Less Is More: Online Spatio-Temporal Selective Learning for Multi-Variable Meteorological Forecasting

Accurate forecasting of high-dimensional meteorological fields remains challenging due to the complex spatio-temporal dynamics of atmospheric systems and the presence of heterogeneous training difficulty across space and lead time. Existing deep forecasting approaches usually optimize all prediction units uniformly, which may overemphasize low-benefit or weakly generalizable supervision signals. To address this issue, we propose Spatio-Temporal Selective Learning (ST-SL), an online training framework that estimates the learnability of each prediction unit by comparing the main model with a frozen reference model and computes the loss only over selected high-benefit spatio-temporal units. To provide an effective forecasting backbone, we further introduce VASTFormer, a variable-aware spatio-temporal Transformer that models cross-variable dependencies, incorporates physics-enhanced Solar Positional Encoding, and captures atmospheric trajectories with an efficient temporal translator. Experiments on the ERA5 reanalysis dataset show that VASTFormer outperforms representative spatio-temporal baselines, while ST-SL further improves accuracy without adding inference-time parameters or computational cost. Compared with the strongest baseline, VASTFormer+ST-SL reduces MSE, MAE, and RMSE by 8.84%, 6.70%, and 4.54%, respectively. Meteorological skill evaluation further shows an average ACC of 0.9801 and RMSESS of 0.8104, and percentile-based extreme-condition evaluations confirm consistent improvements across standard and high-impact forecasting scenarios. These results indicate that selective supervision can improve generalization in dense meteorological forecasting.