Abstract: In order to improve detection accuracy of illegal buildings and illegal occupation targets along river shorelines, based on the Drone-YOLO target detection baseline model, and by integrating a data enhancement module that simulates the river shoreline area characteristics, and by combining the BiFPN feature pyramid network structure, a Drone-YOLO-RCD instance segmentation algorithm based on multi-scale UAV image was designed and implemented. Then, a set of ground object variation detection system was constructed according to the algorithm, and a variation detection method of river shoreline features based on UAV panoramic image was formed. The experimental results on the self-made dataset PGIS_RCD showed that compared with the Drone-YOLO benchmark algorithm, the average accuracy (mAP @ 0.5) of the Drone-YOLO-RCD algorithm on six types of ground objects was improved by 4.4%, which realized the accurate identification of illegal buildings and illegal occupation problems, as well as the detection of ground object changes. The research results have promoted digital management on the ecological environment along river shorelines, and can provide strong technical support for scientifically grasping the range of construction activities along river shorelines.