A Transfer Learning Approach for Estimating All-Weather Daily Net Radiation over the Tibetan Plateau: Site-Scale Evaluation and Spatial Extension

Accurate estimation of daily net radiation (Rn_daily) at high spatial resolution (1 km) over the Tibetan Plateau (TP) is crucial for understanding land surface energy budgets and climate dynamics. This study proposes a densely connected multilayer perceptron (DenseMLP)-based transfer learning framework, with a two-stage strategy (coarse pre-training on GLASS Rn_daily, followed by fine-tuning on limited TP ground observations) using MODIS land surface parameters and auxiliary data to generate 1 km Rn_daily. When evaluated on the training set, the proposed model achieves an overall R2 of 0.87, MAE of 16.06 W m−2, RMSE of 21.94 W m−2, and a near-zero bias of 0.07 W m−2. On an independent test set, the model maintains robust performance with R2 = 0.83, MAE = 17.43 W m−2, RMSE = 22.55 W m−2, and bias = −1.12 W m−2. The method exhibits consistently low bias across individual sites (mostly within ±3.7 W m−2) and accurately captures seasonal variability. When applied to the entire TP for 2018, the 1 km Rn_daily product reveals clear aspect-related terrain effects and a distinct annual cycle. This framework effectively mitigates site-dependent errors, providing a useful reference for long-term Rn product development over the TP.