Extrapolability improvement of machine learning-based evapotranspiration models via domain-adversarial neural networks

Abstract

Machine learning-based evapotranspiration (ET) models capture complex nonlinear relationships but struggle with global extrapolation due to unbalanced data distribution, limiting accurate ET assessments crucial for understanding water and energy cycles. This study used Domain-Adversarial Neural Networks (DANN) to improve the geographical adaptability of ET models by mitigating site-level distributional discrepancies. DANN significantly enhanced ET prediction accuracy, with a median Kling-Gupta Efficiency (KGE) increase of 0.27 (p < 0.001) and with a range from 0.06 to 0.58 for the middle 90% values compared to the traditional Leave-One-Out (LOO) method. DANN proves particularly effective for isolated sites and biome transition zones, reducing errors and avoiding low-accuracy predictions. By leveraging data from resource-rich areas, DANN strengthens the reliability of global-scale ET products, especially in ungauged regions. Future evaluations and improvements are necessary, such as using additional accuracy metrics beyond KGE and focusing on sites located at the intersection of several climate types and sites with unique soil-vegetation-atmosphere processes. This study demonstrates the potential of domain adaptation techniques to enhance the generalization and extrapolation capabilities of machine learning in hydrological predictions.