Linear Variable Convolution and Cross Attention LSTM Method for Radar Echo Extrapolation

Radar echo extrapolation is a critical component of short-term weather forecasting. To enhance the accuracy and adaptability of radar echo prediction, a novel method named LVC-LSTM is proposed, which integrates a Linear Variable Convolution (LVC) module and a Cross-Attention (CA) mechanism into the Spatiotemporal LSTM (ST-LSTM) framework. The LVC module enables dynamic adjustment of the convolutional sampling shape, allowing the network to better capture the irregular and evolving structures of radar echoes, thereby offering improved flexibility and representation capability compared to traditional and deformable convolutions. The CA mechanism introduces a pyramid-based cross-attention structure, incorporating cross-scale embeddings and hierarchical attention blocks, which effectively capture multi-scale features and long- and short-range spatial dependencies inherent in radar echo dynamics. Experimental results on a real-world radar echo dataset demonstrate that the LVC-LSTM model outperforms comparative models across multiple evaluation metrics, including the Critical Success Index (CSI), Heidke Skill Score (HSS), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Structural Similarity Index (SSIM), indicating its strong potential for operational radar echo extrapolation tasks.