Evidence Pack: The Accuracy of Satellite Imagery and Machine Learning for Poverty Mapping

Satellite imagery combined with deep learning can predict local wealth with roughly 55% to 75% accuracy. It offers a highly scalable, low-cost alternative to household surveys in data-poor regions, allowing policymakers to map economic need from space. However, the evidence shows this method hits a hard ceiling at the individual household level, functioning best as a regional triage tool rather than a replacement for human enumerators.[1][6]

The core problem this technology solves is a severe data deficit. Traditional Living Standards Measurement Study (LSMS) surveys are slow, logistically complex, and expensive to conduct. In many developing nations, a decade or more can pass between comprehensive national surveys. Without recent ground-truth data, governments and NGOs are forced to distribute aid and plan infrastructure based on outdated assumptions.[3]

For years, the baseline remote-sensing solution was nighttime lights (NTL) data. Satellites capture luminosity at night, operating on the simple premise that brighter areas are wealthier. While useful at a macroeconomic scale, NTL has severe limitations. It suffers from a "blooming" effect where light spills into adjacent pixels, and more importantly, it fails to distinguish between the poor and the destitute—unlit rural areas all look identical to the sensor.[1][3]

The breakthrough in predictive accuracy came by shifting to high-resolution daytime imagery. But this introduced a new mathematical problem: you cannot simply train a Convolutional Neural Network (CNN) on daytime images to predict wealth directly. Deep learning models require millions of labeled examples to tune their parameters, and the very problem being solved is that ground-truth wealth data is scarce.[1]

To bypass this bottleneck, researchers utilize a mechanism called transfer learning. The CNN is first pre-trained on a massive generic dataset (like ImageNet) to recognize basic shapes and edges. Then, it is trained on a proxy task: predicting the intensity of nighttime lights from daytime images. This forces the model to learn the visual markers of human development—roads, metal roofs, agricultural patterns, and urban density.[1][2]

Once the model learns to identify these structural features, its penultimate layer acts as a feature extractor. Instead of outputting a light prediction, it outputs a dense mathematical vector representing the visual landscape of a specific geographic cluster. The model has effectively learned to "see" infrastructure.[1]

These extracted features are then fed into a simpler statistical model, typically a ridge regression, which is trained on the limited ground-truth LSMS survey data that does exist. This final step maps the visual features (like the ratio of metal to thatched roofs) to actual household consumption or asset wealth metrics.[1][3]

The evidence supporting this method is strong at the aggregate level. When evaluated at the village or cluster level, daytime CNN models explain between 55% and 75% of the variation in local wealth. This represents a massive leap over traditional nighttime lights, which typically explain only 27% to 32% of the variance in the same regions.[1][6]

However, the evidence degrades rapidly when the resolution is pushed to the micro level. The models struggle to differentiate wealth between individual households living in the same visual environment. If a destitute family and a lower-middle-class family both live under similar roofing materials on the same dirt road, the satellite cannot detect the difference in their caloric intake or cash savings.[3][5]

To bridge this gap, researchers have tested alternative data streams, most notably mobile phone metadata. Call detail records (CDRs)—which track call volume, network size, and prepaid top-up patterns—can predict individual wealth with surprising accuracy, capturing behavioral economics rather than just physical infrastructure.[4]

Comparing the predictive power across these methodologies reveals a structural limit to remote sensing. While daytime CNNs more than double the accuracy of nighttime lights, combining satellite data with mobile metadata still leaves roughly 25% of wealth variation unexplained. The models hit an accuracy ceiling because they are measuring proxies, not actual financial health.[4][6]

This unexplained variance represents the "invisible" economy. Informal debt, intra-household inequality, hidden cash savings, and social safety nets cast no shadow from space and leave no digital exhaust. A model trained purely on visual and digital proxies will systematically miss these crucial dimensions of poverty.[5][6]

Another significant weakness in the current evidence is cross-border generalizability. A model trained to associate specific visual features with wealth in Nigeria performs poorly when applied to Malawi without local retraining. The architectural markers of wealth—such as the transition from mud to brick, or thatched to tin roofs—are highly localized and culturally specific.[2][3]

Temporal accuracy also remains an open question. While multi-spectral satellite data has shown promise in tracking structural wealth changes over a decade, it struggles to capture short-term economic shocks. A village hit by a sudden crop failure or hyperinflation will still look identical from space the following week, even as actual poverty spikes.[2][5]

Ultimately, the evidence suggests that machine learning and satellite imagery are best deployed as targeting mechanisms rather than absolute measurements. They allow governments to rapidly identify the poorest districts across a vast country, ensuring that the scarce, expensive human enumerators are sent exactly where they are needed most.[3][6]