Poverty Estimation with Satellite Imagery at Neighborhood Levels Results and Lessons for Financial Inclusion from Ghana and Uganda By Soren Heitmann, Sinja Buri Table of Contents Executive Summary 4 Introduction 6 Data and Methods 7 Ground-Truth Survey data 7 Call Detail Record and Mobile Money data 8 Satellite Images 9 Spatial Segmentation 11 Spatial Boosting 11 Computer Vision Models 13 Results 14 Models 14 Benchmarking 14 Poverty Estimators 15 Exploring and Interpreting Poverty Maps 16 Discussion 24 Lessons for Estimating Welfare with Satellite Imagery and Call Detail Records 24 Application of Poverty Estimation Findings – Financial Inclusion and Beyond 25 Layering heat maps of poverty, telephone usage and financial activity 25 Increasing impact by identifying areas of biggest need and largest reach 26 Improved understanding of the financial behavior and needs of the bottom of the pyramid 26 Application beyond financial inclusion 26 Conclusion 27 References 28 2 Table of Contents Figures Figure 1: Enumeration coverage areas in Northern Uganda 7 Figure 2: Enumeration coverage areas in Ghana 8 Figure 3: Nearest-neighbor spatial boosting 12 Figure 4: Pooled observations of transfer learning model and nightlights model 15 by Jean et al. (2016) Figure 5: Observed vs Predicted PPI Score Distributions in Ghana 16 Tables Table 1: Comparison of Poverty Prediction Models 14 Table 2: Predicted Poverty Statistics, Telephone and Mobile Money Activity (per tower per month) 22 in Chorkor Image Tables Image Table 1: Daytime satellite sampled images: 67m2 resolution from DigitalGlobe 9 Image Table 2: Nighttime satellite sampled images: 750m2 resolution from VIIRS 10 Image Table 3: Comparing bottom and top wealth images in rural Uganda 11 Image Table 4: Aggregated spatial boosting in Uganda: comparing survey and predicted values 13 Image Table 5: Satellite image-based PPI Prediction Mapping – Ghana 17 Image Table 6: Empirical Observations Comparing Poverty Scores and Images – Urban Wealth 18 Image Table 7: Empirical Observations Comparing Poverty Scores and Images – Urban Poverty 1 19 Image Table 8: Empirical Observations Comparing Poverty Scores and Images – Urban Poverty 2 20 Image Table 9: Layering poverty predictions, telephone and mobile money activity in Ghana 21 3 Executive Summary that can, for example, detect features such as cars or trees on satellite images in urban areas. This constitutes a future area of research for the community. In lieu of more advanced machine vision feature detection, this study By successfully reaching historically underserved and employs spatial boosting techniques, which are found to vulnerable populations such as women, the poor, and people improve models for estimating poverty in rural areas where living in rural communities, Digital Financial Services have there are fewer welfare variations among neighbors than contributed to unprecedented growth in financial inclusion in urban centers. Although, with increased urbanization in Sub-Saharan Africa during the past decade. The adoption over time, this is another area that satellite imaging could and usage of DFS -- and the subsequent financial inclusion support. Changes in welfare over time, particularly due to that has resulted -- has helped reduce poverty and increase the availability of financial services, are likely to considerably prosperity throughout the region. Still, service providers outpace observable infrastructure changes, not least due to and development practitioners often lack reliable, detailed, the time needed to construct new buildings. Here, changes and low-cost poverty data that could help them accurately in financial service usage patterns through provider data identify additional communities and individuals who would are expected to yield stronger implications on financial benefit the most from access to financial services. The lack inclusion and livelihood effects on a year to year basis. of data hinders the deployment of services throughout the region and complicates efforts to monitor and evaluate the The study compared various statistical poverty estimation impact that interventions have on poverty. methods and identified the Poverty Probability Index (PPI), which yields better results with satellite imagery. Relying on traditional household surveys for poverty data Low activity levels and variation of transaction behavior is time consuming and expensive. What’s more, by the can make it difficult to use phone and mobile money time the data are collected and analyzed, it is often out of data for poverty prediction. Although, remote sensing date. But there are alternatives for estimating and mapping poverty estimation models can reduce the sample size for poverty with the goal of accelerating and expanding surveys, a broad spectrum of representative ground-truth financial inclusion and helping DFS providers target the survey data is essential for developing and training well poorest. Machine learning algorithms can, for example, performing poverty estimation models. In this respect, be trained to predict poverty based on imagery captured the research finds that remote sensing and geospatial by satellites and from call detail records, which document boosting approaches can be used to improve efficiency and mobile phone usage. For this research study, the IFC- optimization for traditional household survey methods. Mastercard Foundation Partnership for Financial Inclusion However, significant work remains before remote sensing collaborated with the Stanford University Sustainability models can fully replace ground-based surveys. and Artificial Intelligence Lab to advance existing poverty prediction models to generate poverty estimates at This paper also explores the interpretation of predicted neighborhood-level resolution, which is much more refined poverty scores, using PPI estimators, presenting them on than macro-level estimates produced by research to date. heat maps for Ghana at neighborhood-level granularity Satellite Imagery and call detail records (CDR), validated by and layering atop information about telephone and mobile ground-truth surveys, were used to develop models that money activity of users in the same areas to inform targeting can predict poverty in Ghana and Uganda. and monitoring of interventions for poverty reduction and financial inclusion. This visual layering is proposed The study finds that it is possible to make meaningful as a conceptual strategy for how combining techniques welfare estimates based on satellite imagery combined discussed in this study might be used to better quantify with geo-spatial boosting at the neighborhood-level when financial access, financial inclusion reach and support lower levels of precision are acceptable. The study makes providers to better understand customer demographics estimates about poverty demographics in regions that are and size their markets. bounded by cell tower locations. Predicting poverty around cell tower locations allows small area welfare estimation in urban environments where cell tower density is high. Above all, daytime satellite imagery proves to be a good basis for poverty prediction, but significant caveats remain. Models may be improved by adding context-specific segmentation 4 0.35 Mean: 62.5 Median: 63.3 Selected: 55.1 0.3 0.25 Proportion 0.2 0.15 0.1 0.05 0 10 23 36 49 62 76 89 PPI Predictions Executive Summary Visualization of Image Table 8: Mapping PPI predictive scores using the study mapping approach and predictive scores, compared against a satellite image. PPI is the Poverty Probability Index, a standard poverty estimator tool that can translate a PPI score into estimates of multiple benchmarks (eg, $1.90/day or $5/day or access to types of infrastructure). 5 Introduction Financial Inclusion empowers underserved individuals to As light diffuses over large areas, this approach alone participate in the formal economy, facilitates access to provides meaningful interpretation often only at the city- financial services that help businesses grow, and is critical level, or even at more roughly defined coverage areas to achieving economic development policies that aim to of larger administrative districts. As demographic and eliminate poverty. Digital Financial Services support these wealth variations are far more granular - both within development interventions by increasing the breadth of urban neighborhoods and in rural environments - satellite- delivery channels, variety of services, and affordability based poverty estimation models must deliver much more of financial access for consumers and companies. DFS granular estimates to yield sufficient information for are tuned to reaching segments that are historically policy makers to target underserved populations; and for underserved, such as women, rural individuals and the commercial DFS providers to better segment their potential poor. This is especially evident in Sub-Saharan Africa, where customer base and service coverage areas. cell phone penetration reached 44 percent as of 20181, meaning that nearly half of the one billion adults in the Using day-time satellite images provides an alternative region now have the potential to access financial services approach to resolve these granularity issues and deliver through mobile phones. The growing prevalence of DFS on results that are more