| Land is the most basic resource for our living and development and it is also an important part of our country resources. Analysis of the current land ues is the baic for utilization, planning and scientific manage of land rusource. But, it is a complex work and needs a great amount of actual data. Therefore, we need a good method to exrtract the current land use information that can improve efficiency and quality. This work is very important to investigate and update the current land uesage and its changing.Remote sensing technology has features such as wide-rigion, real time monitoring, periodical information, and comprehensive information. All these features give the superiotity for a fast, objective, precise investigating and updating of lan use. At present, the mostly used methods are interpretation and traditional auto-classified technology. But there are some disadvantages. Decision tree have several advantages for remote sensing applications by virtue of their relatively simple, explicit, and intuitive classification structure. Little work has been done to the application of decision tree method on extracting the current land use information using higher resolution SPOT5 imagery.In this research, one popular decision tree algorithm—C5.0 are presented, and one technique known as boosting in machine learning area are introduced. We took lanxi area, zhejiang province, china as an example. The method and result of SPOT5 multispectral land cover classification using decision trees technology was presented in detail.The main conclusion of this study is summed up as follows:(1) Analyzing the results of two different sampling projects which classified by decision tree. The result indicated that training samples should be considered integrity and homogeneous. As it was convenient for conversion from data to rulesets. When we chose some aoi that their DN are exactly the same, it was not property to extract information or rules. Because it is too pieces to summarize rules. Simultaneity, defining the sort that have striking dissimilarity to several selfsame sorts as an absolute sort. Then joined them together after the classification. The result showed that it can increase accuracy.(2) A study has been carried out to find out which combination of characteric data are the best for decision tree classification. The result showed that the accuracy of decision tree classificationimproved by the increasing numbers of training samples. And a good decision tree classification depends on the large numbers of training samples.(3) The tree associated remote sensing image with 4 characteric datas, such as NDVI, texture, elevation and slope, was translated into a clear, quantitative system. Results showd that classification accuracy increased for adding characteric data. It displayed the influence of such phenomena as the same thing with different spectrums, different things with the same spectrum. But different combination of characteric datas gave different accuracy. Elevation and slopes had the most important help for accuracy. NDVI was the least important cause.(4) The paper analysed the structure of decision tree, pruning of decision tree, conversion of decision tree to decision rules. Results showed classification accuracy lied on rationality of tree structure, and it was independent on the complexity of tree. It was easy to understand the knowledge expression when decision tree convert to decision rules. We also tested the behaviour of boosting techniques combined with See5.0 and the result showed that it can increase classification accuracies.(5) The result indicated that decision tree with abundance training samples has higher classification accuracy than maximum likelihood classifier in the land cover classification test. |
PDF Full Text Request