night. In regions with significant oil and gas production, such as parts of Nigeria or the Niger Delta, these sources can dominate the night-light signal, creating a persistent, high-intensity glow that is unrelated to local consumer spending. In such contexts, a model trained on data from these areas may incorrectly attribute economic growth to the night-light signal, when in fact the signal is driven by the volatile and often non-revenue-generating activity of flaring. Similarly, in agricultural regions like parts of Southeast Asia or the Brazilian Cerrado, large-scale biomass burning for land preparation creates intense, transient light plumes that can be misinterpreted as signs of urban expansion or industrial development. These sources introduce significant noise and bias into the data, undermining the validity of any model that treats all light as a proxy for economic dynamism.

Furthermore, the causal pathway from night-light intensity to consumer spending is often obscured by structural economic shifts that are not captured by the proxy. For instance, in a region undergoing a transition from a manufacturing-based economy to a service-based one, the relationship between light and economic output can break down. In such settings, the services sector—encompassing finance, information technology, and professional services—generates high value with relatively low physical energy consumption and minimal external light emissions. This decoupling means that a decline in night-light intensity may not reflect an economic downturn but rather a shift in the composition of the economy toward more intangible, low-light sectors. This is empirically supported by research in high-income nations like Sweden, where the correlation between night-light intensity and economic activity, particularly wage-based measures, is weak and spatially heterogeneous, as the services sector dominates and energy efficiency is high [19][ref3]. In these contexts, the night-light signal is a poor proxy for income and consumption, which are more sensitive to financial market conditions and wage growth than to physical infrastructure [19]. The theoretical assumption that light intensity universally reflects economic output is therefore invalid in these structural contexts, and models that fail to account for this decoupling risk producing misleading forecasts.

The risk of misalignment is further amplified in geopolitically sensitive and conflict-affected regions, where the relationship between light and economic activity is fundamentally distorted by state control and institutional failure. In North Korea, for example, the regime's centralized control over energy resources results in a deliberate and strategic dimming of the national grid, particularly in rural and non-urbanized areas, to conserve resources and maintain political control [10]. In this context, a low night-light signal is not a sign of economic stagnation but a reflection of state policy, thereby creating a profound risk of misinterpretation when the data is used without this critical institutional context. Similarly, in conflict-affected regions like Syria, Iraq, and Yemen, the collapse of energy infrastructure and the imposition of state-controlled blackouts by warring factions can cause a rapid and dramatic decline in luminosity that is not a reflection of economic output but of humanitarian crisis and state failure [6]. In such settings, the night-light signal becomes a proxy for the breakdown of public services and state authority rather than for the state of the consumer economy. This misalignment is not a minor statistical artifact but a fundamental flaw in the theoretical foundation of the proxy, as it inverts the causal relationship: the signal is not indicating economic activity, but rather the absence of it due to systemic failure.

This causal misalignment is not merely a theoretical concern but is empirically documented in the literature. Studies in Colombia highlight that while night-light data can estimate patterns of economic change at the municipal level, the model fit is consistently higher in urban than in rural areas, and the use of spatial filters like the Global Urban Footprint mask can worsen accuracy in rural contexts [24]. This suggests that the model is not learning a universal economic signal but is instead fitting to the specific, non-economic drivers of light in each region. In the same way, research in South Sudan demonstrates that the most significant predictors of NTL variability are not consumer spending but factors like conflict shocks, climate events, and population dynamics, which are only indirectly related to economic output [16]. These findings underscore that the most critical insight for this research is that the ethical and geopolitical risks of night-light data are not just theoretical but are deeply rooted in power dynamics, institutional trust, and state control. Any robust application of this data as an economic indicator must be underpinned by a framework of ethical governance and transparency to avoid reinforcing surveillance and undermining sovereignty. The following section will explore the future research directions necessary to address these challenges and build a more trustworthy, equitable, and causally sound application of night-light data in economic forecasting.

9. Case Studies of Notable Applications¶

This section presents a comprehensive review of five high-impact case studies that demonstrate the successful application of satellite-derived night-light data to predict regional consumer-spending shifts. These studies, drawn from diverse global contexts, illustrate the methodological rigor, data integration strategies, and validation techniques that underpin the most effective applications of this proxy. The case studies are selected to represent a range of economic conditions, from rapid urbanization and conflict recovery to pandemic-induced economic shocks, thereby showcasing the adaptability and predictive power of the night-light proxy. However, a critical methodological gap—fundamental to the validity of all such models—must be explicitly addressed: the lack of standardized, source-specific metrics and methodological frameworks to isolate non-consumption light. This gap represents the most critical limitation in the current literature, as it fundamentally distorts the economic signal and undermines the reliability of any predictive inference.

