A self-supervised deep learning model for enhanced generalization in soil moisture prediction

• A self-supervised Transformer model is developed for soil moisture prediction. • The model achieves enhanced generalization over extreme and unseen conditions. • Latent representations of the model align with underlying soil hydrodynamics. Machine learning models are increasingly developed for soil moisture prediction, but their generalization under diverse and unobserved conditions remains a critical challenge. Existing models often fail in scenarios involving unseen meteorological conditions or soil depths. To address this limitation, we propose a novel approach that combines self-supervised learning with a physics-guided spatio-temporal Transformer architecture to improve generalization. We evaluate our model through synthetic experiments and benchmark its performance against the state-of-the-art Long Short-Term Memory (LSTM) model. Results show that our model consistently outperforms LSTM in capturing extreme soil dryness, adapting to unseen meteorological conditions, and predicting moisture dynamics at unobserved depths. The improved generalization is attributed to the model’s ability to capture spatio-temporal moisture dynamics and reveal hidden patterns by encoding past observations into high-dimensional latent representations. In addition, the model adapts well to complex scenarios, such as extracting infiltration characteristics from datasets without heavy rainfall events and accounting for interactions across soil depths under varying meteorological conditions. Overall, it is demonstrated that self-supervised learning with Transformer-based architectures significantly improves the deep learning of soil moisture dynamics and enhances model generalization. This study provides a robust solution to the challenges posed by variable and complex environmental conditions and represents a promising data-driven approach to facilitating process understanding.