From Space to Economy: Deep Learning Fusion of Remote Sensing Data for Sub-National GDP Estimation

Abstract

Reliable and up-to-date sub-national data on the gross domestic product (GDP) data remain scarce in many regions across the globe. This limits the ability to analyze spatial economic disparities and to design evidence-based development policies. In this study, we investigate whether high-resolution GDP estimates can be derived from multisource remote sensing and auxiliary inputs using a deep learning fusion framework. We combine optical day- and night-time satellite imagery, and auxiliary geospatial layers with several backbone architectures and fusion strategies to assess the robustness of multimodal learning for economic prediction. Across extensive experiments in Brazil, we evaluate multiple encoders (ResNet-18, EfficientNet-B3, and SwinV2-T), alternative fusion layers (concatenation, attention pooling, graph-based multilayer perceptron, and mixture-of-experts), and diverse input modality combinations. Our best-performing model, an EfficientNet-B3 encoder with concatenation fusion using all available input modalities, achieves an R2 value of 0.87 for GDP prediction at a spatial resolution of 5 km×5 km, demonstrating that our multimodal approach effectively captures the complex relationships between spatial patterns and economic activity. These findings highlight the potential of multimodal remote sensing to complement traditional statistical sources by providing spatially consistent, high-resolution representations of economic activity.