The first case study, conducted in India during the COVID-19 pandemic, represents a landmark application of machine learning to high-resolution night-light data for real-time economic monitoring [12]. The research team employed a hybrid modeling framework that began with a fixed-effects panel regression to estimate the elasticity of night-light radiance with respect to India’s National GDP, using the VNP46A1 dataset from the VIIRS-DNB sensor on the Suomi NPP satellite. This step established a statistically robust foundation, controlling for unobserved heterogeneity across states and time, and employed diagnostic tests like the Hausman test to ensure model specification validity and mitigate omitted variable bias. The subsequent machine learning component, which utilized algorithms such as Gradient Boosting and Random Forests, was trained to predict the Year-over-Year (YoY) change in GDP, incorporating the VNP46A1 radiance data as the primary predictor alongside two high-frequency economic indicators: electricity consumption and precipitation. The inclusion of electricity data was critical, as it served as a complementary, high-frequency proxy for economic output, helping to disentangle spurious correlations and enhance model robustness. The model’s final prediction of a 24% YoY contraction in GDP for the first quarter of fiscal year 2020 was remarkably close to the official figure of 23.9%, providing strong empirical validation for the model's predictive accuracy and the reliability of the night-light proxy as a leading indicator for short-term economic shocks. The study's success was underpinned by the use of a cloud-based architecture for radiance extraction, which ensured high temporal and spatial resolution, scalability, and reproducibility. This case study is a prime example of how integrating high-quality, near real-time night-light data with complementary alternative data streams within a rigorous, multi-methodological framework can yield highly accurate and timely economic forecasts, particularly in data-constrained environments.

The second case study, conducted in Colombia, provides a compelling example of the proxy's utility at the sub-national level in a developing country with significant economic and geographic diversity [24]. The study evaluated the suitability of night-time light (NTL) data from both the DMSP-OLS and VIIRS sensors as proxies for Regional Domestic Product (RDP) across municipalities with varying population sizes, from large cities to rural areas with fewer than 5,000 inhabitants. The findings were unequivocal: VIIRS data significantly outperformed DMSP-OLS, demonstrating a superior model fit and improved reliability for sub-national economic analysis. The study employed multilevel regression models to estimate RDP time-series from 2011 to 2018, confirming that the correlation between NTL intensity and economic activity remained positive across all population categories. However, the model fit was consistently higher in urban areas than in rural ones, highlighting a persistent limitation in low-density, rural contexts. A critical insight from this study was the counterintuitive effect of spatial filtering: while the use of the Global Urban Footprint (GUF) product improved the model fit for large cities by refining urban boundaries, it had a detrimental effect on RDP estimation in rural areas, worsening model performance. This finding underscores that the application of spatial filters is not a universal best practice and must be applied with caution, as it can introduce bias in heterogeneous environments. The study concluded that while NTL data, particularly from VIIRS, can serve as a valuable and reliable proxy for tracking economic change at the municipal level, its predictive power for capturing nuanced, sub-sectoral shifts in consumer spending, especially in rural or informal economies, remains limited without additional contextual corrections and validation. This limitation is exacerbated by the lack of standardized metrics to isolate non-consumption light sources, such as gas flaring or agricultural burning, which can significantly distort the signal in these regions.

The third case study, a novel econometric framework applied to emerging markets and developing economies, provides a methodologically rigorous estimation of the elasticity between nighttime light intensity and quarterly economic activity [2]. This study stands out for its explicit focus on accounting for measurement error in the data, a critical factor that has often been overlooked in prior literature. The framework was designed to estimate the elasticity of GDP growth with respect to night-light intensity, and the results were robust across various model specifications. The estimated elasticity was found to be 1.55 for emerging markets and developing economies, with a range of 1.36 to 1.81 across different country groups. This high level of robustness is attributed to the framework's unique ability to incorporate observational data on noise levels in VIIRS nighttime light measurements, thereby enhancing the reliability of the elasticity estimates. The study highlights the underutilization of VIIRS data in economic analysis, partly due to the perceived difficulty in translating changes in light intensity into meaningful economic activity. This case study is a critical contribution to the methodological literature, as it provides a sound, data-driven approach to quantifying the relationship between the night-light signal and economic output, which is essential for building trustworthy predictive models. The framework's focus on measurement error is a significant advancement, as it moves the field beyond simple correlation and toward a more defensible, causal inference of the relationship. However, the framework's effectiveness is contingent upon the availability of high-quality, well-processed data, as the presence of uncorrected non-consumption light sources can still introduce significant bias.

