| Summary: | 碩士 === 中國文化大學 === 地學研究所地理組碩士班 === 101 === Conventional Pixel-Based Image Classification methods use merely grey-level values to classify pixels but ignore other spatial characteristics of ground objects. Object-Oriented Image Classification method connects adjacent pixels according to their spatial relativity in order to build up image objects and then carry out image segmentation according to the segmentation scale and their eigenvalue. This study aims to discuss the segmentation scale of bare land image objects. Multi-Resolution Segmentation is taken as a main method to combine adjacent objects according to Homogeneity Criterion, and then classify those combined image objects. Scale parameters represent the maximum limit when pixels merge into image objects. There are four parameters: color, shape, compactness and smoothness. In this study, we use high-resolution Formosat-2 satellite images to execute image classification on bare land. The land-use map provided by the National Land Surveying and Mapping Center, Ministry of the Interior, is used to compare the accuracy of the classification outcome and to study the accuracy of the results of the effect of image classification under different segmentation scales. Compared to the Pixel-Based Image Classification method, the outcome of Object-Oriented Image Classification method takes spatial relations into account. In this way, the attribute data of GIS and the image classification result can be easily combined and quickly made into a thematic map.
|
|---|
