Developing a physics-guided neural network to predict soil moisture with remote sensing evapotranspiration and weather forecasting

The short-term prediction of soil moisture variation is a decisive indicator of irrigation scheduling and crop management in agriculture. Traditional soil water dynamic models require complex descriptions of water movement and multiple parameters to calibrate for specific fields, which limit the model’s capability of generalization. Machine learning methods based on large sample datasets can automatically learn the most accurate way of predicting soil moisture with numerous related input variables. However, it could be time consuming in training and model optimization to improve performance. Combining the advantages of both methods, we designed a new soil moisture prediction neural network guided by the water transport driving mechanism. The water balance principle is used to limit the training process with remote sensing-based field-scale evapotranspiration, meteorological rainfall and primary soil water changes calculated from a simplified soil water model. By adding the physics layer to neural network, the demand for large datasets and the requirements of training and optimization are reduced. The prediction of soil moisture is at a half-monthly scale, and we tested the model during the winter wheat growing period. The results show that it requires less training capability to achieve high accuracy. Physics-guided neural networks could act as a better framework for parameter prediction in further researches.