aligned with the data required by the continent has been a driving factor in enabling financial policy-makers and DFS providers. This approach was access for poor and underserved individuals, as mobile demonstrated by Jean et al. (2016), using a convolutional money usage has increased from near nil just seven years neural network methodology to identify visible features in ago, to 20.9 percent by 2018. Today, financial inclusion is at high-resolution day-time satellite images, which correlate 43 percent in Sub-Saharan Africa2. While marking impressive with demographic data (e.g., roads, agricultural areas, reach, it is difficult to precisely quantify the extent to which urban environments, building types). the poorest segments are represented in this growth. This study expands the approach through a collaboration Development strategies to accelerate financial inclusion — between Stanford University’s Sustainability and Artificial and commercial providers seeking to scale Digital Financial Intelligence Lab and the IFC-Mastercard Foundation Services — lack access to reliable demographic data on Partnership for Financial Inclusion. The study engages poverty. Collecting data using traditional household surveys questions and areas of further exploration identified in is time consuming, expensive, and data are quickly outdated existing literature to specifically look at using day-time by economic changes and population movements. Using satellite imagine methods to predict poverty at the lowest remote sensing technology, call detail records and machine income segments (e.g., below $1.90, or $5.00 per capita per learning algorithms provides a solution to close this gap. day, using standard poverty threshold benchmarks). Call detail records have been successfully used to predict Here, different poverty estimation models are developed poverty in some countries; both, for models that attempt for two African countries. Multiple measures of poverty are to predict welfare based on call activity only, as well as employed to compare and understand relevance for training for combined models that include telephone data and models of this nature. The study compares modelling remote sensing covariates3. However, relying on CDR data methods and poverty definitions across these two country for regular poverty measurement may be complicated as contexts to learn about trade-offs and optimizations these data are privately managed by service providers. for developing models to predict poverty. The applied Unless data from all main service providers in a country research goal is to support DFS providers and financial is combined, poverty estimation is likely to be biased or inclusion policy interventions with a strategy for enhanced incomplete. information about markets, services and the characteristics of the people who use (or don’t use) these services. The Other methods have also shown promise. Notably, using approach defines demographic segments geographically, night time satellite images to view and measure ground- to establish tangible micro-markets as a unit of analysis, based light emissions that can correlate the magnitude of and then explores these segments with respect to predicted intensity and coverage area cast by light emissions with wealth characteristics, access and usage of digital financial economic activity and general well-being of denizens within services. the coverage zone4. While results from night light images are tantalizing, the level of granularity is low. 1 GMSA 2018 2 Demirgüç-Kunt et al. 2018 3 Steele et al. 2017 4 See for example Gosh et al. 2013 6 Data and Methods Ground-Truth Survey data • The SustainLab Asset-based Wealth Index calculated for this study used principal component analysis This study was implemented in Uganda and Ghana. In on responses to a panel of seven asset ownership both countries, ground-truth poverty data was collected questions within a household survey. The largest using household survey instruments. These instruments resulting principal component was used as an index incorporated modules to assess household poverty value. The hypothesis was that this index would and welfare levels. Instead of directly asking household potentially provide a better method of aggregating respondents about their consumption levels, which are different contributions of variables to derive poverty likely subject to inaccuracies due to seasonal fluctuations levels than a mere sum of scores that weighs different and recall bias, different poverty measurement tools answers to a list of questions, as the PPI methodology were used that eliminated the need to collect detailed does. This method was previously employed by Jean et consumption data. The survey instrument for Uganda al as a poverty predictor in remote sensing models and included a SWIFT (Survey of Well-being via Instant and was therefore used for prediction models in both Ghana Frequent Tracking) poverty estimation module. Whereas in and Uganda for consistency. Ghana, a PPI (Poverty Probability Index) estimation module was used. In addition, information about households’ asset Figure 1: Enumeration coverage areas in ownership was collected in both countries to calculate an Northern Uganda asset-based wealth index, using a similar approach as the SustainLab Index used similar research in this area5. • PPI is a poverty measurement tool to compute the likelihood that a surveyed household is living below a given poverty line based on answers to 10 country specific multiple-choice questions about household characteristics and asset ownership. Questions can also include visual, observable features such as house roofing material (e.g., is your roof tile, thatch, corrugated metal) or if there is an outdoor latrine. The PPI score is a value between zero and 100; it can be calculated for every household. The lower the score, the higher the likelihood of a given household to be poor. Look-up tables convert PPI scores into likelihoods to fall under different poverty lines in a country and may be interpreted for multiple different poverty threshold benchmarks using the same PPI score. • The SWIFT methodology was originally developed The Uganda survey focused strictly on Northern Uganda, to monitor one of the World Bank Group’s goals of one of the poorest areas of the country. In conjunction ending extreme poverty. It helps estimate household with another study investigating the adoption and impact expenditure data and poverty rates in a simple and cost- of DFS to better scale financial inclusion, IFC collected effective manner based on answers to 10-15 general data between November 2017 and January 2018 for 9,037 household level survey questions (e.g. education levels, households within 926 enumeration areas covering the asset ownership and household size). SWIFT models for Ugandan administrative areas of Karamoja, Mid North and specific regions and countries are derived from existing West Nile, and Adjumani (see Figure 1). household budget survey data (multiple rounds of LSMS surveys) indicating which variables are poverty To ensure ability to tune the satellite image-based correlates and should be collected in the core SWIFT modelling, the survey incorporated robust GPS data for survey to then estimate consumption and poverty each surveyed household, at high levels of precision6. This rates. aimed to resolve one of the issues that was previously faced by Jean et al. (2016), which drew on third-party geo- localized survey data that reduced precision by adding up to 10 km of random noise. Here, coordinates were precise within a few meters of survey location. 5 Neal Jean, Marshall Burke, Michael Xie, W. Matthew Davis, David B. Lobell, Stefano Ermon. “Combining satellite imagery and machine learning to predict poverty,” Science, 19 Aug 2016: Vol. 353, Issue 6301, pp. 790-794 6 GPS data achieved high levels of precision overall and the survey data collection methodology implemented robust cross-checking and validation to ensure accuracy and correction of GPS measurement errors. However, due to survey environments in very rural areas, often with farmers, individuals often responded to surveys at community village areas that were proximate to houses but not precisely located at the household for which household information was being reported. 7 Similarly, the Ghana survey ensured precise geo-localized The survey was implemented from December 2017 to survey data. The survey design covered a much larger area, February 2018 for 2,165 individuals within six enumeration spanning across Ghana, rather than the regional focus areas, in which coverage zones ranged between a one implemented in Uganda. Moreover, the enumeration areas and three km radius in seven Ghanaian cities and villages, were more focused on urban centers and densely populated distributed in five principle administrative regions. areas (see Figure 2). Figure 2: Enumeration coverage areas in Ghana. Zones are described in terms of their principle regional cities. 