Improving Land Cover Change Detection Using YOLO Segmentation and Comparative Polygon Analysis

Abstract

This research develops a methodology for analyzing changes in Earth's surface and atmosphere using remote sensing imagery from satellites, aircraft, and drones. The research shows a traditional labeling techniques with refined segmentation specifically tailored for the Land Use Dataset, thereby improving the accuracy of land-cover segmentation predictions. We assessed the performance of various YOLO models—YOLOv5, YOLOv7, and YOLOv8—in our detection framework, with YOLOv8 proving to be the most effective. The YOLOv8 model achieved a mean Average Precision (mAP50) of 42.7% and a Boundary Box Precision of 55.9%.