Author(s): Jeongwoo Han; Tae-Woong Kim
Linked Author(s): Tae-Woong Kim
Keywords: Soil moisture drought Physics-based machine learning Neural ordinary differential equation
Abstract: Under global warming, an increase in the intermittent period of precipitation along with a raised sensible heat has provided a hydrological background that facilitates the evolution of meteorological drought into soil moisture and/or hydrological droughts. Soil moisture, as a key state variable of the hydrologic circulation, plays a significant role in drought propagation, crop growth, and securing environmental diversity. Thus, understanding the temporal behavior of soil moisture and forecasting soil moisture drought are crucial to be prepared for the anticipated droughts and their transition. To forecast monthly soil moisture drought for an extended period, this study developed the Neural ordinary differential equation (NeuralODE), which falls under the umbrella of physics-based machine learning. NeuralODE learns both the mechanistic process (herein, Thornthwaite water balance model) and data features, improving forecast accuracy and interpretability of a deep learning model. NeuralODE trains neural networks to solve the water balance model, based on the initial value problem framework, while minimizing a loss function of input-output relations. Since NeuralODE couples differentiable mechanistic processes (i. e., water balance model) into the backpropagation algorithm, end-to-end differentiable physics-informed learning can be feasible without external coupling of the process-based model. For this study, to forecast soil moisture drought, NeuralODE was trained locally (i. e., for each watershed of CAMELS dataset) across Texas, U. S. A. and showed a median of correlation coefficient (CC) and Nash-Sutcliffe efficiency (NSE) greater than 0.9 and 0.8, respectively, at a 3-month lead time or greater. Besides, when NeuralODE was compared with Long short-term memory (LSTM) and convolutional LSTM (ConvLSTM), NeuralODE outperformed the LSTM-based models in forecasting soil moisture drought. The good forecast accuracy of NeuralODE substantiates its beneficial application to proactive agricultural drought mitigation measures.
Year: 2025