1=Bolgatanga, 2=Tamale, 3=Yendi, 4=Kumasi, 5=Tarkwa, 6=Accra + Tema Call Detail Record and Mobile Money Data 9,037 survey responses, only 222 matched the CDR data. For Ghana, of the 2,165 survey responses, 166 matched the CDR Through IFC project operations, the study incorporated data. With mobile money adoption levels still lying below the anonymized call detail records and mobile money transaction levels of sim card ownership, matching survey and mobile data from mobile network operators (MNOs) respectively money transaction data was even more difficult. In Ghana in Ghana and Uganda. Both operators have significant only 57 household records matched the respective mobile national coverage and the datasets provided customer-level money dataset. Ultimately, too few observations could be information on numbers of incoming and outgoing calls; matched directly to train meaningful prediction models. For SMS volumes; Cash-In transactions; Cash-Out transactions; the CDR models presented in this study, another approach and transfers between mobile money accounts. Information was therefore used for approximation. CDR models were about the geo-localization of activity through cell tower trained with household information that was aggregated and locations was also provided. matched with telephone activity data by cell tower catchment area and not by individual household. Results from these The study expected to identify correspondence between models are presented in Table 1 for the sake of completeness, call detail record data and the household survey data from but accuracy figures are unsurprisingly very low and hold respondent by matching phone numbers across these data little interpretive value due to the poor alignment between sets. The objective was to explore additional CDR-based survey and provider data and extremely small training sample models for predicting poverty. In both countries, the surveys underlying the model. were conducted using a randomized design; and in both cases, meaningful overlaps were not obtained between the surveyed customers and the CDR data. In Uganda, of the 8 Multiple factors may explain the low overlap of survey and Satellite Images mobile network operator data. Different time periods of data collection for survey and MNO data are one factor. Survey Models were developed using both day and night-time respondents may have joined the respective MNO networks satellite imagery. Day-time images were sourced from after CDR data was extracted or may have churned before. DigitalGlobe, with a high resolution of 67m2. Night-time It was moreover determined that survey enumeration areas images were sourced from VIIRS, at 750m2 resolution. Below, poorly overlapped with areas where these operators had example images are shown for Uganda and Ghana. Survey meaningful market share. Even though service was widely regions in Uganda were far more rural, by design. Whereas in available, other providers dominated these markets. This was Ghana, survey coverage included more urban and peri-urban confirmed by survey respondents, who reported using other environments across the country. The ‘ruralness’ in the Uganda providers. Lastly, in Uganda, the rural-focused survey also survey area is more pronounced in the nightlight images, found that only one-third of respondents had phones, which showing scant light signals in most of the images. The more considerably limited the pool of potential correspondences urbanized regions in Ghana, by comparison, show gradients between survey and CDR. This yields an important insight for of deep purple zones (dark, low-light emissions, correlating future research: there may be trade-offs between randomized to low electrification) to bright yellow zones (bright, intense- representativeness of a population sample and ability to light emissions, correlating to people using artificial light in meaningfully correlate demographic statistics with provider homes, offices, cars; general urbanization). These sample data. A stratified survey approach to deliberately over-sample images also illustrate the difficulty of using night-time light individuals who are customers of the service provider should emission imagery for household-level poverty prediction. At be considered. 750m2 resolution, entire neighborhoods can fit under a single pixel, and an entire city within a single image. Even though night-time images were incorporated into the predictive modelling, the coarse resolution of the information yielded little predictive value to improve model accuracy or descriptive power at the granular neighborhood level sought. Image Table 1: Daytime satellite sampled images: 67m2 resolution from DigitalGlobe Uganda Samples Ghana Samples Uganda shows increased prevalence of rural features: farms and open space surrounding single-level buildings. Ghana shows mixtures of features and increased urbanization, neighborhood housing configurations, paved roads and multi-story buildings. 9 Image Table 2: Nighttime satellite sampled images: 750m2 resolution from VIIRS Uganda Samples Ghana Samples The random sampling of representative nighttime images aligns with the day time images: the uniformly dark purple images in Uganda indicate extremely little light emission, meaning few people live in the coverage zone, or those that do aren’t using lighting at night. Whereas in Ghana, the bright yellow and shades of color depict increased urbanization and peoples’ collective usage of lighting within the coverage zone. 10 Spatial Segmentation the environs beyond may be entirely uninhabited. In that case, the poverty prediction model would therefore likely over- The MNO cell tower data was used to create geographical estimate poverty averages for the associated polygon area. segments by constructing Voronoi7 polygons based on the This acknowledged, poverty estimates may still reasonably tower locations. As both network operators had national estimate the average demographics of the individuals within a coverage, this provided a useful method of creating spatial polygon area, since disproportionally more people are likely to zones that covered areas from neighborhoods in densely live in a rural town as compared to the very sparsely populated populated urban areas (where cell tower density is high), up surrounding area. to much larger zones in rural or unpopulated areas (where cell tower density is low, or nonexistent). Although approximate, Spatial Boosting the technique enables grouping poverty estimates by cell tower coverage zone, and by association, to estimate Granular poverty prediction models based on satellite demographic averages of people living within the coverage imagery are challenged by individual images having relatively zone of their nearest “home” tower. For service providers, low-density of signal-rich features in a given image tile. For this allows interpreting customer demographics in terms of example, a grassy image might show a green field whose populations living near cell towers. trimmed grass is recently “mowed” by livestock in a sparsely populated rural area. Similarly, an image of trimmed grass For developing poverty prediction models, physical images might also show the manicured lawn of an upscale residential around cell towers were used, rather than the full geographic neighborhood in an urban area. The figure below illustrates area covered by the Voronoi polygon. This was for the sake this point, comparing top and bottom wealth photo examples of computational simplicity and the cost of accessing and in Northern Uganda. Where the wealthier image is similarly computationally processing high quality daytime satellite rural and depicting agricultural features, it also shows nuances images (Ghana, for example, would be represented by with more refined looking fields, higher quality thatch roofing approximately 53 million individual high-resolution image on out-buildings, and in the bottom-right corner, the cropped files). In this analysis, satellite images were downloaded portion of a larger building with an angular blue roof. In this from DigitalGlobe and VIIRS, around the geographical sense, it may be unclear if trimmed grass per se; or thatched areas corresponding to the GPS coordinates collected in roofs per se, correlate with income (let alone the ability of the household surveys conducted and the mobile network machine learning algorithms to identify such features). The operator tower sites. problem arises: which visual patterns generate signals to pay attention to. In high-density urban areas, higher tower density results in much smaller polygons, permitting the satellite picture around the tower to suitably represent the demographics of the overall area, typically a neighborhood or even smaller coverage zone. However, in rural areas, the Voronoi polygons are far larger due to low cell tower density. A shortcoming of the methodology is that it may not be reasonable to assume that the area seen on the satellite image around the tower is representative of the broader region. In fact, it may not be: in rural areas, a tower could be placed centrally in a town, while Image Table 3: Comparing bottom and top wealth images in rural Uganda Bottom 1% of surveyed wealth Top 1% of surveyed wealth 7 A Voronoi decomposition is a method of segmenting space around a set of points such that the borders of the resulting polygon area are equidistant from other adjacent points. Any given point is thereby located at the weighted center of the polygon, in relation to its neighboring points. 