GeoChange: Predicting Socioeconomic Change from Earth Observation

Earth observation (EO) has enabled increasingly fine-grained mapping of socioeconomic conditions, but many policy applications require monitoring how local conditions change over time rather than estimating levels at a single date. Accurate point-in-time estimates do not necessarily yield accurate change estimates, as date-specific errors can propagate when estimates across dates are differenced. We study socioeconomic change prediction at policy-relevant administrative scales and introduce GeoChange, a hierarchical graph model that directly predicts changes in gross regional domestic product (GRDP) and population from paired EO observations and regional statistics. GeoChange aggregates visual information from fine-resolution image tiles to administrative districts, models spatial dependencies across districts, and uses a bottleneck and learned gate to selectively regulate high-dimensional satellite representations. Across 227 South Korean administrative units, GeoChange achieves $R^2$ values of 0.610 for GRDP change and 0.817 for population change, performing on par with the strongest adapted baseline under a common input setting. These results highlight the importance of treating socioeconomic change as a direct prediction objective and selectively fusing EO and regional information for administrative-scale monitoring.