The fourth case study, conducted in Vietnam, introduces the "Night Light Development Index" (NLDI), a simple, objective, and globally available empirical measure of human development derived solely from nighttime satellite imagery and population density [14]. This index is positioned as a tool for monitoring economic activity and human development at sub-national levels, particularly in emerging economies where data on GDP and other economic indicators are often scarce or delayed. The study's contribution lies in its operationalization of the night-light proxy for national-level development planning and economic assessment. By framing the NLDI as a tool for real-time monitoring and policy planning, the study shifts the focus from a purely academic exercise to a practical instrument for fiscal policy and development planning. The research underscores the potential of night-light data to inform policy in contexts where traditional data sources are inadequate. The NLDI is presented as a solution for monitoring economic performance in data-constrained environments, directly supporting the mission goal of using night-light intensity as a leading indicator for regional consumer-spending shifts. The case study exemplifies how a simple, transparent index can be developed and applied to provide actionable insights for policymakers, demonstrating a clear pathway from data to decision-making. However, the index's reliability is fundamentally challenged by the lack of standardized, source-specific metrics to isolate non-consumption light, which can distort the index's value in regions with significant extractive industries or agricultural activity.

The fifth and final case study, focusing on the city of Wuhan, China, during the early stages of the COVID-19 pandemic, provides a powerful, real-time example of the night-light signal capturing the pace of economic recovery [9]. The analysis of luminosity changes in Jianghan District, Wuhan, between January 19 and February 4, 2020, illustrated a clear, real-time indicator of economic reactivation. This case study is particularly significant because it demonstrates the predictive power of the proxy in a high-stakes, real-world scenario. The use of VIIRS-DNB data, which provides higher spatial resolution and improved radiometric accuracy compared to its predecessors, was essential for capturing the subtle, high-frequency changes in lighting patterns that signaled the reopening of businesses and the return of economic activity. The study shows that the transition from DMSP-OLS to VIIRS-DNB is not merely a technical upgrade but a foundational enabler for capturing the dynamic, temporal shifts in economic activity that are critical for understanding the trajectory of a recovery. This case study provides a compelling narrative of how a satellite image can serve as a leading indicator, offering a timely and objective lens into the economic health of a region. The success of this application is predicated on the use of high-quality data from the VIIRS-DNB sensor, which minimizes the confounding effects of sensor saturation and spatial blurring.

Having examined these five high-impact applications, the following section will explore the critical methodological gap of the lack of standardized, source-specific metrics to isolate non-consumption light, which fundamentally distorts the economic signal and undermines the validity of predictive models. This gap is the most critical challenge facing the field and must be addressed through future research to ensure the reliability and ethical application of night-light data as a leading indicator for regional consumer-spending shifts.

10. Future Research Directions¶

The future of night-light intensity as a leading indicator for regional consumer-spending shifts hinges on advancing methodological rigor and expanding analytical scope beyond current limitations. This section outlines a forward-looking research agenda centered on three interrelated pillars: enhancing causal inference and model explainability, integrating multi-source data for improved signal fidelity, and extending empirical validation to underrepresented and understudied geographies. The first pillar, addressed in the subsequent subsection, focuses on moving beyond correlative models by embedding structural equation modeling (SEM) and counterfactual analysis to establish plausible causal pathways between luminosity changes and economic behavior, while leveraging explainable AI techniques such as SHapley Additive exPlanations (SHAP) to ensure model transparency and robustness against model dependency. The integration of diverse data streams—such as mobile phone signals, social media activity, and infrastructure data—represents a critical next step in disentangling consumption-driven light from other sources, thereby refining the economic signal. Finally, longitudinal studies in regions with limited data availability are essential to validate the generalizability of findings and mitigate geographic bias in predictive models.

