Enhanced daily retrieval of Evapotranspiration from a Transformer-based deep learning model combined with satellite passive microwave and optical observations

Evapotranspiration (ET) is a key component influencing Earth’s hydrological cycle and energy balance. Although the integration of deep learning models with multi-source remote sensing data represents a cutting-edge approach for quantitative ET retrieval from end-to-end, few studies have developed a daily-scale deep-learning model driven simultaneously by satellite multi-channel passive microwave and optical observations. This study proposed a novel lightweight model to achieve daily ET estimations, which combined convolutional neural networks (CNNs) for local temporal feature extractions and Transformer’s self-attention mechanism for long-term temporal dependencies. We employed Chinese Fengyun-3D satellite passive microwave Emissivity Difference Vegetation Indices (EDVIs) at X-, Ku-, and Ka- bands as the indicator of hierarchical vegetation water content, and MODIS optical Enhanced Vegetation Index (EVI) as proxy of greenness foliage, which offered distinct and complementary vegetation characteristics for the model learning. The hybrid CNN-Transformer model was trained and tested over 108 global forest, cropland, wetland, savanna and grassland flux sites, and compared with three widely-used deep learning models (Transformer, CNN-LSTM, and LSTM). Results indicated that the CNN-Transformer model achieved daily ET estimation with an overall root mean square error (RMSE) < 35% and bias < 5%, particularly in the Northern Hemisphere (R 2 = 0.86, RMSE = 0.52 mm day⁻ 1 or 34%). The model presented a more robust capability of generalization (R 2 = 0.71, RMSE = 0.58 mm day⁻ 1 or 35%) than other deep learning models (R 2 = 0.59–0.66; RMSE = 0.62–0.70 mm day⁻ 1 or 38–42%) in the Southern Hemisphere. Notable enhancement of our model presented during April-May transient periods and mid-growing seasons when vegetation states and meteorological conditions changed rapidly. The synthetic use of satellite microwave and optical vegetation indices provided additional gains for the CNN-Transformer model, thereby enhancing daily ET estimation. The spatiotemporal pattern of our model agreed closely with that of FLUXCOM-X ET over the Northern Hemisphere. These findings underscore the value of the hybrid model consisting of temporal convolution-attention mechanisms together with satellite multi-source vegetation data for robust daily ET retrievals, holding a great potential for global applications.