Predicting Drought, Preventing Famine: Advanced Soil Moisture Analysis for Climate Resilience

A two-phase methodology combining temporal Graph Neural Networks (GNNs) with physics-based data assimilation to create continuous global soil moisture coverage from satellite missions, addressing critical gaps in drought monitoring and food security applications.

Problem Statement

Surface soil moisture data represents a valuable input for food security, famine early warning system networks, as well as food action collaborative programs. Satellite-based soil moisture retrieval has advanced significantly in recent years through missions like the Soil Moisture and Ocean Salinity (SMOS) and NASA's Soil Moisture Active Passive (SMAP). These missions provide critical data on surface soil moisture (SSM) at depths of 3–5 cm, which is vital for crop yield estimation, drought monitoring, irrigation management, and urban water planning.

Currently, no truly coherent, comprehensive global soil moisture database with systematic quality control exists. This challenge is even more pronounced for root-zone soil moisture (RZSM) at depths of 30–110 cm, which traditionally requires ground-based measurements that are far too sparse to calculate RZSM at global or even regional scales.

Proposed Solution

Phase 1: Machine Learning Gap-Filling

We apply temporal GNNs techniques to interpolate data gaps in daily soil moisture products from satellite missions like SMAP, creating continuous global coverage. This machine learning (ML) approach is particularly well-suited for hydrologic applications characterized by high spatial and temporal heterogeneity, as ML methods can better capture complex nonlinear relationships in heterogeneous systems compared to traditional process-based or physical models.

Phase 2: Physics-Based Data Assimilation

We employ these enhanced SSM datasets as initial conditions for estimating RZSM profiles using Reduced-Order Variational Data Assimilation. This phase leverages physics-based simulation to couple the assimilation of SSM data from the first phase, yielding more accurate estimates of soil moisture profiles at global scale.

Tools & Techniques

• Satellite data retrieval and processing from SMAP mission for global soil moisture monitoring
• Temporal Graph Neural Networks for spatio-temporal soil moisture gap-filling and interpolation
• Physics-based soil moisture modeling for root-zone estimation and profile simulation
• Reduced-Order Variational Data Assimilation for integrating satellite observations with physical models

Research Team & Collaborations