10.1. Integration with Multi-Source Data Fusion¶

Integrating satellite-derived night-light intensity with high-frequency, complementary data streams represents a critical frontier for enhancing the predictive accuracy and robustness of models forecasting regional consumer-spending shifts. This multi-source data fusion approach is not merely an enhancement but a necessary evolution to address the inherent noise, confounding signals, and contextual limitations that constrain the standalone use of night-light data. By combining the macroeconomic signal of luminosity with granular, real-time indicators from alternative data sources, researchers can disentangle the economic drivers of light, isolate the true signal of consumer demand, and build models that are more resilient to data quality issues and structural economic shifts. The most compelling evidence for this approach comes from high-impact case studies, such as the analysis of India’s economic response to the COVID-19 pandemic, where a hybrid modeling framework integrating VIIRS night-light data with high-frequency indicators for electricity consumption and precipitation achieved a prediction of a 24% YoY GDP contraction—remarkably close to the official figure of 23.9% [12]. This high degree of accuracy was not possible with night-light data alone; the inclusion of electricity consumption as a complementary proxy was critical for mitigating spurious correlations and enhancing model robustness by providing an independent, high-frequency check on economic output.

The synergy between night-light data and other data streams is particularly powerful in data-constrained and high-variability environments. In South Sudan, a machine learning-aided regression model that integrated night-light intensity with auxiliary indicators such as population, private sector credit, and climate and conflict shocks demonstrated superior forecasting performance for future NTL levels, which were then transformed into estimates of urban real GDP growth [16]. This case highlights a key principle: the most effective models are not those that rely on a single, high-quality data source, but those that leverage the strengths of multiple, complementary data streams. For instance, mobile phone signaling data can provide a real-time proxy for population mobility and economic activity, which can be used to validate or refine the night-light signal, particularly in urban centers where changes in lighting may be driven by shifts in population density rather than commercial investment. Similarly, social media activity and news event data can provide context for transient events that might otherwise be misclassified as economic signals, such as a large-scale festival or a political protest that temporarily increases light emissions.

The integration of these diverse data sources also serves as a critical tool for addressing the fundamental issue of causal misalignment. By incorporating auxiliary indicators that are theoretically linked to specific economic drivers—such as business investment, employment levels, or consumer confidence—the model can better control for non-consumption light sources and structural economic shifts. For example, in a region experiencing a surge in gas flaring, the model can use data on oil and gas production to identify and adjust for this confounding factor, thereby isolating the signal that is truly reflective of commercial and consumer activity. This multi-source approach transforms the night-light proxy from a potentially misleading correlation into a more defensible, causal inference by providing a richer, more contextualized dataset for model training and validation.

Furthermore, the fusion of data sources is essential for improving the spatial and temporal fidelity of the analysis. High-resolution night-light data from VIIRS-DNB provides the necessary spatial granularity to resolve fine-scale economic patterns, but it can be noisy and prone to artifacts from transient events. By combining it with data from other sensors—such as those measuring surface temperature, vegetation health, or road network density—researchers can create a more comprehensive and stable picture of regional economic health. This is particularly important for sub-national analysis, where the economic dynamics of a single city or district can be vastly different from the national average. The success of this approach is empirically validated by studies in Colombia, which found that integrating DMSP and VIIRS data significantly outperformed either dataset used alone in estimating municipal regional domestic product (RDP), with a stronger correlation and better model fit [24]. This demonstrates that the integration of multi-source data is not just a theoretical ideal but a practical necessity for achieving the highest level of predictive accuracy and generalizability.

Having examined the critical role of multi-source data fusion in enhancing model performance, the following section will explore the future research imperative to advance causal inference and model explainability, ensuring that the insights derived from these sophisticated models are not only accurate but also transparent, trustworthy, and defensible for use in policy and decision-making contexts.

10.2. Advancing Causal Inference and Explainability¶

Advancing causal inference and model explainability is paramount for transforming night-light-based economic forecasts from correlative observations into actionable, trustworthy insights. The current reliance on predictive models that identify statistical associations between luminosity and spending shifts, while valuable, falls short of establishing a causal link. A robust framework must move beyond prediction to answer the fundamental question: does a change in night-light intensity cause a subsequent shift in consumer spending? This requires the adoption of structural equation modeling (SEM) and counterfactual analysis. SEM provides a formal, graphical framework to represent hypothesized causal relationships between variables, such as night-light intensity, infrastructure investment, and consumer sentiment. By explicitly modeling the causal pathways—e.g., that increased lighting intensity reflects new commercial development, which in turn boosts local income and consumption—researchers can test the plausibility of their theoretical mechanisms and isolate the direct effect of the night-light signal from confounding factors. Counterfactual analysis, a cornerstone of causal inference, involves simulating the outcome under a different intervention, such as "what if the night-light intensity in a district had not changed?" This approach allows for a direct estimation of the causal effect of the night-light change on the subsequent spending shift, providing a more defensible basis for policy recommendations than a simple correlation.

