Structured Pixels: Satellite Imagery as the Cause in Causal Effect Estimation

Abstract

We present Structured Pixels (SP), a causal inference model that positions satellite imagery as a cause/treatment in a causal graph, rather than merely a proxy for outcomes or confounders. Built on the generalized Robinson decomposition and a two-step, R-learner-inspired algorithm, SP uses learned latent representations to partial out confounding influences and isolate the causal effect. Its modular training pipeline supports integration with diverse machine learning models across domains. We evaluate SP using semi-synthesized datasets on two tasks: the impact of environmental conditions on mosquito populations and the influence of coastal characteristics on dark vessel prevalence. SP consistently outperforms baseline methods, and its learned representations capture meaningful environmental patterns. We further demonstrate SP’s applicability by re-examining the relationship between deforestation and agricultural productivity with real-world data; the results align with prior work. These findings highlight SP’s potential to advance GeoAI for environmental monitoring and resource management.