Science-Guided Multi-Task Deep Learning for Emulating APSIM Simulations for Root-zone Soil Moisture Forecasting

Abstract

Soil moisture is a critical variable for water management, drought monitoring, and crop risk management, yet direct measurements below the surface are sparse and difficult to scale. We present a deep learning framework that forecasts root-zone soil moisture generated with a process-based agroecosystem simulation model, APSIM. Our approach forecasts multi-layer soil moisture profiles down to 1m using static site descriptors and daily meteorological forcings. Our approach unifies (i) large-scale physics-based simulation data generation over diverse counterfactual soil properties, irrigation strategies, and weather information, (ii) a controlled benchmark spanning Temporal Convolutional Networks, and Mamba-style state space models, and (iii) a multi-task training objective that predicts both absolute moisture levels and step-wise changes (deltas). The deltas formulation anchors forecasts to the last observed state and focuses learning on day-to-day process rates, improving stability across depths and forecast steps. Experiments on a large APSIM-derived dataset with 11 depth layers evaluate accuracy under a standard held-out test split, spatial generalization to unseen stations, and temporal generalization to a future year. Across architectures, delta-aware training consistently improves forecasting performance relative to direct level prediction and simple baselines, with the strongest gains appearing under distribution shift.