Remote Sensing Image Classification Based on Adaptive Watershed Segmentation and Improved Support Vector Machine

摘要

This paper proposed an object-oriented classification method of remote sensing image based on adaptive watershed segmentation and improved support vector machine. Firstly, the image was segmented based on adaptive watershed transformation. We made four improvements to overcome the over-segmentation: a new method of noise reduction based on wavelet transformation and the Butterworth low-pass filter, marker extraction based on the scale, gradient and edge information, calculation of region expansion cost based on spectrum characteristic and fractal dimension, the strategy of adjusting segmentation threshold adaptively. Secondly, homogeneity regions were automatically classified based on improved support vector machine. Our work included two aspects: the improved SVM multi-class classification based on binary tree, and the SVM classifying algorithm combined with ISODATA. ISODATA clustering analyzing algorithm could provide high quality training samples for SVM algorithm. Then the improved SVM multi-class classification was carried out to construct more accurate classifiers. From the experiments, the proposed method can not only improve the accuracy of classification result, but also automate the classification process.