Soil Physics-Informed Neural Networks to Estimate Bimodal Soil Hydraulic Properties

Abstract

Pedotransfer functions (PTFs) are widely used to estimate soil hydraulic properties (SHPs) from easily measurable soil properties. However, most existing PTFs are based on unimodal hydraulic models, which fail to capture the bimodal behavior of hydraulic properties caused by soil structure. In this study, we developed new PTFs using two bimodal hydraulic models and introduced a soil physics–informed neural network to embed these models into the training process. The results showed that the new PTFs effectively represented bimodality in hydraulic conductivity curves, achieving a root mean square error of 0.531(K in cm/d) on the test set, compared with 0.626 for unimodal models. They also improved predictions of soil water retention curves but struggled with bimodal cases for some samples, likely due to the limited number of bimodal retention curves in the training data set. Evaluation on an independent data set showed that the error for hydraulic conductivity predicted by the new functions was about one-third that of conventional approaches. In addition, the proposed soil physics–informed neural network, which directly optimizes SHPs, outperformed the conventional approaches that optimize fitted model parameters. We also found that whether water retention and hydraulic conductivity are optimized separately or simultaneously has a large effect on performance. Nonetheless, the lack of explicit soil structure information in the input data, along with limited measurements near saturation, continues to constrain accuracy. This emphasizes the need to develop a more comprehensive soil hydraulic database.