Satellites and machine learning are transforming how we see poverty on the ground. In recent years, Earth Observation (EO) data and machine learning have been used to create detailed maps that estimate economic hardship. These tools have shown impressive accuracy in recent studies, often between 70 - 90%, driving growing interest and investment in this emerging field. However, researchers warn that today’s models can appear precise while missing the true realities of life on the ground, risking an oversimplification of complex, dynamic and nuanced local conditions. In a study led by Dr Gary Watmough in the School of GeoSciences, researchers caution that many of the current “black-box” models – those created directly from data by algorithms that obscure how different variables are combined – may not capture the true multifaceted nature of poverty. The paper, 'A perspective on the interpretability of poverty maps derived from Earth Observation,' published in Science of Remote Sensing, argues that models built on raw satellite imagery must balance accuracy with explainability and be linked to local contexts if they are to be trustworthy and effective. How is Earth Observation used for poverty mapping? There are two main approaches to using Earth Observation data for poverty mapping.The first extracts interpretable features from satellite images, such as roads, buildings, or vegetation, which act as representations for poverty and wealth. These features are then statistically related to poverty data measured at household, village or district level through survey or census data. This approach allows researchers to understand and explain how the model reaches its conclusions, building confidence in the results. However, it often relies on local expertise and can be challenging to apply at large scales. The second method, which several studies have recently taken, uses raw satellite imagery fed into deep learning models. These models also depend on household or census survey data for training and validation, but instead of using predefined features, the algorithm automatically learns patterns from the raw satellite images. This can achieve high levels of accuracy but offers less transparency about how results are produced, making it harder for users to understand or validate the reasoning behind a specific poverty value. The accuracy and usefulness of the poverty maps derived from raw satellite imagery depend on how well the models reflect local realities, and maps that appear precise can still misrepresent the complex reality of poverty on the ground. Transforming Earth Observation images into poverty maps How can the models be improved? While Earth Observation data holds great promise for poverty monitoring and decision-making, its use remains limited. To realise its full potential, models must be developed and tested further to ensure they are robust, interpretable and ready for real-world application. The study urges a shift towards interpretable and operational models that make their reasoning clear and can be scrutinised by local communities and decision makers. Rather than relying on hidden patterns in raw imagery, models should use meaningful geospatial features such as road access and quality, building footprints and materials, land use, crop health, flood exposure and access to services. When these features can be directly linked to real-world conditions, the resulting maps become more transparent, trustworthy and useful for those they aim to support.As governments and aid organisations face tightening budgets and rising needs – from climate shocks to rapid urbanisation – misleading poverty maps risk misdirecting scarce resources. The researchers argue that truly transformative poverty mapping will depend not only on technical accuracy, but on the ability of communities and policymakers to understand and trust the results. The next generation of poverty maps, they suggest, will succeed not just when they predict well, but when they make clear why they predict as they do, turning satellite data into tools that can genuinely inform and improve lives on the ground. We wanted to try to emphasise the great progress made in this field of research whilst also highlighting what needs to happen next in order for people to trust our maps. Accuracy is important, but if no one can explain why the map looks the way it does, many will not feel comfortable using it especially if those decisions involve targeting millions of dollars of investment in a particular location.Realistically speaking we need more and better training data for our models, but this is prohibitively expensive, so we need to find ways to get our models trained with existing datasets whilst also ensuring they are understandable to a wide audience. Dr Gary Watmough Senior Lecturer, School of GeoSciences Read the full paper A perspective on the interpretability of poverty maps derived from Earth ObservationWatmough, G., Brockington, D., Marcinko, C. L. J., Hall, O., Pritchard, R., Berchoux, T., Gibson, L., Delamonica, E., Boyd, D., Mlambo, R., Ó Héir, S. & Seth, S., 1 Dec 2025, In: Science of Remote Sensing. 12, 100298. Related links View Dr Gary Watmough’s Edinburgh Researcher Explorer profileEarth Observation for Poverty – EO4Poverty website Publication date 07 Nov, 2025
