Comparative Analysis of YOLO, Faster-RCNN and RetinaNet Object Detection Models for Satellite Imagery Analysis

This study delves into the comparative analysis of two leading object detection models, YOLO, Faster-RCNN and RetinaNet, within the realm of satellite imagery analysis. Utilizing the “Ships/Vessels in Aerial Images” dataset from Kaggle, featuring satellite imagery of marine vessels, the research investigates model performance and trade-offs. The dataset comprises images with a characteristic for resolution of 768 × 768 pixels and 3 color channels, later augmented to 416 × 416 pixels. The evaluation includes metrics such as mean Average Precision (mAP), Intersection Over Union (IoU), and processing times. YOLO achieved an IoU of 0.47 and mAP of 0.399, with preprocessing, inference, and post-processing times recorded at 0.4ms, 33.4ms, and 1.6ms, respectively. The Faster-RCNN model attained an IoU of 0.679, F1-Score of 0.543, and mAP50-95 scores of 0.335, 0.560, and 0.680 for small, medium, and large objects, respectively. Conversely, RetinaNet model attained an IoU of 0.971, F1-Score of 0.889, and mAP50-95 scores of 0.044, 0.399, and 0.647 for small, medium, and large objects, respectively. To summarize, this research offers significant perspectives on the characteristics of performance and limitations of YOLO, Faster-RCNN and RetinaNet models in analyzing satellite imagery, laying the groundwork for future developments in computer vision research.