Improving actual evapotranspiration estimation under water stress by fusing solar-induced fluorescence, photochemical reflectance index, and interpretable deep learning models

Accurately estimating actual crop evapotranspiration (ET c act ) is critical for evaluating crop water use efficiency and simulating regional water–carbon cycles, particularly under the increasing frequency of climate change and drought events. Traditional meteorology-based ET c act estimation methods often fall short in capturing the physiological processes of photosynthesis and energy partitioning in vegetation. In contrast, two spectral indicators—solar-induced chlorophyll fluorescence (SIF) and photochemical reflectance index (PRI)—provide additional information related to crop light use efficiency and short-term stress responses. This study developed a multi-source driven framework for ET c act estimation by integrating meteorological and spectral data, based on continuous field observations of rainfed winter wheat from 2021 to 2024. Machine learning (ML) and deep learning (DL) models were used to evaluate the impact of multiple feature combinations, and Shapley additive explanation (SHAP) analysis was applied to interpret model behavior. The results showed that: (1) Correlation and maximal information coefficient analyses identified net radiation (Rn), SIF, and air temperature (Tair) as the dominant controls of ET c act variability, while the contribution of PRI became more evident under drought conditions; (2) Scenario S2 (Rn, SIF, PRI, and Tair) achieved the best balance between feature dimensionality and estimation accuracy; (3) The Temporal Convolutional Network (TCN) achieved the highest performance, with an average R 2 of 0.93 and RMSE and MAE values of 0.78 mm·day −1 and 0.88 mm·day −1 , respectively, surpassing traditional ML and Long Short-Term Memory models; (4) SHAP analysis revealed an increase in PRI’s contribution under drought, supporting its complementary role to SIF in representing water-stress-related physiological responses. By integrating meteorological and spectral information, this study proposes an ET c act estimation framework with enhanced physiological interpretability and highlights the synergistic value of combining SIF and PRI inputs. These findings provide technical pathways and theoretical support for crop water use efficiency monitoring, drought response analysis, and regional evapotranspiration modeling. • Developed a multi-source framework integrating remote sensing and meteorological data to estimate evapotranspiration. • Temporal convolutional network with four optimal inputs gave best accuracy (coefficient of determination 0.93). • Photochemical reflectance index contribution rose markedly under drought via energy dissipation regulation.