From ungauged to poorly gauged sites: A scalable deep learning framework for reference evapotranspiration forecasting through regional generalization and transfer learning

Short-term reference crop evapotranspiration (ET 0 ) forecasting is critical for prospective irrigation scheduling, but reliable local models are difficult to establish without long meteorological records. To address this problem, we developed regional generalized models based on four architectures: CNN, BiLSTM, CNN-BiLSTM, and CNN-BiLSTM with multi-head attention (CNN-BiLSTM-MHA) using observations from 15 meteorological stations in southeastern China. We evaluated the models under three data-availability scenarios. In the site-specific scenario, each model was trained and evaluated using data from the same station. In the ungauged scenario, the target station was excluded from regional model training. In the poorly gauged scenario, the regional model was fine-tuned using one year of observations from the target station. We found that ET 0 forecasting performance depended on matching model complexity to meteorological input richness. The input scheme that incorporated weather forecast information consistently produced the best overall performance. In this scheme, CNN-BiLSTM-MHA achieved the best performance for site-specific forecasting, with an average 1–7-day RMSE of 0.26 mm d −1 across 15 stations. For the ungauged scenario, the network-wide mean RMSE across the four architectures ranged from 0.54 to 0.63 mm d −1 . Fine-tuning with one year of local observations further improved performance at poorly gauged sites. CNN-BiLSTM-MHA produced the largest gains, reducing RMSE by 48.58% and 49.75% at the two target stations. The corresponding RMSEs were 0.34 and 0.29 mm d −1 . SHAP analysis further showed that air temperature, particularly forecast maximum temperature, was the most influential contributor to model predictions. Overall, this study established an ET 0 forecasting framework that adapts to local data availability. Across four deep learning architectures, we demonstrated the feasibility of applying regional generalized models to ungauged sites within the same hydrometeorological region and fine-tuning these models with one year of local observations at poorly gauged sites. This framework could extend irrigation scheduling and agricultural water allocation services to regions with uneven monitoring capacity.