Research On Target Classification For Forest Change Detection Based On High Resolution Remote Sensing Images

This paper aims at the application of forest change detection in Dianjun region of Yichang city,Hubei province,and adopts GF-2 remote sensing image data to carry on the contrast analysis,solves the vegetation refinement classification and the related forest resources change information discovery and the extraction.A target classification study for the reliable and effective flow realization of forest change detection was carried out with GF-2 remote sensing images in 2015 and 2016 as the data source,with the help of data processing software such as ENVI5.3,Arc GIS10.4.1,e Cognition9.0.By means of Orthorectification,image registration and image fusion,the multi-scale segmentation parameter evaluation function is defined based on the absolute value of neighborhood difference and standard deviation ratio,which is used to determine the segmentation scale,shape factor and compactness in image segmentation.Then we use the neural network classification method to determine the optimal feature combination based on object classification,and use the object-based multi-level classification method to classify the remote sensing image.Finally,the forest category in the two remote sensing images classification results is detected.The results show that in the process of classification,the total accuracy of multi-level classification based on multi-scale segmentation objects combined with supervised classification can reach 98.66%,and the multi-level classification is also obviously better than that without stratification.The main contents and results of this paper are summarized as follows :1)The classification of vegetation as the main target in Yichang Mountain area,and the application scheme of forest change detection in high resolution remote sensing image based on object classification.The steps are: image preprocessing,image segmentation,feature combination contrast and selection,selection of training samples,classification contrast,change detection and change detection post-processing,in which image segmentation,feature selection and classification scheme construction as the main research content.2)Includes the selection of segmentation parameters under multi-scale conditions and the selection and evaluation of segmentation scale,and the feature selection module uses a classification method to compare the classification results of any combination of features,and selects the most suitable feature feature combination of remote sensingimages in this study area.The features include spectral characteristics of the image itself,normalized vegetation index(normalized differential vegetation index,ndvi),normalized water body index(normalized difference water index,ndwi),texture,principal component and so on.from the traditional classification method,we first select the classification method with the highest image coarse classification accuracy,that is,the maximum likelihood classification,the overall accuracy can reach 96.77%,and the kappa coefficient can reach 0.92.Then the neural network classification,maximum likelihood classification,k-nearest neighbor classification(k-nearest neighbor,knn)are compared based on the optimal feature combination,and the result of object-based knn classification has the highest accuracy,and its kappa coefficient can reach 0.97.3)The object-based multi-level classification scheme uses the segmentation multi-scale and the interval delineation of the eigenvalues for stratification,this paper is divided into three layers.The first two layers are fuzzy membership classification,and the threshold of different object features is set to extract vegetation.The first layer extracts non-aqueous bodies when the mean value of near-infrared band is greater than 156 and the ndwi is less than 0.28.The second layer extracts vegetation from non-aqueous bodies when the ndvi is greater than 0.12.4)Classification situation can obviously see that the multi-level classification accuracy is higher and the kappa coefficient is larger,at the same time,the multi-scale according to the principle of layer by layer decline,the final fine classification results are the best,and the overall accuracy can reach 0.98.Under the good classification result,the classification chart of the two phase image is further compared and analyzed by using the envi5.3 tool,and the change detection result is obtained.The vector result is imported into the arcgis for the change detection post-processing,and the final change vector small class is obtained.Proposed method of forest change detection based on object classification has a standard and perfect flow,high efficiency and feasibility has certain reference value.At the same time,the proposed multi-level classification method based on multi-scale segmentation has high precision and good feasibility,and can also be widely used in the classification of other application scenarios.