Enhancing subseasonal 2 m temperature forecast correction and downscaling over north China based on terrain-informed deep learning
Accurate high-resolution temperature predictions at the subseasonal-to-seasonal (S2S) timescale are vital for applications in agriculture, energy management, and disaster mitigation. However, current S2S forecasting capacity is hindered by declining predictive skill with increasing lead time and limited capability in capturing extreme events and fine-scale spatial variability, particularly over regions with complex terrain. To bridge this gap, this study develops a deep learning (DL)-based framework, termed Terrain-ResNet-U-Net (TRU-Net), for the downscaling (from 0.5 ∘ to 0.1 ∘ resolution) and bias correction of 2 m air temperature forecasts over North China. Unlike existing approaches that treat topographic information as auxiliary input, TRU-Net integrates terrain information through a dedicated feature extraction module and incorporates it into a U-Net architecture with a ResNet34 backbone and a spatial attention mechanism, enabling enhanced nonlinear feature learning and improved representation of terrain-induced spatial heterogeneity. Comparative experiments show that TRU-Net consistently outperforms the original forecasts for daily 2 m temperature prediction at the S2S timescale. Averaged across the six lead times, TRU-Net reduces the root mean squared error by 22.97% and increases the Pearson correlation coefficient and Accuracy by 25.16% and 24.48%, respectively, relative to the original forecasts. It also outperforms representative baseline methods, including Linear Regression, U-Net, and several DL baselines. The largest overall improvements are observed at lead times of 3–5 weeks, while the ablation results confirm the contributions of terrain information and the proposed multi-scale terrain feature extraction and fusion strategy. Similar advantages are maintained for representative high-impact events, where the proposed framework also provides positive correction and downscaling effects under these challenging conditions. These results indicate that an effective terrain-aware framework can substantially improve S2S temperature bias correction and downscaling over North China.