| The remote sensing image classification is always an crucial part in the remote sensing area. Therefore, the research on how to resolve the problem of the subcategory of the cover type, especially the vegetation type and how the precision can be up to par is of great significance. The traditional classification method of the remote sensing image is based on the classification of pixel which could not use the space and texture of the image efficiently and the results of the classification could arise the problem of "salt and pepper" and low precision effects. However, object oriented classification method is based on the classification of object by integrating multi-temporal and multi-source remote sensing data, and by making use of spectrum, texture and spatial information, etc. This new way could overcome the weakness of the traditional way. But, as it is based on the high-resolution image, there are few researches on remote sensing monitoring of land use and land cover in the wild area which is covered with complex feature by adopting the medium-resolution image.According to this, besides of using other auxiliary data such as DEM, this article uses the HJ-1CCD and Landsat5TM medium-resolution remote sensing image as the main data resources, and the eCognition as the software platform to extract the subcategory information of remote sensing image in Jingjinji area. As the research area covers a wild range with complicated cover types, this research use the method of division of the operating block to classification and chooses two typical area (a plain and a mountainous area) as the research area. For each area divide the area into different parts, then set different segmentation parameters for the varied terrain according to the multi-scale segmentation to get the multilayer extraction. After completion of the classification of each operating block, we can get the classification results of the land cover in the researching area by splicing and processing. The main results are as follows:1. After the multi-scale segmentation test on the plain and the mountainous area, the appropriate segmentation scale for the various cover types could be obtained. By combining the experience and the extracting aim, the fittest segmentation scale and parameters could be got. Then by adopting the multi-scale segmentation method, we can extract the land cover with unified spectral signature on large segmentation scale20such as on the wet land or bare soil, etc. Later we can extract the information on smaller segmentation scale10for the vegetation, non-vegetation and some subcategory types. And it is much better to extract the information of the small land cover type on smaller segmentation scale5, such as in the coniferous forest or the forest land of plain terrain, etc. According to the extraction aim to extract the land cover in different layers, the mix of different land covers could be avoided efficiently and the different land covers can be distinguished well. It will be beneficial for extracting the information based on the object-oriented classification.2. On the basis of the multi-scale segmentation, by integrating multi-temporal, multi-source remote sensing data and other auxiliary data such as DEM, we can extract the land cove information in Jingjinji area under the guiding of spectral signature and spectral angle classification method. By this way, we can solve the problem of how to extract the subcategory information in the wild complicated area by using the medium-resolution remote sensing image better. The characteristics of the object-oriented classification method include spectrum, geometry, texture and spatial information, etc. For example, we can use the length-width ratio to distinguish the living land and traffic land, reservoir swag and river. In most cases, we extract the information of the land cover based on the spectral signature. Regarding to the information extraction of evergreen coniferous forest and deciduous broad leaved forest, we can use winter image to distinguish them effectively. And for the information extraction in the drought land and living land in the mountainous area, by the aids of DEM and slope, the odds of making mistakes between the living land and drought land could be decreased. While for the small range of the forest in the plain area, it is much difficult to extract the information of it. So by comparison, this paper adopts the spectral angle classification to extract the information of forest in plain from the drought land better.3. Compared the confirmation of the precision of the results with the traditional maximum likelihood and the object-oriented classification method, it shows that the precision of the object-oriented classification method is relatively high and this method can be applied to the subcategory information extraction in the wild complicated area by adopting the medium-resolution image.The total accuracy of plain area based on the object-oriented classification method is91.51%, and the Kappa coefficient is0.89, which is18.82%higher than that which the accuracy is only72.69%based on the traditional maximum likelihood. The precision of all kinds of the extraction information has been enhanced, especially the precision of the forest in the plain area is enhanced a lot by adopting the method of spectrum angle classification. And the total accuracy of mountainous area based on the object-oriented classification method is83.65%, and the Kappa coefficient is0.81, which is15.83%higher than that which the accuracy is only67.82%based on the traditional maximum likelihood.By using the multi-temporal data such as the image in the unplanted or harvest time and the auxiliary data such as DEM, the precision of the extraction in the drought land and coniferous forest has been enhanced a lot. According to the precision evaluation in the Beijing, Tianjing and Hebei area, the first-order accuracy and the second-order accuracy is90.43%and82.97%respectively which can meet the requirements of the precision of the project and image classification. Therefore, object-oriented classification method could be used to extract the subcategory information by using the medium-resolution remote sensing image in the wild complicated area and could provide a technical process for reference. |
PDF Full Text Request