Aerial Image Segmentation Using YOLO v8 and SAM Pipeline

Aerial image segmentation, important for understanding image content, faces significant challenges such as feature scale variation, resolution issues, spectral inconsistencies, and class imbalance. This paper introduces a robust pipeline that combines YOLOv8 for object detection and the Segment Anything Model (SAM) for refining segmentation. The study addresses detection attributes, augmentation methods, and dataset creation for reliable modelling. Trained on a custom sub-meter aerial image dataset, the model achieves moderate accuracy with MAP@50 of 0.68 and MAP@50-95 of 0.51. It demonstrates strong potential for applications in urban planning, disaster management, and geospatial analysis. By enhancing segmentation accuracy and efficiency through cutting-edge deep learning techniques, this method supports critical geospatial applications.