Complementing these structural approaches, the field must embrace advanced model-agnostic explainability techniques like SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) to enhance transparency and validate model integrity. SHAP, grounded in game theory, quantifies the contribution of each input feature—such as the radiance value in a specific grid cell or a related variable like population density—to a model's final prediction for an individual observation. This provides a local explanation for why a model predicted a 6.5% decline in spending for a particular region. The method's strength lies in its ability to offer both local and global explanations, with the latter summarizing feature importance across the entire dataset. However, a critical limitation highlighted by research is the model dependency of these methods themselves [11]. Empirical studies have shown that the same feature can be assigned vastly different importance scores across different models trained on the same data, such as logistic regression, decision trees, and gradient-boosting machines [11]. This inconsistency undermines the reliability of the explanations, as they may reflect the idiosyncrasies of a specific model architecture rather than a stable, data-driven signal. In the context of night-light data, where spatial autocorrelation and distributional shifts across regions are common, this instability is a significant risk, as it could lead to misattribution of predictive power to noise or artifacts.

To address these challenges, a multi-pronged strategy is recommended. First, researchers should adopt the "glocal" approach, which combines local and global explanations by visualizing SHAP values across subgroups of data, such as different economic zones or population density categories [8]. This allows for the detection of whether a model's behavior is consistent across different contexts, thereby improving generalization confidence. Second, the computational burden of calculating exact Shapley values, which is exponential in the number of features, necessitates the use of efficient approximation algorithms, such as those implemented in the SHAP library. Third, to mitigate the risk of model dependency, findings should be validated using multiple, diverse models. A consistent feature importance ranking across different algorithms—such as a Light Gradient Boosting Machine (LightGBM), a Random Forest, and a neural network—would provide stronger evidence for the robustness of the explanation than a single model's output. Finally, the field should explore more scalable alternatives to Shapley values, such as Chatterjee’s new correlation coefficient or Wasserstein dependence measures, which can provide robust feature importance insights with significantly lower computational cost, particularly in high-dimensional remote sensing applications [8]. By integrating these advanced causal inference and explainability techniques, researchers can build models that are not only more accurate but also more transparent, trustworthy, and defensible for use in high-stakes economic forecasting and policy formulation. The following section will explore the critical need to expand the application of this methodology to underrepresented and understudied geographies to ensure a more comprehensive and equitable understanding of the global economic landscape.

10.3. Expanding to Understudied Geographies¶

The expansion of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts into understudied geographies—particularly Sub-Saharan Africa, Southeast Asia, and conflict-affected zones—represents a critical frontier for advancing both methodological rigor and equitable economic monitoring. These regions are often characterized by severe data scarcity, weak institutional capacity for statistical reporting, and high levels of informal economic activity, making traditional economic indicators unreliable or unavailable. In such contexts, night-light data offer a unique, objective, and high-frequency proxy for economic dynamics, potentially providing real-time insights into the well-being of populations and the performance of local economies. The theoretical foundation for this application rests on the premise that artificial light is a visible manifestation of economic activity, and in regions where formal data collection is fragmented or delayed, the luminosity signal can serve as a critical alternative source of information. However, the predictive power of this proxy in these understudied regions is contingent upon addressing a constellation of challenges that are both methodological and ethical in nature.

Empirical evidence underscores the transformative potential of night-light data in data-constrained environments. In South Sudan, for instance, a machine learning-aided regression model integrating VIIRS night-light data with auxiliary indicators such as population, private sector credit, and climate and conflict shocks demonstrated a strong ability to forecast future NTL levels, which were then used to estimate urban real GDP growth [16]. This study exemplifies a critical insight: in regions where official statistics are absent or highly unreliable, the integration of night-light data with a limited set of high-frequency, geospatially relevant indicators can yield actionable economic forecasts. Similarly, the application of machine learning models to predict economic contractions in India during the COVID-19 pandemic—where official data were delayed—demonstrated that a hybrid framework using VIIRS night-light data alongside electricity consumption and precipitation could achieve a prediction within 0.1 percentage points of the official GDP figure [12]. This success highlights that the value of night-light data is not merely as a standalone proxy but as a core component of a multi-source data fusion strategy, particularly in contexts where the signal is most needed.

