An Object-Oriented Classification Method for Detectdion of Large Wild Herbivores:a Case Study in the Source Region of Three Rivers in Qinghai

Luo Wei;Shao Quanqin;Wang Dongliang;Wang Yangchun;Key Laboratory of Land Surface Pattern and Simulation(IGSNRR) ,Chinese Academy of Sciences(CAS);Institute of Mountain Hazards and Environments(CAS);

Wild large herbivores;;UAV image;;Multi-resolution segmentation;;Template matching;;Fuzzy logical classification

The source region of the Three Rivers in Qinghai is in western China in the hinterland of the Qinghai Tibet Plateau in southern Qinghai Province. Elevation ranges from 3 500 to 4 800 meters and climate conditions are harsh. In this region,it is difficult to use traditional wildlife monitoring methods. Based on aerial images acquired by unmanned aerial vehicles( UAVs) in this region in July 2016,we used object-oriented image analysis to automatically extract information on large wild herbivores and we analyzed quantitative statistics. First,the contours of the animals in the images were roughly separated from the background using a multi-scale segmentation technique.Next,typical samples of the target animals were selected to generate matching templates to classify and detect the segmentation results and to locate suspected target objects. Then,the spectral features and shape features of the target animal objects in the images were extracted to construct a feature knowledge base and to screen the detection results. Finally, we evaluated accuracy by combining the number of animals extracted from the statistics with those extracted from the results of visual interpretation. This method had both high extraction speed and high precision. The method is expected to significantly reduce or even replace part of the ground survey for wildlife,and improve the efficiency and accuracy of wildlife survey.