Improved estimation of daily surface all-wave net radiation from FY3D MERSI-II data based on Pseudo-label learning

Surface all-wave net radiation (Rn) plays a pivotal role in land–atmosphere energy exchange and redistribution, and modulates global water and heat balance and energy circulation. Although direct estimation of Rn from satellite top-of-atmosphere (TOA) observations has shown strong potential, existing methods have largely relied on single-sensor observations and sufficient training samples. This limitation has become increasingly critical as MODIS approaches the end of its mission, while the expected replacement by FY-3D MERSI-II (hereafter MERSI) still lacks enough samples for stable global model development. To address this issue, we first proposed a global Rn estimation framework based on the length ratio of daytime (LRD) for sensors providing TOA observations from the visible to thermal infrared bands. Within this framework, MODIS-only and MERSI-only models were then built separately for daily net radiation (Rn_daily) estimation; however, the MERSI-only model showed higher uncertainty due to the limited number of training samples available over a shorter period. To improve the MERSI-only model, we used the MODIS-only model to generate pseudo-labels for numerous unlabeled MERSI pseudo-site samples and combined these pseudo-labeled samples with limited ground-labeled MERSI samples to train a new model using a newly proposed progressive loss-weighting strategy, namely Dynamic Weight Adjustment Training (DWAT). Validation against Rn_daily measurements showed that the obtained DWAT model outperformed the MERSI-only model in terms of generalization, robustness, and estimation accuracy, yielding the overall validation and independent-validation root-mean-square-error (RMSE) values of 21.68 and 23.95 W/m², respectively. The largest improvement was found at high latitudes, with the independent-validation RMSE reduced by up to 3.36 W/m². Moreover, the DWAT model demonstrated better predictive accuracy and strong spatial mapping ability. Overall, this approach provided a practical way to generate accurate and continuous global Rn datasets from different sensors and offered a scalable solution for cross-sensor applications in remote sensing.