Typhoon localization detection algorithm based on TGE-YOLO.

Area of Science:

• Meteorology and Atmospheric Science
• Computer Vision and Machine Learning
• Remote Sensing Technology

Background:

• Satellite cloud maps present challenges in accurately identifying complex cloud features, leading to imprecise typhoon localization and detection.
• Existing algorithms struggle to balance feature fusion, computational efficiency, and precise localization for typhoon analysis.

Purpose of the Study:

• To develop an advanced typhoon localization algorithm, TGE-YOLO, that overcomes limitations in current satellite-based typhoon identification methods.
• To enhance the accuracy of typhoon center localization and improve overall target detection performance using deep learning.

Main Methods:

• The TGE-YOLO algorithm is built upon the YOLOv8n model, incorporating a novel TFAM_Concat module in the neck network for superior feature fusion.
• It utilizes GSConv convolution to reduce computational load and improve global information aggregation, alongside an improved Efficient Intersection over Union (EIoU) loss function for enhanced positional accuracy.

Main Results:

• TGE-YOLO achieved a typhoon identification mean average precision (mAP) of 87.8%, outperforming Faster R-CNN, YOLOv5s, YOLOv9s, and YOLOv11n.
• The model demonstrated a mean square error (MSE) of 0.115 for typhoon center localization and a detection speed of 416.7 frames per second (FPS).

Conclusions:

• The TGE-YOLO model offers a significant advancement in typhoon localization and detection accuracy and speed.
• Its efficient performance provides crucial technical support for real-time typhoon monitoring, promising more timely and accurate typhoon warnings.