Preliminary Study On Remote Sensing Classification Of Land Use Based On Object - Oriented Decision Tree Algorithm

Remote sensing technology is the main means of land use / land cover information acquisition,in the process of getting the land use / land cover information using remote sensing image classification methods, method of selecting appropriate or not directly affect the accuracy of classification.It is only based on automatic classification method of pixel in the computer in the traditional remote sensing classification method,in this way only the spectral information participate in the the classification, result in that the classification accuracy is not up to expectations.Thus, it is necessary to explore new method. Through the analysis and summary of domestic and foreign scholars, obtained the way that taking the data of Landsat-8 OLI as the date source,taking a part of Shangri-La as the test area to carry out the land use / land cover classification based on the way of object-oriented decision tree. The main research contents and conclusions are as follows:1. The process of decision tree classification based on object oriented. Research shows that this method comprises the following steps: data pre processing、image segmentation and object feature extraction、training sample selection and data mining to construct decision tree rule、image classification. Image segmentation is the key step,the reason why is that different methods of segmentation results are not the same,at the same time, image segmentation directly affects the subsequent steps in the process of classification; in addition, auxiliary function of different characteristic variables of different image classification, select which variable to the auxiliary image classification to a large number of experiments to do comparative analysis.2. Experimental study of the test area in Shangri-La.(1)Image segmentation and feature extraction: through the experiment we know that using multi-scale segmentation method for image object image segmentation is the most suitable for classification of object, among these datas, the value of segmentation scale is 40, shape weight is 0.1and the compact degree weight is 0.6;besides, selectting seven bands(B1-B7) 、four vegetation index(NDVI、RVI、DVI and MSAVI)、one water index(NDWI)、the first three principal component analysis of the variables(PC1-PC3)、one shape features(Length/Width) and eight texture features, a total of 27 characteristic variables of the original image aided classification.(2)The classification accuracy evaluation: the classification results of Kappa coefficients of basing on C5.0 decision tree of fifth groups of feature combination algorithm and CART decision tree algorithm, fourth groups of feature variable combination are 0.844 and 0.769, the total classification accuracy are 85.8% and 79.0%,these datas are higher than other results and Maximum likelihood classification results(which Kappa coefficients is 0.699 and total classification accuracy is 72.4%). The results show that the method that the decision tree classification method based on object oriented can be good used in data classification, moreover,accuracy is higher than traditional methods.That is to say it is a kind of effective means of land use classification.(3)Analysis of different land types: the classification results of the method of C5.0 shows that: the measure of area farmland is 287.6892 km2、forest is 5632.4912 km2、grassland is 700.9399 km2、Construction land is 89.562 km2、Unused land is 40.7102 km2、wetland is 60.2905 km2、water is 84.56689 km2、glacier snow is 189.0577 km2 and other lands are 639.7567 km2. Each lands have different rules, among these rules, glacier snow mainly distributed in the regional of B1>687.846 and DEM>3614.86.Although the method of basing on object-oriented decision tree classification has high precision,however, in the process of selecting the optimal feature combination,we need doing a large number of experiments and spending more time. How to efficiently select optimal feature combinations of variables need to be further studied. Key Words: Remote sensing;Object-oriented;Decision Tree algorithm; Classification;...