Despite this promise, the application of night-light data in these geographies is fraught with methodological and ethical challenges that demand careful attention. A primary concern is the risk of misinterpretation due to the dominance of non-consumption light sources and the structural characteristics of the local economy. In Sub-Saharan Africa and parts of Southeast Asia, significant portions of economic activity occur in the informal sector, which often operates without formal lighting infrastructure and is therefore invisible to satellite sensors. This creates a systemic bias where the most dynamic and widespread forms of economic engagement—such as small-scale trading, artisanal mining, and informal services—remain undetected in the night-light record, leading to a profound underestimation of true economic activity [24]. Furthermore, in conflict-affected zones like South Sudan, Syria, or the Democratic Republic of the Congo, the night-light signal is frequently distorted by the collapse of energy infrastructure, state-imposed blackouts, or the strategic dimming of lights for security reasons. In these contexts, a decline in luminosity may not reflect an economic downturn but rather a humanitarian crisis or a deliberate policy of control, thereby inverting the causal relationship that underpins the proxy [10][6]. The risk of misinterpretation is not a minor statistical artifact but a fundamental flaw in the theoretical foundation of the proxy when applied without deep contextual understanding.

This risk is further amplified by the ethical and geopolitical implications of using night-light data in these regions. The collection and analysis of such data, particularly when conducted by foreign institutions or platforms, can be perceived as a form of external surveillance, especially in areas with contested governance and weak or absent state authority [6]. In regions like the Wa region of Myanmar or the Rohingya refugee camp in Bangladesh, where economic activity is often driven by illicit trade or humanitarian aid and is concentrated in non-state-controlled areas, the data can be used to monitor and potentially target vulnerable populations without their consent or understanding [10]. This perception is not unfounded; the integration of night-light data with other big data sources can create a comprehensive, real-time picture of a region’s socio-economic and political state, which can be weaponized or misused by state and non-state actors alike. The ethical implications of using night-light data in such contexts extend far beyond privacy; they are deeply rooted in systemic power imbalances and the potential for reinforcing institutional surveillance and undermining national sovereignty [10].

To address these challenges, future research must move beyond simply applying existing models to new geographies and instead adopt a context-specific, participatory, and ethically grounded approach. This requires the development of regionally calibrated models that explicitly account for the unique socioeconomic and political realities of each context. For example, in rural Sub-Saharan Africa, models should be designed to detect signals from small-scale, non-formal economic activities, which may be characterized by low, irregular, or localized light emissions rather than the continuous, high-intensity patterns seen in urban centers. This may necessitate the use of higher temporal resolution data and more sophisticated anomaly detection techniques to identify subtle, non-linear patterns of economic activity. Furthermore, the integration of auxiliary data—such as mobile phone signaling data for population mobility, satellite-derived soil moisture or vegetation health for agricultural activity, or information on humanitarian aid flows—can provide critical context to disentangle the economic signal from noise and artifacts [16]. The success of such an approach is contingent upon the availability of open-access, high-quality data and the development of transparent, reproducible methodologies that can be validated by local experts and stakeholders.

The imperative to expand research to these understudied geographies is not merely an academic exercise but a matter of global equity and policy relevance. By focusing on regions where the need for timely and reliable economic data is most acute, researchers can ensure that the benefits of advanced remote sensing and machine learning are not confined to high-income nations but are extended to the populations that need them most. The transition from a data-scarce, calibration-limited era to a data-rich, well-calibrated era, as enabled by VIIRS, is a necessary but insufficient condition for this expansion. The ultimate goal must be to build a framework of ethical governance and transparency that ensures the responsible and equitable use of night-light data, thereby avoiding the reinforcement of existing power imbalances and the surveillance of already vulnerable populations. The following section will explore the critical need to advance causal inference and model explainability to ensure that the insights derived from these models are not only accurate but also trustworthy and defensible for use in high-stakes policy and decision-making contexts.

11. Conclusion¶

The application of satellite-derived night-light intensity as a leading indicator for regional consumer-spending shifts represents a significant advancement in economic monitoring, particularly in the context of data scarcity and real-time forecasting. This comprehensive review has synthesized the state-of-the-art, demonstrating that the transition from legacy DMSP-OLS to the high-fidelity VIIRS-DNB sensor has been transformative, resolving critical data quality issues such as sensor saturation, spatial blurring, and radiometric drift that previously invalidated many earlier studies [15][20]. The empirical evidence is unequivocal: models trained on VIIRS data exhibit dramatically superior predictive power, with robust, positive elasticities between night-light intensity and economic indicators like GDP and regional domestic product, even in complex, heterogeneous contexts [15][24]. This foundational improvement in data quality has enabled the development of sophisticated methodological frameworks, from time-series and panel data models to advanced machine learning and spatial-temporal frameworks, which collectively demonstrate the capacity to forecast consumer-spending shifts with a reliable lead time of one to three months [12][9]. The most compelling evidence for this predictive validity comes from high-impact case studies, such as the accurate prediction of a 24% YoY GDP contraction in India during the COVID-19 pandemic, which closely aligned with the official figure of 23.9% [12], and the real-time monitoring of economic reactivation in Wuhan, China, which provided an early signal of recovery several weeks before official data [9].

