A comparative analysis of deep learning models for accurate spatio-temporal soil moisture prediction

Soil moisture (SM) is essential for energy and water exchange between soil and atmosphere. Accurate prediction of its spatio-temporal occurrence is critical for climate, hydrology, and agriculture. This study fine-tunes and evaluates state-of-the-art deep learning models for spatio-temporal SM prediction in the North China Plain, including Convolutional Long Short-Term Memory (ConvLSTM), Memory in Memory (MIM), Predictive Recurrent Neural Network (PredRNN), and Cubic Recurrent Neural Network (CubicRNN). The models were trained using the ERA5 climate reanalysis and the CN05.1 meteorological datasets. The performance was analyzed by different accuracy metrics. The results show that ConvLSTM has the lowest errors in both time series and spatio-temporal prediction. However, the prediction performance is related to the magnitude of SM. All models demonstrated increased prediction errors as SM values exceed 75%, suggesting limitations in the models’ ability to predict extreme events. This study underscores the significance of choosing suitable deep learning models and suggests exploring ensemble approaches in the future.