EvapoDeep: A Dual Deep Learning Framework Utilizing GNSS Data for Evapotranspiration Modeling and Predictive Analysis

Modeling and predicting reference Evapotranspiration (ET₀) is crucial for effective water resource management and the protection of natural ecosystems. While the Penman–Monteith (PM) equation is widely recognized for its high accuracy in estimating ET₀, it requires extensive meteorological data, which is often unavailable at many synoptic stations. In contrast, the Thornthwaite (TH) model is simpler and more widely used but suffers from limited accuracy. Previous studies have demonstrated that calibrating the TH model with additional variables can improve its accuracy; however, these efforts have been largely preliminary. This study addresses the accuracy gap by developing EvapoDeep, an advanced model for ET₀ modeling and prediction that leverages deep learning techniques. Initially, the difference in ET₀ (DET₀) between the TH and PM models is calculated. This DET₀ is then modeled using an advanced Convolutional Neural Network-based deep learning method. The modeling component of EvapoDeep uses Global Navigation Satellite Systems (GNSS)-based Precipitable Water Vapor data from California, utilizing one year of GNSS measurements. In the final phase, the prediction component of EvapoDeep employs the generated ET₀ time series to predict ET₀ for the subsequent day using Long Short-Term Memory (LSTM) networks, demonstrating the model’s ability to handle temporal dependencies through its LSTM component. The EvapoDeep model significantly improved accuracy, achieving a root mean square error (RMSE) of 0.32 mm, compared to the TH model’s RMSE of 1.09 mm, representing a 70% improvement. In addition, the next-day ET₀ predictions yielded an RMSE of 0.29 mm, underscoring their high predictive performance.