Study Of The Scale Issues In The Information Extraction Of Forest Vegetation Based On Remote Sensing

In this research we use two kinds of remote sensing images which are ETM+and SPOT-5 of Zhenjiang City of Jiangsu Province as main data sources. It is mainly about the scale issues in object-oriented image analysis, and extracts the information of forest vegetation in different ways and scales, with the help of remote sensing software such as ERDAS9.2, eCognition8.0, ArcGIS9.3, etc. In the extraction progress, the method of object-oriented multi-scale image segmentation and fuzzy classification, as well as related scale issues has been studied deeply. And then the extraction result is compared to the traditional pixel-based method.The main contents concluded as follows:(1)The scale issues involved in this paper include:selection of segmentation scale, scaling, scale effect and selection of optimal scale program. Unlike previous study, the selection of optimal scale is not only selecting the appropriate segmentation scale before segmenting, but more important is to find out optimal scale scheme according to the extraction results.(2) Making principal component transformation merge between multispectral images and panchromatic image, is pixel-based scaling down. At the same time, resample SPOT 2.5m spatial resolution of panchromatic band by cubic convolution, and scaling up to 5m resolution of panchromatic image, which is a basis data for fusion. Use the ETM+and SPOT-5 remote sensing image to do the fusion work, and obtain four kinds of images:SPOT 10m+SPOT 2.5m, ETM 30m+ETM 15m, ETM30m+SPOT 2.5m, ETM 30m+SPOT 5m. The spectral information and spatial information of fusion image have been enhanced.(3) After several tests and considering the efficiency and effectiveness of information extraction, finally choose the ETM 30m+SPOT 5m fusion image, and construct images hierarchy as scale 100, shapes 0.1, compactness 0.4; scale 200, shape 0.2, compactness 0.4; scale 300, shapes 0.2, compactness 0.3.(4) Use maximum likelihood method to extract vegetation information from TM30m+SPOT2.5m fusion image with minimum spatial scale and maximum spectral information. The result shows the total forest area is 28836.7ha, the overall accuracy is 74.67%, and the forest kappa index is 0.7778, the overall kappa index is 0.6379, visual valuation scores 83. Many areas has not been classified as forestland but divided into other vegetation. The extraction of object-oriented fuzzy classification improved accuracy, also make the extraction results more actual. There are 2272 objects on scale 100, forest area is 38913.8ha, the overall accuracy is 85.34%, and the overall kappa index is 0.7852, visual valuation scores 150.727 objects on scale 200, forest area is 39356.4ha, the overall accuracy is 84.14%, and the overall kappa index is 0.7695, visual valuation scores 134.346 objects on scale 300, forest area is 32925.5ha, the overall accuracy is 82.23%, and the overall kappa index is 0.7308, visual valuation scores 108.(5) Comparing and analyzing the extraction results of different scales with object-oriented method show that the optimal scale scheme of forest vegetation information extraction is:ETM 30m+SPOT 5m fusion image with scale threshold in 100～200, shape parameter between 0.1～0.2, color parameter between 0.8～0.9, compactness parameter between 0.3～0.4, smoothness parameter between 0.6～0.7.This research indicates that the object-oriented information extraction is more precise than the traditional method of pixel-based extraction, and the information extracted under different scales is more accord with the objective reality situation, it fully displays the superiority of the application of object-oriented method to analyze image information extraction scale problem.