Remote Sensing Image Classification Based On Multi-Source Data

At present,as one of the most important fields of global change research, the study on dynamic monitoring of land use/land cover change (LUCC) and getting its information quantificationally based on remote sensing technology have been the frontier. The automatic information extraction and classification of remote sensing image is the key for us to realize the aim. This paper takes a case study on the land use classification of the outskirts of yulin city in Shaanxi Province and deals with the TM image classification respectively based on MLC, Decision tree and SVM method integrated the information of spectrum and texture.The main results are as follows:(1)Constructed the new bands based on TM image using MNF, K-T and NDVI. The new bands as the new data including former four bands of MNF, former three bands of K-T and NDVI band will be used to class.(2)In order to extracting texture information, the TM image was firstly analyzed by principle component transform, and secondly, the PCA1 and PCA2 image were analyzed with Gray Level Co-occurrence Matrix using eight texture features. Discussing the texture windows, the result showed that the windows size of 5x5 is the better.(3)Classified the image by using three methods of MLC, Cart and SVM based on information of spectrum and texture. Evaluated the three results of the classification by comparing and analyzing the chart and the Kappa coefficient, the result show that: SVM with information of spectrum and texture had a precision appraisal score of 86.67% and a Kappa coefficient of 0.8501, which were significantly higher than others. Experiments demonstrate that there were relatively complete landmark in land-use image without post-processing with high accuracy and less classification error. Under the same conditions, SVM can make better use of multi-source data and had higher precision and adaptability. Compared with SVM only based on spectrum information, the result shows that texture information can improve precision.(4)Analyzed the reasons of low classification precision by using the Cart methods to classify image, the reasons were as follows: some samples selected were not representative, which cause samples impure and data noisy. If the landmark is few, then there are fewer samples to be selected, so it leads to some rules constructed by less data; the principle of acceptance or rejection affects the result when decision tree was translated into rules. We can use many methods to improve the purity of the samples, selecting the samples evenly, clearing the data noise and the data which are independent of the classification target to eliminate the effect of the selected samples in experiments.