SoilCast: An Interactive Machine-Learning System for Forecasting Soil Health under Climate and Management Scenarios

Abstract

We present SoilCast, an interactive machine-learning system for real-time exploration of long-term soil health outcomes under alternative climate and agricultural management scenarios across Europe. SoilCast replaces computationally expensive process-based environmental simulations with pretrained surrogate models learned from more than 15 million simulator-generated samples. The system employs CatBoost-based state-transition emulators that predict changes in soil organic carbon, soil nitrogen, and crop productivity over successive 20-year periods, enabling recursive multi-step forecasts through 2100. To support reliable scenario analysis, SoilCast visualizes uncertainty intervals derived from empirical rollout errors on held-out spatial test regions, capturing the accumulation of forecast error over time. The application demonstrates how scalable surrogate modeling can transform complex scientific simulators into responsive decision-support systems for sustainable land management.