Performance Evolution of Yolo Models in Remote Sensing Images
Irfan Hassan, Xinyou Zhang
Abstract
Object detection (OD) in Remote Sensing (RS) satellite imagery presents unique challenges due to high variability and complex backgrounds. These complexities in RS data often lead to increased omission and false detection rates in object recognition. This paper presents a relative study of the performance of YOLO models (v5, v8, v9, v10, and 11) on two datasets: Airbus Aircraft and SkyFusion. Through a detailed ablation study, we analyze the impact of hyper-parameter tuning and dataset-specific, fine-tuning techniques for the performance of YOLO models and their variants. Experiments are conducted on datasets of varying sizes, including a smaller dataset of 103 images and a larger dataset of over 2995 images. The results highlight the improvements across YOLO versions in overcoming RS challenges, such as object sizes, high variability, and diverse contextual dependencies. This study emphasizes YOLO's adaptability towards RS tasks, which is highlighted in terms of fine-tuning strategies and performance optimization according to different scales of datasets.