11 The ideal goal of remote sensing poverty prediction would be In rural areas that are sparsely populated, this assumption is to shine a viewpoint over any spot on earth, accurately extract stronger, as more geographic distance will likely be traveled visual information, and derive estimates of the demographic before large changes are observed in the population income norms of the individuals living in the area. Presently, the characteristics. In urban areas, a household might also be technology does not yet permit ex-ante predictions – ground- expected to evidence similar characteristics as immediate truth data is necessary to train predictive models with known neighbors, but the rate of change between a less-wealthy true data points. Research of this nature therefore requires and relatively more-wealthy neighborhood might be more relatively large survey samples. A challenge – also faced by this sudden. Indeed, this assumption is borne out in the results, study – is that surveys must be representative not only of the where spatial modelling in Uganda shows a higher R-squared demographic population, but also of the visual space. That is, value as compared to Ghana, where the small-scale spatial to have data for individuals across the income spectrum – and variation is higher, leading to lower explanatory power. also to have an additional dimension of the spectrum of visual environments in which they live. For example, what a low- The spatial boosting approach is depicted in Figure 3 below, income house and field in a rural area looks like, as compared where the central satellite image’s asset-based wealth index to a high-income house and field in a rural area. score is estimated by using the nearest neighbors whose scores are calculated using known survey data. In Uganda, surveys expressly focused on rural low-income areas and generated GPS coordinates for images around low- income households. The result was thin survey data of wealthy comparators with which to train models to differentiate the Figure 3: Nearest-neighbor spatial boosting visual cues associated with the spread of welfare and poverty levels. Somewhat conversely, in Ghana, survey data focused on urban areas, which resulted in fewer samples of what rural demographic variations looked like. However, as Ghana surveys had much broader geographic coverage, surveys and images were far more diverse compared with Uganda and generated a stronger set of features. To ameliorate the issue of survey visual variation, a method of spatial boosting was employed to estimate income demographic information for visual areas that did not have an explicit survey data point. This was done by creating Voronoi polygons between known household survey GPS coordinates, a strategy similar to the spatial segmentation employed at the level of cell tower locations. Here, images that were not directly associated with household GPS coordinates inferred poverty levels as a weighted average of several closest neighbors, by assuming that households near known survey locations were likely to evidence similar income demographics. 12 Aggregating the spatial boosting across the survey In Ghana, 8-nearest neighbors clustering was used. Image enumeration areas is depicted in Figure 3, for Uganda. The Table 4 below compares spatial boosting at the aggregate strongest signal was detected at an aggregation of 10-nearest level of the entire survey coverage area in Uganda: the visual neighbors. differences are quite small between the actual survey data and the predicted scores at the level of the Voronoi polygons. Image Table 4: Aggregated spatial boosting in Uganda: comparing survey and predicted values Ground-Truth Survey Score Poverty Prediction Score Gradients Gradients Computer Vision Models The satellite models presented below are so-called ResNet models. ResNet models are convolutional models that The satellite models that are presented in this study are through their design address the common challenge when models that predict poverty levels based on features derived training networks with multiple layers that normally let from satellite imagery through programmed visual pattern model performance saturate or decrease with the addition recognition. The models are convolutional neural networks, of layers (vanishing gradient problem). For this study, ResNet which are a classification of machine learning algorithms. models were used that had been pre-trained for pattern Meaning, with appropriate training, the computer can recognition from generic images and they were further fine- effectively learn to “see” relevant features in the associated tuned with additional layers and simple extensions based on images. the relevant country datasets in Ghana and Uganda and the study’s objectives to see features that correlate with income Convolutional neural networks are neural networks with demographics. multiple mathematical layers (between input and output layers) that can recognize visual patterns directly from pixel images with minimal processing since they filter pixel connections by proximity. 13 Results A variety of prediction methods were explored in the Uganda Models and Ghana contexts, as well as using different poverty score metrics as dependent variables for the models. A comparative In Uganda, using a combined model approach yielded an table of key models is shown below (Table 1) specifying each R-squared value of 0.28 using an asset-based wealth index as model by listing the category of features that were included in the outcome poverty metric. In Ghana, the highest R-squared the model (from satellite imagery, through spatial boosting or value observed was 0.2, using an asset-based wealth index as derived from call detail records) as well as the poverty metric the dependent variable. In basic terms, this means that the that was predicted respectively. Models that use spatial predictive models are able to explain 28 percent and 20 percent boosting in combination with satellite imaging yielded the of the variation in poverty observed in Uganda and Ghana, most explanatory power in both countries. respectively. Generally, these figures are not considered especially strong indicators of explanatory power. However, in the context of explaining differences in welfare from one neighborhood to the next, even a small percentage may offer meaningful insight. Table 1: Comparison of Poverty Prediction Models MODEL POVERTY METRIC R2 Ghana Satellite Asset-based wealth index 0.01 Spatial Asset-based wealth index 0.20 Satellite + Spatial PPI 0.15 CDR PPI 0.07 Uganda Satellite Asset-based wealth index 0.14 Spatial Asset-based wealth index 0.23 Satellite + Spatial Asset-based wealth index 0.28 CDR SWIFT 0.01 Benchmarking Explanatory power falls between 1x and 2x poverty line, suggesting difficulty in identifying visual signals to segment Among the poorest demographics, these results are gradients of poverty. Noting that the authors’ approach comparable to previous work conducted by Jean et al. (2016). yielded scores for larger geographic areas, the model was able In that study, the pooled results of the day-time satellite image to achieve R-squared results of up to 0.6 across all ranges of model yielded R-squared values of approximately 0.10 to 0.25 income demographic clusters, notably increasing explanatory for the set of poorest clusters below the poverty line of $1.90 power at levels greater than $5.00 per capita per day income per capita per day (see Figure 4). (i.e., approximately 3x and above). 14 Figure 4: Pooled observations of transfer learning model and nightlights model by Jean et al. (2016) 0.6 International poverty line 2x poverty line 3x poverty line 0.5 Probability of SME 0.4 0.3 0.2 transfer learning 0.1 nightlights 0.0 0 20 40 60 80 100 Poorest percent of clusters used This research explored poverty estimates at more granular Below, predicted poverty scores, their interpretation, and household and neighborhood levels. As previously noted, the comparison with actual images are explored and discussed CDR-based models were inconclusive due to poor ability to in more detail by the example of the PPI predictions of the match phone customers and survey responses and acquire a Ghana satellite model. Although model results primarily statistically robust sample. Therefore, the models produced incorporated the SustainLab asset-based index approach and results at the relatively granular resolution of neighborhoods, provide some comparability across the Uganda and Ghana as defined by areas in proximity to cell phone tower locations contexts, the PPI was considered to offer more interpretive at a variable resolution of the Voronoi polygons. As spatial power due to the ability to resolve PPI index scores across resolution of poverty estimates was variable, depending on multiple poverty benchmarks. Further, in the course of this cell tower density, the results are not directly comparable to study, some exploratory analysis suggested the design of the the more constant resolution discussed in Jean et al. (2016). PPI survey might better correspond to visual features that can Nevertheless, models achieving R-squared values of 0.28 be resolved by vision models. This may be one area where and 0.2 may be considered reasonable, given the nature of future research might specifically focus on identifying features the input data and granularity of estimates sought, and that that tools like PPI have established as statistically significant estimates were specifically targeting the lowest clusters