Improving soil surface evaporation estimates with transformer-based model

Abstract

Soil surface evaporation (E) is an important component of evapotranspiration from barren or sparsely vegetated (BSV) areas, and accurately estimating E in areas with limited water resources remains challenging due to the complexity of influencing factors. In this study, a large number of global ground-based measurements from bare soil conditions were collected, and a transformer-based model based on transformer architecture was developed to estimate E. The estimated instantaneous E achieved an R value of 0.73–0.96 and an RMSE of 0.03–0.05 mm/h, outperforming the process-based model, in which the surface evaporation resistance is considered as a function of soil moisture in exponential form or power form. The RMSE value of estimated E was low when the soil was relatively dry, indicating that the model is suited for water-limited regions. Furthermore, the transformer-based model was applied to BSV regions in Northwestern China, producing spatial patterns that were not only reasonable but also more detailed and consistent with river distributions. Compared to the other two products (GLEAM and BESS), the spatial annual mean E from our model and BESS were similar, while GLEAM's result was significantly lower, particularly in summer. Our findings suggest that applying deep learning to E simulation can greatly improve the accuracy and help overcome current challenges related to unclear mechanisms and the lack of universal modeling approaches.