Field-Level Agricultural Disaster Warning and Loss Assessment Based on Multi-Spatio-Temporal Fusion Network

Extreme weather poses a severe challenge to global food security, and timely, accurate agricultural disaster warnings can effectively mitigate crop yield losses. Traditional agricultural disaster warning models often rely on sparse meteorological station observations, making it difficult to capture micro-meteorological variations at the field level. Moreover, they frequently overlook the decisive role of crop stress tolerance in disaster occurrence. This study proposes a multi-spatio-temporal fusion network (MSTF-Net) featuring a unique dual-tower architecture. One tower utilizes high-dimensional remote sensing features to capture crop growth status and micro-topography on the forecast date, while the other tower employs long short term memory (LSTM) to process “past 30 days + future 7 days” meteorological time-series data, simulating the dynamic evolution of environmental stress. Within a unified framework using Sentinel-2 imagery, this approach simultaneously achieves field-level disaster warning, cause attribution, and loss assessment to mitigate yield losses from disasters. Results demonstrate that MSTF-Net achieves 12% higher accuracy compared to LSTM models using only meteorological data and multilayer perceptron (MLP) models using only remote sensing data. The model maintains high-precision early warnings (AUC > 0.85) and loss assessments within a 7-day window, meeting crop growth scheduling needs. In summary, the proposed MSTF-Net model delivers effective field-level agricultural disaster warnings, offering a feasible pathway to mitigate agricultural disaster losses.