A Multi-Class Object-Level Change Detection Method for Identifying Human Disturbance in Ecological Red Line Areas
The delineation of ecological red lines,which define areas where industrialization and urbanization are prohibited,holds great significance for environmental conservation.To ensure the protection of ecological red line areas,it is essential to identify human disturbances accurately.Traditional methods face challenges in pre-cisely detecting land cover changes caused by human interference and distinguishing the classes of objects before and after the changes.Segmatic change detection methods based on deep learning suffer from issues such as ex-cessive false positives and difficulties in obtaining training samples.To address these challenges,this paper pro-poses a multi-class object-level change detection method for precise identification of human disturbances within ecological red line areas.The proposed method consists of two parts:object-level binary change detection and scene classification.The object-level binary change detection network utilizes YOLOv5 as the underlying frame-work to extract features from the pre-and post-change images,fuse the features,and output the changed re-gions in the form of bounding boxes.The scene classification network,based on MobileNet v2,accurately clas-sifies the pre-change and post-change images corresponding to the changed regions.High-resolution satellite images from the Li River Ecological Protection Zone are used as the dataset to identify 27 types of human distur-bance activities.Experimental results demonstrate that the object-level change detection network achieves APIoU=.50 of 68.8％and APIoU=.50:.O5:.95 of 57.2％for change region extraction.The top-1 accuracy for human distur-bance activity recognition reaches 91.81％,and the top-5 accuracy reaches 99.83％.The results indicate that the two-step approach of object-level change detection and scene classification improves the effectiveness of change region extraction and overcomes the limitation of insufficient training samples for multi-class change de-tection.This approach provides effective support for the identification of human disturbance activities in ecologi-cal red line areas.
Ecological red lineHuman disturbanceDeep learningChange detectionMulti-classObject-lev-elScene classification
万方数据
维普
