Research On Forest Change Detection Methods Based On Remote Sensing Images

Forest is an important natural and strategic resource in our country.For better protection of forest resources and timely and accurate information on dynamic forest changes,the annual update of forest survey by the National Forestry and Grassland Administration for overall planning and unified deployment has become an important task for forestry departments at all levels.The forest inspection work generally uses domestic high-resolution satellite remote sensing data,combined with manual interpretation,to extract the forest change areas in two years.Due to the tight schedule and heavy tasks,the traditional survey method still has the problems of high labor cost and huge workload.The use of remote sensing images for forest change detection has advantages of low cost and high efficiency,however there are many change detection methods based on remote sensing data with uneven detection effects,so it is especially critical to find a universally applicable and efficient detection method.In this study,Cangwu County,Wuzhou City,Guangxi Zhuang Autonomous Region is taken as the study area,and the 2018 and 2019 images of GF-1,GF2 and ZY-3 satellites are selected as the experimental data,considering the forest inspector work,which requires fast pre-processing and stitching of the national multi-source remote sensing data in a short time,it is difficult to realize the accurate matching of the color of multi-source remote sensing data,so this paper also adopts the simplest data pre-processing method to make the experimental data closer to the remote sensing images in the inspector work,and explored the forest change information extraction.The results show that:(1)Using band difference method,vegetation index method,maximum likelihood classification method and support vector machine classification method to extract and analyze forest change areas in Cangwu County,the Kappa Coefficients are 0.55,0.43,0.61,0.58,and the accuracy rates of positive forest change extraction are 0.19,0.10,0.22,0.22,and the accuracy rates of negative forest change extraction are 0.38,0.16,0.27,0.35,the results show that the accuracy of forest change information recognition by several methods is low,and the detection results are difficult to be applied to production practices.(2)Based on the deep learning theory,we construct a forest change detection network,which is much better than the above four algorithms in forest change detection.Finally,the Kappa Coefficient of the detection results is 0.94,the mean intersection-over-union is 0.88.The accuracy rate is 0.88,the recall rate is 0.91,and the F1 score is 0.89 in positive forest change detection;the accuracy rate is 0.89,the recall rate is 0.93,and the F1 score is 0.91 in negative forest change detection.Both the positive and negative detection of the forest are very satisfactory,and this method is less influenced by the data source and data processing methods.Meanwhile,the positive change detection results of this network can also provide important data support for the future evaluation of carbon sequestration effect and plantation survival in national forests.(3)Among the forest change detection of remote sensing data from the GF-1,GF-2 and ZY-3 satellites,the GF-2 image and ZY-3 image have the most satisfactory detection results.In the case of less clouds and better data quality,using the data from the GF-2 and ZY-3 satellites,combined with the classification algorithm to extract forest change regions,we can also get more satisfactory results,but still a big gap compared with the algorithm of deep learning.