Satellite-Based Estimation of Soil Geologic Properties Using Physics-Guided Machine Learning

The Central Valley region in California is one of the most agriculturally active regions in the United States. Aquifer storage capacity has been permanently reduced and land subsidence has been severe as a result of decades of intense groundwater withdrawal to support crop production. The local geology of the soil, especially the coarse-grain ratio (CGR), strongly influences the degree of subsidence. Usually, comprehensive borehole surveys are necessary to gather CGR data, however this data collection process is expensive and geographically limited. In this work, we present a physics-guided machine learning (ML) approach that integrates satellite-based observations of land subsidence and groundwater storage (GWS) change with effective stress theory and poroelasticity to infer the CGR of the soil. Our results show that the proposed framework can estimate soil CGR with sufficient accuracy, demonstrating that geologic composition can be inferred from satellite observations without relying on in-situ measurements. The presented approach supports more informed groundwater management and climate resilience by combining data-driven modeling with physical constraints to map soil geologic features in areas without direct ground measurements.