of poverty estimators. observed income. For Ghana, using the PPI poverty estimator, the predicted Poverty Estimators distribution compares favorably with the observed PPI results from the survey. Across the 1,262 Voronoi polygon coverage Whereas other research has focused on income estimates in areas in Ghana, the model predicts a median PPI score of 63.3. a more absolute range across populations (such as predicting This is consistent with a median PPI score of 63 observed by a specific income value), this study incorporated different the household surveys.8 Figure 5 shows that the distribution poverty estimation methods, PPI, SWIFT and an asset-based of observed PPI scores and the distribution of predicted scores wealth index, to estimate poverty prevalence more generally. are very similar, centered around a score value of 62-63, with The PPI and SWIFT approaches achieve this by providing a most score variation happening ten score points below and statistical estimate that a household is simply above or below above this value, and slightly skewed toward higher (non- a given poverty line. Focusing at more granular spatial levels poor) scores. A score of 60-64 means that nine percent of the of urban neighborhoods results in lower power of models population is likely to fall below the $2.50/day poverty line; to explain the range of approximated levels of household and about 52 percent are likely to fall below the higher $5.00/ consumption and poverty incidence but the models show a day poverty line. reasonable ability to impute overall prevalence of poverty. 8 Statistical lookup tables that convert PPI scores (here between 60 and 64) into the corresponding likelihoods of falling below different poverty lines in a country are produced by Innovations for Poverty Action and are available here: https://www.povertyindex.org/country/ghana. 15 Figure 5: Observed vs Predicted PPI Score Distributions in Ghana Distribution of: PPI Predictions Mean: 62.2 Observed PPI distribution: Median: 63 10 Median 63 8 A high PPI score corresponds to a 6 lower probability of being poor. Percent 4 2 0 20 40 60 80 100 PPI Score 0.35 Mean: 62.5 Median: 63.3 Predicted PPI distribution: 0.3 Median 63.3 0.25 Proportion 0.2 0.15 0.1 0.05 0 10 23 36 49 62 76 89 PPI Predictions Exploring and Interpreting Poverty Maps In this manner, poverty estimation is more granular at a neighborhood level in higher density urban areas; whereas Using the results obtained in this study, poverty maps are in rural areas, polygons are far larger. Zooming-in on urban presented at varying national, regional and localized scales centers in Accra and Kumasi shows the granular nature of by using the cell tower geo-segmentation approach. Image the polygons, whose geographic area becomes smaller as Table 5 presents the mapping of predicted PPI poverty scores cell tower density increases. Many map areas do not have in Ghana at the country level. A total of 1,262 polygons are predicted poverty levels (they are filled with a gray checked visualized, nationally. pattern): satellite images were not available country-wide at the resolution used; some areas faced processing errors that Shaded polygons are established using mobile network resulted in incomplete mapping; and as already discussed, only provider cell tower locations, where darker shades show areas around cell towers attempted generating estimates, as estimates of low poverty incidence; lighter shades, higher computing several million images far exceeded the coverage incidence of poverty. With greater cell tower density to of the network and survey data. serve more densely-populated urban areas, polygon sizes become more granular, as do predictive score coverage areas. 16 Image Table 5: Satellite image-based PPI Prediction Mapping – Ghana Zoom into Accra Municipality Zoom into Center of Kumasi Map of entire country This Image Table visualizes the mapping tool developed for this project, providing shared coverage zones corresponding to the Voronoi polygon segmentation approach. The sequence of images illustrate the ability to zoom-in from country-level to neighborhood-level coverage areas. Here, PPI scores are depicted (darker is higher PPI score prediction, meaning higher wealth; lighter non-checked areas show low scores and therefore increased predicted prevalence of poverty). As discussed elsewhere, the tile-based mapping approach enables layering multiple indicators of interest. In terms of satellite imaging, urban areas are also more Visually exploring urban areas in Accra helps to make this feature-rich in terms of buildings and roads, while in rural point, while also illustrating the application (and challenges) areas there may be more grasslands or uninhabited areas. Yet, of the poverty estimation models combined with maps urban areas also have much more demographic diversity in segmented by the Voronoi estimation zones. The Image smaller areas, meaning neighboring households may be less Table 6 depicts one of Accra’s wealthiest areas, serving as an similar in terms of welfare, despite sharing common visual empirical example of high-income visual features. features in a satellite image. 17 Image Table 6: Empirical Observations Comparing Poverty Scores and Images – Urban Wealth Distribution of: PPI Predictions 0.35 Mean: 62.5 Median: 63.3 0.3 Selected: 64.7 0.25 Proportion 0.2 0.15 0.1 0.05 0 10 23 36 49 62 76 89 PPI Predictions Map: https://earth.google.com/web/@5.65555391,-0.111776,33.4206017a,314.48224383d,35y,0h,0t,0r Trasacco Valley is recognized as one of Accra’s wealthy neighborhood10; and also, the Southern area of Achimota, neighborhoods9. Selecting this area specifically on the which notably includes the slum area of Abofu (see lightly- predicted poverty map shows above-average PPI predicted shaded low-income estimated coverage area, highlighted with scores, although only modestly so. This clearly shows an orange-border polygon area). A satellite image snapshot of limitations of the model’s accuracy, with a predicted score the zone covered by Google Maps shows visual differences in of 64.7 – an improbably low prediction corresponding to 43 the housing density and construction of buildings, particularly percent below $5.00/day. This sort of discrepancy may likely clustered around the crossing highways. Identified through be an artifact of the spatial boosting approach combined with the predictive satellite mapping, the predicted PPI score in this satellite imaging. Zooming-in on the coverage area, multi- area is 53--on the lower end of the distribution of values across story single family houses are seen, lawns and swimming the country (see orange line in distribution chart). With a mix pools, which are expected to correlate with near-zero poverty of more affluent housing stock and slum areas, this score for the coverage area. indicates probability of 68 percent of denizens in this area to fall below the $5.00/day poverty line. One of the poorest zones predicted within the greater Accra area, depicted below (Image Table 7), includes the Northwest section of the Abelemkpe, a relatively wealthier 9 https://www.africa.com/a-million-gets-you-in-ghana/ 10 https://en.wikipedia.org/wiki/Neighborhoods_of_Accra 18 Image Table 7: Empirical Observations Comparing Poverty Scores and Images – Urban Poverty 1 Distribution of: PPI Predictions 0.35 Mean: 62.5 Median: 63.3 0.3 Selected: 53 0.25 Proportion 0.2 0.15 0.1 0.05 0 10 23 36 49 62 76 89 PPI Predictions Map: https://earth.google.com/web/@5.61074734,-0.2240724,20.66296749a,1087.02950268d,35y,0h,0t,0r Another example of a poorer neighborhood in Accra is depicted Although the predicted poverty level for this area is at the below in Image Table 8 at different zoom levels. Chorkor is lower end of the distribution poverty scores across the county a fishing village at the coastline in Accra. The corresponding (see distribution in Image 8), it is not as low as expected for a satellite image shows a densely populated neighborhood at slum area known for its high levels of poverty and lower access the coastline in Accra. It is a fishing village struggling with to infrastructure. This result may be explained by a closer look poor sanitation, access to water and power infrastructure, at the Google Maps satellite snapshot. Apart from the slum and waste management. The predicted PPI score for this area area at the bottom half, the upper part of the image shows is 55.1 which corresponds to a 60.3 percent likelihood for the less dense housing structures surrounded by more greenery population living there to fall below the $5.00/day poverty suggesting higher levels of wealth. Indeed, this upper part of line. the image shows a university and a hospital campus. 