Graph-Transformer for Spatiotemporal Soil Moisture Forecasting Using Multimodal Remote Sensing Data

Soil moisture is a key variable in the hydrological cycle, essential for drought monitoring, water resource management, and agriculture decision-making. However, accurate forecasting of soil moisture remains a challenge due to the nonlinear nature of soil water dynamics and the complex influencing factors. While deep learning models such as Long Short-Term Memory and Transformers have advanced temporal modelling, they struggle with spatial dependencies across diverse geographic locations, as their architecture primarily focuses on temporal dynamics. To address this limitation, we propose a novel hybrid framework that integrates graph neural networks with state-of-the-art time series models. We introduce a Graph-Transformer architecture for improved soil moisture forecasting by jointly modelling spatiotemporal dependencies. Our model forecasts soil moisture at two depths including 0–7 cm, referred to as surface soil moisture (SSM), and 7–28 cm, representing the root-zone soil moisture (RZSM). Evaluation using four years of multimodal data from 19 sites across Australia shows that the proposed Graph-Transformer model outperforms the standard Transformer, achieving improvements of up to 10% in the coefficient of determination (R2) and 8% in Root Mean Square Error (RMSE) for SSM, and up to 11% and 9% inR2and RMSE, respectively, for RZSM. These findings highlight that integrating graph architecture with temporal models can potentially enhance soil moisture forecasting across diverse geographic regions, benefiting effective drought monitoring and agricultural planning.