However, this progress is counterbalanced by a fundamental and unresolved methodological gap that undermines the reliability and generalizability of the proxy for its intended purpose. The most critical insight from this synthesis is that the predictive power of night-light data for consumer-spending shifts is fundamentally compromised by the lack of standardized, source-specific metrics and methodological frameworks to isolate non-consumption light. This gap is not a minor technical detail but the central obstacle to achieving defensible, actionable insights. Non-consumption light sources—such as gas flaring in extractive industries, agricultural burning for land clearing, and the activity of large fishing fleets—create persistent, high-intensity signals that are unrelated to commercial or residential economic activity and can confound the economic signal [16][6]. In regions with significant extractive industries, such as parts of Nigeria or the Niger Delta, these sources can dominate the night-light record, leading models to incorrectly attribute economic growth to the signal when it is actually driven by volatile, non-revenue-generating activity [16]. Similarly, in agricultural regions like parts of Southeast Asia, large-scale biomass burning creates intense, transient plumes that can be misclassified as signs of urban expansion or industrial development [16]. The absence of standardized, source-specific metrics to detect and correct for these confounding signals means that the predictive relationship between night-light intensity and consumer spending is systematically distorted, rendering the most sophisticated models vulnerable to spurious correlations and fundamentally limiting their utility for accurate, real-world forecasting.

This methodological flaw is compounded by profound socioeconomic limitations and ethical risks that constrain the ethical and equitable application of this tool. In understudied geographies, including Sub-Saharan Africa, Southeast Asia, and conflict-affected zones, the informal economy dominates and energy poverty is widespread. In these contexts, the night-light signal is a poor proxy for true economic activity, as it systematically underestimates the vast informal sector and can be distorted by state policies or infrastructure collapse, leading to a fundamental misalignment between the observed signal and its intended economic driver [10][24]. The ethical and geopolitical risks of using this data are not theoretical but are deeply rooted in power dynamics, institutional trust, and state control. The use of this data, particularly when conducted by external institutions, can be perceived as a form of surveillance, especially in regions with weak governance and vulnerable populations, potentially reinforcing institutional power imbalances and undermining national sovereignty [10]. Any robust application of this data as an economic indicator must therefore be underpinned by a framework of ethical governance, transparency, and, crucially, deep, contextual understanding of the socioeconomic and political realities on the ground. The future of this field lies not in seeking a universal, one-size-fits-all model, but in developing context-specific, multi-source data fusion frameworks that integrate night-light data with auxiliary indicators like mobile phone signals, electricity consumption, and humanitarian aid flows to enhance accuracy and provide essential context [16][12]. This approach, coupled with a commitment to explainability and ethical rigor, is essential to ensure that the promise of night-light data as a leading indicator is realized not only as a tool for improved economic forecasting but also as a force for more equitable and trustworthy global development. The following section will explore the critical need to advance causal inference and model explainability to ensure that the insights derived from these models are not only accurate but also trustworthy and defensible for use in high-stakes policy and decision-making contexts.