19 Image Table 8: Empirical Observations Comparing Poverty Scores and Images – Urban Poverty 2 Distribution of: PPI Predictions 0.35 Mean: 62.5 Median: 63.3 Selected: 55.1 0.3 0.25 Proportion 0.2 0.15 0.1 0.05 0 10 23 36 49 62 76 89 PPI Predictions Map: https://earth.google.com/web/@5.5334772,-0.23087813,18.66830879a,2254.39682943d,35y,0h,0t,0r Both examples of neighborhoods presented above (Image • Map A.1 depicts the satellite-based predicted PPI scores for Table 7 and 8) are areas where welfare levels were predicted Ghana. Darker shades visualize higher scores in respective solely based on the underlying satellite imagery. None of the areas, which translate into lower predicted poverty survey data used for model training was collected in those incidence in those polygons. areas. The fact that both wealthy and poor areas are covered by the polygons respectively, explains moderately low predicted • Map B.1 visualizes call activity for users of a Ghanaian poverty incidence and provides anecdotal evidence that the mobile network operator. The map shows gradients of satellite model for PPI estimation aligns to some degree with telephone calls incoming to respective smaller areas. observed characteristics. Darker areas depict relatively more calls received. These two examples also illustrate the complexity in • Map C.1 shows mobile money transaction activity. The generating granular neighborhood-level estimates, especially map shows gradients of the total value of transfers that in more urban environments precisely because of the rapid are being sent and received in a respective area. The higher changes that may occur between low- and high-income the value of transfers per month, the darker the shade. segments, the visual features that characterize them, and general lack of border boundaries (e.g., a political or administrative line). Ultimately, the goal of this research is to explore the interplay of poverty and Digital Financial Services. Image Table 9 shows how poverty heat maps can be meaningfully compared to layers of telephone and mobile money activity. Three metrics are selected for comparison: 20 Image Table 9: Layering poverty predictions, telephone and mobile money activity in Ghana Map A.1–Poverty Prediction Map B.1-Telephone Activity Map C.1-Mobile Money Activity (PPI score) (Lighter=Poorer) (Number of Incoming Calls) (Total Value of Transfers) Map A.2 – Zoom into Accra Map B.2 - Zoom into Accra Map C.2- Zoom into Accra Across the maps, the same polygon areas are shaded,11 mobile money activity layer as adoption levels are still largely enabling the ability to directly layer transactional data atop lacking behind cell phone ownership. Zooming-in on urban poverty estimates. Moreover, as polygons are approximating areas (see for example Map B.2 zoomed into Accra in Image mobile network operator service areas, insights are equally Table 9) shows again the granular nature of the polygons, valuable to providers seeking a better understanding of whose geographic areas become smaller as cell tower density consumer segments with respect to their service areas. increases. As a result, these urban zones also show significant gradients of calling and mobile money activity between them Regarding call behavior and mobile money transaction at this level of resolution. activity, nationally, darker-shaded urban areas show increased activity, as would be expected. This is most pronounced for the 11 Polygons with missing data are again filled with a gray checked pattern. Different reasons can explain missing data. No available high-resolution satellite imagery, processing errors or no recorded call or mobile money activity during given time period in respective polygon. 21 To illustrate the feature layering approach with a concrete values across locations in Ghana. By definition each Voronoi example, one area that was discussed before (capturing parts polygon constitutes a geographic area with a single cell tower of the poor village of Chorkor in Accra) is again highlighted at its geometric middle. Therefore, values may be interpreted with orange boarders (Image Table 9). Table 2 lists the in terms of volume of activity per tower per month. corresponding poverty, telephone, and mobile money activity metrics for the selected areas, comparing them to the median Table 2: Predicted Poverty Statistics, Telephone and Mobile Money Activity (per tower per month) in Chorkor EXAMPLE POLYGON ZONE METRIC GHANA MEDIAN (PART OF CHORKOR) Poverty Statistics Predicted PPI Score 55.1 63.3 $5.00/day poverty rate - PPI interpretation 60.3% 52.1% $1.90/day poverty rate - PPI interpretation 3.6% 1.5% Telephone Activity Number of outgoing calls per month (per month and user) 104 47.5 Number of incoming calls (per month and user) 41 15 Outgoing call duration (total tower minutes per month) 129 hours 69 hours Incoming call duration (total tower minutes per month) 61 hours 29 hours Number of incoming SMS per month (total per tower) 26,600 7,900 Number of outgoing SMS per month (total per tower) 81,000 39,000 Mobile Money Activity Mobile money transfer (average amount per month) $21 $19.5 Mobile money cash in (average amount per month) $13 $15 Mobile money cash out (average amount per month) $17 $16 The predicted poverty incidence in the selected polygon is Overall, this shows that in this specific area, a community with an estimated 60.3 percent of the population living below with higher poverty prevalence also shows much higher the $5.00/day poverty line, which is an eight percentage point telephone usage; and similar mobile money activity patterns higher poverty rate than the median value in Ghana. But (slightly biased toward cash-out, suggesting net inflows into despite low welfare levels, the area still shows high levels of the community). telephone activity. Across all call activity metrics, the values for the specific polygon are more than twice as high as the As previously observed, this neighborhood shows a mix median values. In other words, the cell tower in this area of features that expect to correlate with higher and lower hosts a highly active userbase, as compared to tower and user poverty prevalence (eg., slum areas adjacent to areas with communities elsewhere. single family homes and lawns). It is impossible to identify wealth characteristics at the individual user level to know Regarding mobile money activity, results differ depending if the telephone and mobile money patterns are driven by on the metric. Mobile money activity in the area is higher wealthier demographics or poorer demographics or evenly than the countrywide median with respect to the volume of distributed across all users. However, what is known – and cash-outs and mobile money transfers; whereas the average what is important from the perspective of both providers and cash-in amount is lower than in the majority of other polygons policy makers is this: the community depicted here shares a across the country. common infrastructure. 22 Telephone statistics are reported in terms of the traffic served computational approach that explores relationships among by the tower at the geometric center of the polygon; cash-in these types of data would identify “hot spots” of interest and cash-out statistics are similarly reported in terms of the according to specific strategic interests for providers or tower that intermediated the agent float balances to facilitate policy-makers. Digital financial service providers and donors the service transaction. Consequently, any commercial or might use the layering approach to compare even static or developmental interventions designed to expand financial slowly changing poverty baseline estimates with a variety of access will reach communities that access those services different indicators that help to monitor and identify areas via this shared infrastructure. It is therefore meaningful to for targeted interventions to reduce poverty and to increase articulate the reach of financial services with respect to the financial inclusion. Remittance rates as well as other metrics demographic make-up of the communities to share the of (net) financial inflows and outflows of neighborhoods, and “home” network tower in their neighborhood. especially population numbers that better estimate financial reach and micro-market sizing, are interesting indicators for The Chorkor neighborhood discussed here was selected layering atop of poverty rates. simply by having a low-scoring PPI prediction for the sake of exemplifying a layered analytic approach between poverty models, GSM and DFS activity. Overall, a refined 23 Discussion Lessons for Estimating Welfare with Satellite Imagery tower12. It is reasonable to assume that tower density is and Call Detail Records proportional to population density (or at least provider subscriber density). That is, providers are incentivized to It is necessary to measure welfare and poverty levels regularly, put more towers where increased service is need. Doing at high spatial resolution, at high temporal frequency, and at so refines coverage polygons into smaller geographic low cost. The increasing availability of day- and night-time spaces, importantly characterized by the people using satellite images and powerful deep learning algorithms has the shared access point. The area defined by the Voronoi introduced new methods to predict poverty and welfare levels. polygon therefore describes the DFS usage statistics of This research aimed to further these methods, specifically by the denizens since the tower intermediates transactions increasing the granularity of analysis to smaller areas. performed by users and agents. Overall, the study finds that it is possible to identify 3. Daytime satellite imagery improves poverty prediction, meaningful welfare estimates at neighborhood-level but caveats remain. Nightlight satellite imagery can resolution. However, these estimates are likely to lack provide baseline estimates for regional poverty, but precision. A