| Forest type identification is an important question in satellite remote sensing monitoring.Since low-resolution remote sensing images can’t meet the accuracy requirements of forest type identification,high-resolution remote sensing images are contained rich spectral information,spatial geometry information and texture feature information.Therefore,in recent years,high-resolution remote sensing images have been the focus of research in the field of forest tree species identification.The traditional pixel-based classification method can’t effectively utilize the rich spatial and texture feature information in the image,and also has the problem of "classification noise",which cannot accurately extract the forest tree species information.The object-oriented classification method is an extraction method based on image segmentation object features,which can make full use of the spectral,texture and spatial geometry information of high-resolution remote sensing images.In this thesis,based on GF-1 remote sensing images,the object-oriented multi-level information extraction is carried out on the ground objects in the Engebe ecological demonstration area.The results of the study are as follows:(1)Using a multi-scale segmentation algorithm,different segmentation scales are established according to the characteristics of different objects.Through the comparative analysis of scale,shape factor,spectral factor,smoothness and compactness,the segmentation parameters suitable for different ground features are selected.Using the ESP scale evaluation parameter tool,and according to the optimal segmentation scale principle and visual interpretation method,the segmentation parameters suitable for different features of the ground objects were selected to lay the foundation for further classification.The results show that the optimal segmentation scale,shape factor and compactness factor of the three levels are 222/0.5/0.5,167/0.5/0.6,81/0.4/0.6.(2)The object-oriented classification method based on multi-scale segmentation eliminates the influence of non-vegetation and cultivated land on the forest land,and can better classify the forest land in detail and avoid the mixed classification phenomenon caused by different objects.Through the training samples,the optimal features are screened out during classification,and the nearest neighbor classification and random forest classification are performed respectively.Use the systematic sampling method to generate random sample points in the study area,build a confusion matrix according to the actual categories and classification results,and evaluate the accuracy of the classification results.The results show that the overall classification accuracy of the nearest neighbor classification of remote sensing images in 2020 and 2021 is 88.58% and 87.26%,and the overall classification accuracy of random forest classification is 92.95% and 92.02%.The Kappa coefficients of the nearest neighbor classification are 0.86 and 0.85,and the Kappa coefficients of the random forest classification are 0.92 and 0.90.Applying the object-oriented classification method in this study can overcome the confusion problem in the traditional classification method,thereby providing a better direction for the classification of high-resolution images.(3)The object-oriented post-classification comparison method was used to detect changes in the study area,and the change information was extracted according to the post-classification comparison method.The accuracy of the change detection results is evaluated by the method of confusion matrix.The results showed that the land types in the study area did not change much in the two periods of remote sensing images.Mainly reflected in the decrease of 0.52% in cultivated land,1.72% in sandy land and 1.92% in grassland.In the entire demonstration area,sandy land and grassland account for a larger proportion,reaching 68.57%,and forest land accounts for 12.90%. |
PDF Full Text Request