Scholarlink

<p>Accurate prediction of cumulonimbus radar echo distribution remains challenging because most existing methods emphasize motion extrapolation or generic spatiotemporal learning while insufficiently representing the irregular boundaries and structural evolution of convective echoes. This limitation often leads to blurred boundaries, degraded structural consistency, and failure to capture nonlinear meteorological events. To address these issues, we propose CSDP-Net, a boundary-aware spatiotemporal prediction network that combines Cumulonimbus Boundary-Sensitive Convolution (CBS-Conv) with a Cumulonimbus Temporal-Spatial Feature Fusion (CTSFF) module. CBS-Conv aligns with the morphology-aware characteristics of cumulonimbus-related echoes, and CTSFF refines deep semantic information to enhance prediction accuracy. Extensive experiments demonstrate the effectiveness of CSDP-Net. On the GCAPPI-2.0 dataset, CSDP-Net achieves the lowest Mean Squared Error (MSE) while maintaining a highly competitive Structural Similarity Index (SSIM). Competitive performance is also observed on the Moving MNIST benchmark. Ablation studies further validate the efficacy of the proposed modules, underscoring their complementary roles in local boundary representation and global structural consistency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Morphology–aware spatiotemporal prediction of cumulonimbus–related radar echoes using CSDP–Net

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Deyi Wang,
  • Xu Wang,
  • Zhenhong Cheng,
  • Xin Chang,
  • Hongping Yuan,
  • Jixiang Yang,
  • Yi Liang,
  • Jiaming Hu,
  • Ting Zhang

摘要

Accurate prediction of cumulonimbus radar echo distribution remains challenging because most existing methods emphasize motion extrapolation or generic spatiotemporal learning while insufficiently representing the irregular boundaries and structural evolution of convective echoes. This limitation often leads to blurred boundaries, degraded structural consistency, and failure to capture nonlinear meteorological events. To address these issues, we propose CSDP-Net, a boundary-aware spatiotemporal prediction network that combines Cumulonimbus Boundary-Sensitive Convolution (CBS-Conv) with a Cumulonimbus Temporal-Spatial Feature Fusion (CTSFF) module. CBS-Conv aligns with the morphology-aware characteristics of cumulonimbus-related echoes, and CTSFF refines deep semantic information to enhance prediction accuracy. Extensive experiments demonstrate the effectiveness of CSDP-Net. On the GCAPPI-2.0 dataset, CSDP-Net achieves the lowest Mean Squared Error (MSE) while maintaining a highly competitive Structural Similarity Index (SSIM). Competitive performance is also observed on the Moving MNIST benchmark. Ablation studies further validate the efficacy of the proposed modules, underscoring their complementary roles in local boundary representation and global structural consistency.