joint spatial/satellite model provides the they are less useful in rural areas that do not have highest explanatory power, which combined interpolated much variation or nightlight signals due to lower levels geo-marked survey data with machine vision feature of electrification. Daylight satellite imagery provides identification. This demonstrates that there are components a better alternative in many cases, although high- of estimated wealth that are detectable through satellite resolution imagery is not always available for all regions imagery. While ground surveys are still necessary to develop in a country and individual high-resolution satellite country-specific models, adding remote-sensing information images are unlikely to have uniformly distributed can reduce the sample sizes needed for detailed poverty features that carry meaningful signals of poverty or estimation. wealth. Enough variation in wealth exists across the visual space to make wealth estimation with day time The following are key lessons learned from this study about satellite imagery difficult without a robust ability to different data sources, methods, and challenges depending detect and identify the features that characterize the on context and targeted levels of granularity: visual space. There is room for ample improvement for 1. Evaluating welfare at neighborhood levels and with high granular level poverty estimation, especially for urban spatial resolution may be valuable when lower levels neighborhoods. In this study, higher R-squared values of precision are acceptable. A rough understanding of for rural Uganda are due to the relatively lower rate of income can meaningfully segment geographic areas that change across adjacent satellite images. Urban Ghana’s are below (or above) specific poverty threshold, such as rapid feature changes across smaller geographic space in this case where poor versus not-poor can characterize results in fewer salient features (or conflicting features) neighborhoods or provide estimators for service reach to in the visual space. Here, further research might focus demographic segments. More so, when rough estimates on specialized feature detection, such as models that can help to describe highly variable wealth demographics can detect cars, prominent urban characteristics and among neighbors in densely populated areas, it may be other indicators of wealth. possible to approximate general poverty preponderance In the course of this research, it was conjectured that additional within that neighborhood, rather than a specific per- feature detection research might prioritize identifying features household value. that correspond to visually-identifiable features of poverty 2. Poverty prediction at the level of Voronoi polygon-based survey tools. Specifically, the PPI methodology uses some cell tower locations allows small area welfare estimation statistical measures that have strong visual determinants in urban environments. Predicting poverty at lower than (such as a building’s roof material, or whether there is an regional levels raises the question of how to segment outdoor latrine). While a challenging problem to solve, it space – simply, where do boundaries exist? Political is nevertheless reasonable to train a visual model to see a or map boundaries may or may not exist in a national thatch roof or metal roof or tile roof, for example, and perhaps context, especially for smaller towns. More importantly, recognize features like community washrooms or cisterns how political boundaries are drawn is unlikely to reliably for potable water. Whereas poverty survey tools driven by characterize demographic features of people who consumption data, ownership or household expenditure live within that zone. Using cell towers to segment provide less direct opportunity to “see” these types of features geographies is beneficial, as the Voronoi polygon approach in the visual space and interpret them accordingly. groups populations in terms of proximity to nearest 12 Cell tower location data is available publicly or may be purchased by independent organizations that map infrastructure locations, globally, such as OpenCellID (https://www.opencellid.org/). 24 An area of future research might therefore seek to train visual Application of Poverty Estimation Findings – Financial models specifically to recognize observable features present in Inclusion and Beyond PPI (or other methods) to improve prediction accuracy. Research shows that telephone usage13 and increased social 4. Spatial boosting is particularly helpful to improve models network size14 are strong predictors for active uptake and use of for rural poverty estimation. The ostensible goal of using Digital Financial Services. Moreover, Digital Financial Services remote-sensing to estimate poverty is to reduce the boost financial inclusion and contribute to poverty reduction time and expense of ground surveys. This study found and improved livelihood indicators. DFS tend to be adopted that using spatial boosting helps to address this, as first among higher income demographics, particularly urban meaningful estimates can be inferred for non-surveyed youth. They scale by diffusing from early adopters and are likely image locations by weighting surveyed observations to grow along remittance corridors or social networks (such from nearest-neighbors. This approach was found to be as urban laborers who send money home to families in rural more effective in rural areas that are less likely to have areas).15 Identifying these corridors is key, and normalizing the substantial variations in welfare over short distances. use of DFS is a means of scaling financial inclusion. Tracking Whereas in urban areas, large disparities of wealth were financial flows into (or out of) areas with low-income welfare observed among neighbors, posing a significant challenge estimates may help to monitor the reach of financial inclusion for training machine vision models. and help target areas of greatest need. 5. Representative sampling may not meaningfully overlap Given that DFS can play a significant role in diminishing poverty, with provider data. This study also tried to predict poverty it is important to be able to accurately identify locations levels in neighborhoods with call activity data, expecting where the poor live for the purpose of deploying targeted randomly selected survey respondents to be sufficiently financial inclusion strategies, as well as for monitoring the represented in provider data to model CDR usage and use of financial services, and observing the impact they have wealth demographics. For both countries, the survey on the population. Current national survey methods are slow results effectively showed that providers had relatively low and costly, meaning that observing and measuring reach and market share in the enumeration zones. Although results change is likely to take place on multi-year time lines. Even are presented in Table 1 for the sake of completeness, there techniques that employ remote-sensing perfectly, while less is little interpretive value due to very low coincidence expensive, may also be slow to observe poverty changes. between CDR users and survey respondents. Therefore, However, indicators of financial empowerment represented future similar research should conduct minimal baseline through provider data change much faster, at the rate of surveys to understand general market share when usage and uptake. attempting to use provider data and then design the full survey to over-sample in a statistically controllable Layering heat maps of poverty, telephone usage and manner to ensure adequate coincidence between data financial activity sets. Financial inclusion insights can therefore be obtained 6. Broad-spectrum representative ground-truth survey from comparing different data layers of telephone usage, data is essential for training poverty estimation models. financial transaction activity and poverty levels to deepen the Breadth here implies that ground-truth welfare data understanding of how they are interconnected. A base layer encompasses the range of economic well-being within of poverty estimates is fundamental to drive these types of the population. Data should come from households with insights, which can benefit providers and policy makers alike. sufficient geographical dispersion so that the number The maps presented in this report depict single-variable layers of areas they fall in are high enough to train machine to illustrate the approach. But further research is necessary to learning models if the variance is too small. Additionally, computationally aggregate population, income and financial breadth also implies that images selected also encompass activity estimates to quantify the reach and scale of financial the range of buildings, roads, fields, farms and relevant inclusion meaningfully at the national-level and at a more features that are representatively associated with the granular scale. Population layers are also critical to further spectrum of welfare. The variation should ensure sufficient this approach and should be considered equally in further feature capture to identify and differentiate a wealthy research. household’s manicured lawn from a poorer household’s adjacent pasture, for example; equally for urban areas, to ensure that the variety of visual features are captured along with the variety of wealth segments that may correlate with those features. This problem was evident in this study’s focus in rural Uganda, particularly since (by design) surveys focused on the poorest households, but this also resulted in a relative lack of wealthy households against which to compare and train models. 