Evaluating YOLOv8 Variants for Object Detection in Satellite Image

Object detection in Remote sensing images enables crucial functionalities in various fields, including agriculture, environmental monitoring, urban planning, and disaster management. While traditional methods face challenges in speed and accuracy, deep learning has emerged as a powerful solution. This paper compares YOLOv8 variants (nano, small, medium, and large) for ship detection in high-resolution satellite images. The YOLOv8 model is a one-stage object detection that utilizes a Cross Stage Partial Darknet-53 (CSPDarknet53) backbone for feature extraction. A Path Aggregation Network - Feature Pyramid Network (PAN-FPN) neck for multi-scale feature fusion, and a decoupled head for final predictions. Its speed and accuracy can be adjusted by selecting from different model variants, which vary in size and complexity. The Ship dataset was chosen due to its challenging characteristics, including high-resolution imagery (30–50 cm) from Google Earth, diverse viewpoint variations, occlusions, cloud cover, shadows, varied lighting conditions, and cluttered marine backgrounds. mean average precision (mAP), recall, precision, F1-score, training time, and model size were utilized in evaluating YOLOv8 variants. The results demonstrate that YOLOv8s achieved the best balance between accuracy and efficiency, with an F1-score of 96.3% and a mAP50-95 of 70.4%. Although YOLOv8n demonstrates the highest processing speed, its detection performance is marginally inferior. In contrast, Larger models (YOLOv8m and YOLOv8l) do not show significant improvements in accuracy despite increased computational cost. The results provide insights into the effectiveness of each model for ship detection, enabling decisions for selecting the optimal model based on the balance between accuracy, speed, and resource utilization.