Global extreme poverty has fallen dramatically over the past two centuries, yet hundreds of millions remain impoverished, underscoring the need for scalable monitoring tools. In Mexico, poverty metrics are available only sporadically in terms of time and space (e.g., every 5 years at the municipal level), making it difficult for decision-makers to access reliable, up-to-date, and sufficiently detailed information, highlighting the need for higher-resolution, timely methods. To address this problem, we propose a two-stage approach that combines socioeconomic and Earth Observations-based data. Initially, a machine learning model maps census variables to official poverty indicators belonging to a multidimensional model, yielding fine-scale poverty estimates. A census-based model trained with eXtreme Gradient Boosting (XGBoost) achieved a determination coefficient (R2) of approximately 0.842, indicating strong agreement with official poverty figures and providing high-resolution proxies. Afterward, we use features based on remote observations to predict these poverty estimates at a 469 m grid scale. In this case, advanced foundation models outperformed other machine learning (ML) approaches, achieving an R2 of 0.683. While foundation models enable more accurate, fine-scale poverty mapping and could accelerate poverty assessments, their use comes at a heavy price in terms of carbon emissions.
Salas et al. (2026) studied this question.