Abstract
To address the problems of complex cloud features in satellite cloud maps, inaccurate typhoon localization, and poor target detection accuracy, this paper proposes a new typhoon localization algorithm, named TGE-YOLO. It is based on the YOLOv8n model with excellent high-low feature fusion capability and innovatively achieves the organic combination of feature fusion, computational efficiency, and localization accuracy. Firstly, the TFAM_Concat module is creatively designed in the neck network, which comprehensively utilizes the detailed information of shallow features and the semantic information of deeper features, enhancing the fusion ability of features at each layer. Secondly, the GSConv convolution is used to replace traditional convolution to reduce the computational cost of the model and effectively aggregate global information. Finally, the Efficient Intersection over Union (EIoU) loss function is improved to enhance its sensitivity to positional errors and optimize the error distribution, thereby improving the model's accuracy in capturing the position at the center of the typhoon. The experimental results show that the proposed TGE-YOLO model outperforms Faster R-CNN, YOLOv5s, YOLOv9s, and YOLOv11n, with the typhoon identification mean average precision (mAP) reaching 87.8%, the mean square error (MSE) of typhoon center localization at 0.115, and the detection speed at 416.7 frames per second (FPS). The model's rapid inference capability and efficient performance provide technical support for typhoon monitoring, which is expected to improve the timeliness and accuracy of typhoon warnings in practical applications.