Comparison On Methods Of Extracting Rodent Damage Information In Zoige Grassland Based On Low-altitude Remote Sensing

It has a serious impact on life,production and ecological environment that rodent damage occurs frequently in Zoige grassland.In this paper,the method of extracting grassland rodent damage information based on UAV image is studied to provide some theoretical and technical references for grassland rodent damage monitoring.It is of great significance for the control of rodent damage and sustainable development of grassland resources in Zoige grassland.In this study,two small pastures with serious rodent damage were used as experimental areas in Xiaman Township of Zoige County.Firstly,aerial photographs of the experimental area were obtained by the DJI 4Pro UAV.And four low-altitude remote sensing orthophoto images were producted through Pix4 D Mapper processed aerial photographs.They include images of zokor damage areas in spring,zokor damage areas in summer,marmot damage areas in spring and marmot damage areas in summer.Secondly,four methods,gray threshold segmentation,optimal color texture,rule-based object-oriented,and BP neural network,were used to extract rodent damage information from four images.Then,the accuracy of extraction results was evaluated from two aspects: spatial accuracy and quantitative accuracy.Finally,the advantages of each method were summarized through comparative analysis.And the best extraction method for different rodent damage information in different season images was obtained.The main conclusions of this study are as follows:(1)The image of zokor damage area in spring: The classification accuracies of optimal color texture,rule-based object-oriented,and BP neural network nethods is high,and the producer accuracy and user accuracy are more than 90%.There are some omissions in extracting the center part of the mound by gray threshold segmentation.And the producer accuracy is only 73.03%.The accuracy of the rule-based object-oriented extraction results is 96.30%,the area deviation is 0.16,and the shape deviation is 0.13.It is the most accurate method for extracting zokor mound in spring images.(2)The image of zokor damage area in summer: The extraction results of gray threshold segmentation and object-oriented method have many omission pixels,and the producer accuracy is only about 72%.The number and position accuracy of soil mound spot are only 57.14% and 50.00%.The producer accuracy and user accuracy of BP neural network method are above 85%,and the accuracy of mound spot is 90.00%,the area deviation is 0.30.It is a relatively optimal method to extract zokor mound in summer image.(3)The image of marmot damage area in spring: The producer accuracy of the holes is 50% which is extracted by BP neural network method.The user accuracy of the holes is 37.77% which is extracted by optimal color texture method.The number and position accuracy of the above two methods are only 27.59% and 23.81%.The number and position accuracy of hole radiation areas is less than 40% which is extracted by every methods,and the fragmentation degree of the spots is high.In the extraction result of rule-based object-oriented method,the producer accuracy of the marmot holes and hole radiation areas are 83.02% and 74.46%,the user accuracy are 100.00% and 99.79%,the number and position accuracy are 52.17% and 38.46%,and the area deviation and shape deviation of the hole radiation areas are 0.13 and 0.29.It is a relatively high precision method to extract marmot damage imformetion.(4)The image of marmot damage area in summer: All the methods have good accuracy in extracting the hole radiation areas.Except for the gray threshold segmentation method,the classification accuracy of other methods are more than 90%,and the number and position accuracy are more than 80%.For marmot holes,the extraction accuracy of each method is quite different.There are many commission pixels in the extraction results of the optimal color texture and object-oriented method,and the user accuracy are only 45.10% and 65.81% respectively.The number and position accuracy of gray threshold segmentation and optimal color texture are poor,only 59.46% and 18.52% respectively.The number and position accuracy of the holes extracted by BP neural network is 74.19%,the producer accuracy and user accuracy is 82.37% and 99.95%.The area deviation and shape deviation of the hole radiation areas is 0.24 and 0.23,respectively.It is the best method to extract the marmot damage information in summer images.