Study Of Large Areas Land Cover Classification Based On C5.0Data Mining Algorithm

In last ten years, extreme weather occurs frequently, which has brought greatendangerment to the world. The effects of the global change have been the focus of attentionfor all mankind. With the development of modern high-tech like space technology,information technology, computer technology and so on, remote sensing which has thecharacteristics of macro and comprehensive, multi-spectral, multi-temporal and all-weather,has rapidly developed into a kind of synthetical technology that has a variety of applications.It is played an important role in many fields of national economy construction. In recent years,the international community has been committed to the use of space remote sensing techniquefor global land cover and its changes. So far, only the United States and the European Unioncomplete the global land cover data with spatial resolution of1km and300meters. But theaccuracy is low, both the classification system and spatial and temporal resolution are difficultto effectively meet the study requirements of the global change and earth system. Aiming atthe urgent need of the study of China’s global change and earth system model, in2009,Science and technology department start the863key program named” Study on remotesensing mapping and key technology of global coverage of surface”.So, aimed at the need of geographical conditions and on the basis of the sub-topic of the863key program “the product development of land cover data in Oceania/Antarctica”, thispaper, using Landsat ETM+image data, apply the C5.0data mining algorithm to the landcover classification. The main research contents and innovation points as follows:(1) This paper expounds the main classification process on the basis of the C5.0algorithm by studying area of Australia. By making full use of the multiple feature fusion data,such as the spectral information and various characteristic variables (K-T and NDVI), we canimprove the precision of remote sensing image classification.(2) In this paper, large area land cover products is made according to eco-geographicaldistricts as the basic unit. Samples were selected by the eco-geographical districts, on the onehand, it can reduce the time collecting samples, on the other hand, it can ensure thecomprehensiveness of the samples to reduce the unreasonable situation of image classificationedge matching because of the phase difference. In addition, classification rules establishedcan apply to all the images of the eco-geographical districts to avoid creating rules for eachimage and this mathod can improve the efficiency. (3) For the problem of the difference of feature type proportion in image and theclassification accuracy and by the study of the extraction order of object classes and accuracyevaluation, the paper proposes reasonable classification strategy that is from coarse to fine toreduce the processing workload and improve the efficiency.(4) In addition, based on the completion of land cover classification of the Australiandata in2000, this paper completes the land cover products in2010by use of change detection,so that it can avoid reclassification, reduce the workload and save time. By means of theclassification experiments of two typical geographical regionalization of ecological, this paperproves the good adaptability, high precision of classification and good stability of the C5.0classification method which is suitable for producing the land cover classification productionof large area.