References¶

1. VIIRS Nighttime Light. Available at: https://eogdata.mines.edu/products/vnl/ (Accessed: September 01, 2025)
2. Measuring Quarterly Economic Growth from Outer Space in: IMF Working Papers Volume 2022 Issue 109 (2022). Available at: https://www.elibrary.imf.org/view/journals/001/2022/109/article-A001-en.xml (Accessed: September 01, 2025)
3. https://www.tandfonline.com/doi/full/10.1080/19475683.2019.1570336. Available at: https://www.tandfonline.com/doi/full/10.1080/19475683.2019.1570336 (Accessed: September 01, 2025)
4. NOAA NPP VIIRS. Available at: https://ncc.nesdis.noaa.gov/VIIRS/ (Accessed: September 01, 2025)
5. https://www.tandfonline.com/doi/full/10.1080/01431161.2017.1342050. Available at: https://www.tandfonline.com/doi/full/10.1080/01431161.2017.1342050 (Accessed: September 01, 2025)
6. https://www.sciencedirect.com/science/article/abs/pii/S014362281731281X. Available at: https://www.sciencedirect.com/science/article/abs/pii/S014362281731281X (Accessed: September 01, 2025)
7. Stock Price Time Series Data Forecasting Using the Light Gradient Boosting Machine (LightGBM) Model | Hartanto | JOIV : International Journal on Informatics Visualization. Available at: https://joiv.org/index.php/joiv/article/view/1740 (Accessed: September 01, 2025)
8. Emanuele Borgonovo, Elmar Plischke, Giovanni Rabitti. (2024). The many Shapley values for explainable artificial intelligence: A sensitivity analysis perspective. European Journal of Operational Research.
9. Applications of nighttime light data in international development research. Available at: https://www.3ieimpact.org/blogs/applications-nighttime-light-data-international-development-research (Accessed: September 01, 2025)
10. North Korea: average yearly night-time luminosity. Sources: Earth... | Download Scientific Diagram. Available at: https://www.researchgate.net/figure/North-Korea-average-yearly-night-time-luminosity-Sources-Earth-Observation-Group_fig1_339990994 (Accessed: September 01, 2025)
11. A Perspective on Explainable Artificial Intelligence Methods: SHAP and LIME. Available at: https://arxiv.org/html/2305.02012v3 (Accessed: September 01, 2025)
12. Using satellite images of nighttime lights to predict the economic impact of COVID-19 in India. Available at: https://pmc.ncbi.nlm.nih.gov/articles/PMC9128334/ (Accessed: September 01, 2025)
13. Mathias Kraus, Daniel Tschernutter, Sven Weinzierl, Patrick Zschech. (2024). Interpretable generalized additive neural networks. European Journal of Operational Research.
14. (PDF) Measuring Economic Growth Using Night Light Data: A Case Study of Vietnam. Available at: https://www.researchgate.net/publication/384359843_Measuring_Economic_Growth_Using_Night_Light_Data_A_Case_Study_of_Vietnam (Accessed: September 01, 2025)
15. https://www.sciencedirect.com/science/article/abs/pii/S0304387820301772. Available at: https://www.sciencedirect.com/science/article/abs/pii/S0304387820301772 (Accessed: September 01, 2025)
16. https://www.sciencedirect.com/science/article/abs/pii/S0040162524001951. Available at: https://www.sciencedirect.com/science/article/abs/pii/S0040162524001951 (Accessed: September 01, 2025)
17. The relationship between night-time light and socioeconomic factors in China and India. Available at: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0262503 (Accessed: September 01, 2025)
18. (PDF) Night-Time Light Data: A Good Proxy Measure for Economic Activity?. Available at: https://www.researchgate.net/publication/283263879_Night-Time_Light_Data_A_Good_Proxy_Measure_for_Economic_Activity (Accessed: September 01, 2025)
19. Night-Time Light Data: A Good Proxy Measure for Economic Activity?. Available at: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0139779 (Accessed: September 01, 2025)
20. Fifty years of nightly global low-light imaging satellite observations. Available at: https://www.frontiersin.org/journals/remote-sensing/articles/10.3389/frsen.2022.919937/full (Accessed: September 01, 2025)
21. https://www.sciencedirect.com/science/article/pii/S2772662223000048. Available at: https://www.sciencedirect.com/science/article/pii/S2772662223000048 (Accessed: September 01, 2025)
22. Lights and GDP relationship: What does the computer tell us?. Available at: https://link.springer.com/article/10.1007/s00181-023-02377-y (Accessed: September 01, 2025)
23. https://www.pnas.org/doi/10.1073/pnas.1017031108. Available at: https://www.pnas.org/doi/10.1073/pnas.1017031108 (Accessed: September 01, 2025)
24. https://www.sciencedirect.com/science/article/pii/S235293852100183X. Available at: https://www.sciencedirect.com/science/article/pii/S235293852100183X (Accessed: September 01, 2025)
25. https://www.sciencedirect.com/topics/earth-and-planetary-sciences/dmsp-satellite. Available at: https://www.sciencedirect.com/topics/earth-and-planetary-sciences/dmsp-satellite (Accessed: September 01, 2025)
26. https://www.imf.org/en/Publications/fandd/issues/2019/09/satellite-images-at-night-and-economic-growth-yao. Available at: https://www.imf.org/en/Publications/fandd/issues/2019/09/satellite-images-at-night-and-economic-growth-yao (Accessed: September 01, 2025)