A Comparative Analysis of YOLOv8 and YOLOv11 for Tropical Cyclone Detection: The Impact of Data Augmentation on Model Performance

Tropical cyclones, also known as TCs, are a category of severe weather events considered to have a significant weather impact, making it extremely important to detect them accurately. Despite the advantages presented by deep-learning techniques compared to the usual forecasting approaches, the use of the best model is still a concern. This study is a comparative analysis of the performance between two state-of-the-art, light-weight models, namely YOLOv8n and YOLOv11n, on the detection task of tropical cyclones. Four experiments of the above-mentioned models were conducted on the INCYDE satellite images, using a normal unaugment dataset compared to an augmented normal dataset, applying standard geometric transformation techniques, namely flipping and rotation. The experimental results demonstrate a significant improvement on the YOLOv11n model, using the unaugment dataset, resulting in a mAP@.50–.95 score of 0.921 and a detection rate of 0.994. Unexpectedly, the standard augmentation approach deteriorated the performance on both architectures.