CSDP-Net: An Accurate Approach for Cumulonimbus Spatiotemporal Distribution Prediction

Abstract

Accurate prediction of Cumulonimbus distribution based on radar echo data is crucial for advancing weather forecasting. Existing studies often focused on tracking Cumulonimbus movement, overlooking the inherent feature priors within the data. To address this limitation, this paper proposes the Cumulonimbus Spatiotemporal Distribution Prediction Net (CSDP-Net), which is designed to predict Cumulonimbus distribution by integrating both temporal and spatial features while explicitly embedding shape priors of Cumulonimbus. CSDP-Net introduces two key innovations: (1) Cumulonimbus Boundary-Sensitive Convolution, which aligns with the morphological characteristics of Cumulonimbus, and (2) a Cumulonimbus Spatiotemporal Feature Fusion Module, which refines semantic information to enhance prediction accuracy. Extensive experiments demonstrate the superior performance of CSDP-Net. On the Moving MNIST dataset, CSDP-Net achieves an average improvement of 80.30% in Mean Squared Error (MSE) and a 2.25% gain in Structural Similarity Index (SSIM) compared to state-of-the-art methods. On the GCAPPI-2.0 dataset, CSDP-Net outperforms existing approaches with a 70.27% reduction in MSE and a 1.97% improvement in SSIM. Ablation studies further validate the efficacy of the proposed modules, underscoring their contribution to the model's robustness and precision.