STADEN: A Dual‐Branch Deep Learning Framework With Multithreshold Loss for Robust Radar Echo Extrapolation

Radar echo extrapolation is a key method for weather nowcasting. Deep learning has become a hotspot of radar echo extrapolation, but it suffers from problems such as prediction fuzziness and low prediction metrics, mainly due to two reasons: firstly, existing deep learning methods still have insufficient capabilities in learning spatial–temporal features and focus on critical regions. Second, atmospheric motion is chaotic and uncertain, and only the confidence interval of the future atmospheric state can be predicted. Existing pixel‐level loss functions such as MSE loss require the predicted values to be completely consistent with the gold standard values numerically, resulting in the model being unable to converge to the optimal solution. To address these issues, we propose the Spatial–Temporal‐Attention Dual‐path Extrapolation Network (STADEN) model and a novel multithreshold loss function (MTL). STADEN introduced a dual‐branch network based on the inception and metaformer to enhance the ability to learn spatial–temporal information. MTL allows predicted values to fall within a ground‐truth interval to achieve the minimum loss value, enhancing model robustness. In the 2‐h extrapolation experiment on the Sichuan dataset, STADEN achieved critical success index (CSI) metrics of 0.488, 0.388, and 0.211 at thresholds of 15, 25, and 35, respectively, outperforming all comparative methods. MTL were applied to all comparative methods directly, and improve CSI metrics by 0.078, 0.058, and 0.024. Robustness experiments on the HKO‐7 dataset further demonstrate that both STADEN and MTL achieve optimal performance. These results highlight the potential of STADEN and MTL for practical applications in radar echo prediction.