13 IFC 2016; Blumenstock et al. 2015. 14 Mattson and Stuart 2018 15 IFC et al. 2017, Aga and Martinez Peria 2014 25 Increasing impact by identifying areas of biggest need Application beyond financial inclusion and largest reach This study was more specifically focused on identifying This research finds that satellite imaging can be used to poverty as a basis to compare and assess with respect to the meaningfully segment welfare levels at neighborhood- prevalence of Digital Financial Services. However, the need level granularity, although with relatively low precision. for regular and granular poverty prediction with the help of While these models offer only modest ability to explain the satellite imagery and call activity data goes of course beyond variation in wealth at granular levels, the ability to segment financial inclusion. Layering publicly-available population and rank poverty estimates can identify key areas to focus information16 onto cell-tower location-based polygons allows on, potentially advancing both commercial and financial for example to approximate estimates for how populations inclusion strategies. Further, layering poverty estimate data with different income demographics access Digital Financial can identify financial inclusion engagement opportunities Services. Applications in other domains is also possible, by (i.e., high cell coverage, low welfare) or populations that are assessing the proximity to access different services as well as particularly underserved that donors may seek to strategically the coverage density provided to populations within an area target (i.e., low cell coverage, low welfare). In this manner, of interest. Other areas of application include for example poverty estimates such as those obtained through this study agriculture and infrastructure. can provide viable insights, despite the low precision of results: relative rank of welfare estimate is sufficient to provide directional information on financial inclusion targeting and reach, as does a categorical assessment of poverty prevalence above-or-below a given threshold. Improved understanding of the financial behavior and needs of the bottom of the pyramid Comparing poverty estimates with financial activity data helps to explore the scale of financial inclusion among the poorest income demographics. Providers seeking to better understand their own markets and customer demographics may gain insight into how services are used across geographies and income demographics. Such as whether money is sent and received from high-to-low predicted poverty areas or vice versa; or to quantify the volume of activity with respect to these parameters or relative per-capita metrics within coverage areas. Do these wealthier and poorer segments make phone calls to each other? Do remittances flow from one to another? If so, to what degree? If not, how may remittance and communication corridors be described in terms of the demographic characteristics of sender and receiver zones? 16 Such as Center for International Earth Science Information Network https://www.ciesin.columbia.edu/data/hrsl/ or WorldPop http://www.worldpop.org.uk/data/summary/?doi=10.5258/SOTON/WP00098 26 Conclusion Neighborhood level poverty estimation with satellite The ability to map small area poverty estimates and to imagery is possible, although aided significantly by spatial combine them with layered financial transaction data, boosting techniques that draw on traditional survey data. as explored in this study, provides opportunities for Combined, the coverage of surveys is effectively increased development professionals and Digital Financial Services substantially, enabling smaller sample sizes to yield more providers alike to identify and quantify engagement, information. While the precision of poverty estimates is particularly among the poorest individuals. Equally, to limited, the ability to segment and rank geospatial areas identify opportunities where high engagement on telephone in terms of welfare is nevertheless insightful. Further work channels or other demographic characteristics may signal is needed to refine the models developed in this study and opportunities to strategically engage underserved markets to develop research of this nature into insights for service that are likely to adopt and benefit from improved services. providers. However, the basic building blocks are here to start using them. Even directional information on estimates of wellbeing can help to direct better understanding of financial inclusion reach. 27 References Aga, Gemechu Ayana; and Martinez Peria, Maria Soledad (2014). “International Remittances and Financial Inclusion in Sub- Saharan Africa”. Policy Research Working Paper No. 6991. World Bank Group, Washington, DC. Blumenstock, Joshua (2016) “Fighting Poverty with Data.” Science 353 (6301), 753-756 Blumenstock, Joshua; Cadamuro, Gabriel; On, Robert (2015). “Predicting Poverty and Wealth from Mobile Phone Metadata.” Science 350 (6264), 1073-1076 Blumenstock, Joshua (2018). “Don’t forget people in the use of big data for development”, Nature, 561 (7722), 170-172 Blumenstock, Joshua (2018). “Estimating Economic Characteristics with Phone Data.” AER Papers and Proceedings, 108, 72-76 Demirgüç-Kunt, Asli, Leora Klapper, Dorothe Singer, Saniya Ansar, and Jake Hess (2018). The Global Findex Database 2017: Measuring Financial Inclusion and the Fintech Revolution. World Bank: Washington, DC. - Global Findex 2017 https:// globalfindex.worldbank.org/ Donaldson, Dave and Storeygard, Adam. (2016) “The View from Above: Applications of Satellite Data in Economics.” Journal of Economic Perspectives – Volume 30, Number 4, 171-198 Fung, Vincent (2017). “An Overview of ResNet and its Variants.” Medium – Towards Data Science. https://towardsdatascience. com/an-overview-of-resnet-and-its-variants-5281e2f56035 Ghosh, T.; Anderson, S.J.; Elvidge, C.D.; Sutton, P.C. (2013). “Using Nighttime Satellite Imagery as a Proxy Measure of Human Well-Being.” Sustainability 2013, 5, 4988-5019. GSMA (2018). “The Mobile Economy” - https://www.gsma.com/mobileeconomy/wp-content/uploads/2018/05/The-Mobile- Economy-2018.pdf IFC (2016). “Find the Gap: Can Big Data help to increase Digital Financial Services Adoption?” Partnership for Financial Inclusion IFC, Cignifi and Airtel Uganda (2017). “Mobile Wallet Usage Study – Apply CDR models to increase Mobile Wallet Adoption & Activity.” Netmob. Book of Abstracts, p.11 http://www.netmob.org/www17/assets/img/bookofabstract_oral_2017.pdf Jean, Neal; Burke, Marshall; Xie, Michael; Davis, W. Matthew; Lobell, David B.; Ermon, Stefano (2016). “Combining satellite imagery and machine learning to predict.” Science 353 (6301), 790-794. Mattson, Carolina and Stuart, Guy (2018). “Understanding Key Mobile Money Users.” MFO PPI – Poverty Probability Index - https://www.povertyindex.org/ Steele, Jessica E.; Sundsøy, Pål Roe; Pezzulo, Carla; Alegana, Victor A.; Bird; Tomas J., Blumenstock; Joshua; Bjelland, Johannes; Engø-Monsen, Kenth; de Montjoye, Yves-Alexandre; Iqbal, Asif M.; Hadiuzzaman, Khandakar N.; Lu, Xin; Wetter, Erik; Tatem, Andrew J. and Bengtsson, Linus (2017). “Mapping poverty using mobile phone and satellite data.” Journal of the Royal Society Interface –14 (127) - https://royalsocietypublishing.org/doi/10.1098/rsif.2016.0690 Suri, Teevnet; and Jack, William. 2016. The long-run poverty and gender impacts of mobile money. Science, 354 (6317):1288-1292. World Population Review Ghana - http://worldpopulationreview.com/countries/ghana-population/ - Estimate as of January 31st, 2018 (reference date during survey data collection) Yoshida, N., R. Munoz, A. Skinner, C. Kyung-eun Lee, M. Brataj, W. Durbin and D. Sharma (2015). “SWIFT Data Collection Guidelines Version 2” AUTHORS Soren Heitmann leads the applied research and learning program for the Partnership for Financial Inclusion. His background is in data science, development economics and cultural anthropology. Sinja Buri is a data operations analyst for the Partnership for Financial Inclusion. Her research focuses on digital financial service customer behavior and demographics and applying insights for product development and strategy. CONTRIBUTING AUTHORS Guanghua Chi, doctoral student at the UC Berkeley School of Information and Nikhil Desai, software engineer at Google (formerly a researcher at the Stanford Sustainability and Artificial Intelligence Lab) also contributed to this report. ACKNOWLEDGMENTS IFC and the Mastercard Foundation Partnership for Financial Inclusion are grateful to Marshall Burke and Joshua Blumenstock for their scientific input and oversight of the model development for this research project. To the Bill & Melinda Gates Foundation for supporting the research engagements whose data are included in the analyses. Additional appreciation to Shafique Jamal for his support in mapping and visualizing poverty predictions and to Gary Seidman and Lesley Denyes from IFC for their editorial support. April 2019