基于对比自监督学习的热带气旋中心定位方法

Contrastive Self-Supervised Learning-Based Tropical Cyclone Center Localization Method

摘要: 热带气旋中心定位是台风路径预测和灾害预报的重要步骤。现有的热带气旋中心定位模型主要基于监督学习，数据标注成本高且未充分利用热带气旋在卫星云图上显著的结构特征。为此，本文提出一种基于对比自监督学习的热带气旋中心定位模型(SSLTCL)，首先利用大量无标签的热带气旋样本对比学习气旋特征，再利用少量高精度的热带气旋标注样本回归预测台风中心位置，协同训练气旋特征提取模型和气旋中心回归模型。试验结果表明，SSLTCL能有效地定位不同强度等级的台风中心，平均绝对误差(MAE)为0.210˚，优于其他主流模型，且定位精度和召回率随台风强度等级的增大而增大。此外，模型可为台风预报提供台风中心定位支持，并对热带气旋进行有效检测，检测准确率、召回率均达97%以上。

Abstract: Tropical Cyclone (TC) center localization is a crucial step in typhoon track prediction and disaster forecasting. Existing tropical cyclone center localization models are primarily based on supervised learning, which involves high data annotation costs and fails to fully utilize the prominent structural features of tropical cyclones in satellite cloud imagery. To address this, this paper proposes a contrastive Self-Supervised Learning-based Tropical Cyclone Localization (SSLTCL) model. The SSLTCL framework first employs contrastive self-supervised learning to extract cyclone features from a large volume of unlabeled tropical cyclone samples. Then, it fine-tunes the model using a small set of high-precision labeled samples to regress and predict the typhoon center location. This approach co-trains the cyclone feature extraction model and the center regression model in a synergistic manner. Experimental results demonstrate that SSLTCL can effectively locate the centers of typhoons with different intensity levels, achieving a Mean Absolute Error (MAE) of 0.210˚, outperforming other mainstream models. Notably, the localization accuracy and recall rate improve as typhoon intensity increases. Additionally, the model provides real-time tropical cyclone center localization support for typhoon forecasting and achieves robust cyclone detection, with both precision and recall rates exceeding 97%.