Coastal Land Covers Classification Of Remote Sensing Images Based On Data Minning Technology

Supposed by 908 Projects of Marine comprehensive investigation and assessment in China, in the view of data mining technique, coastline interpretation method and coastal land covers classification method from remote sensing images are studied.For coastline interpretation method, since shoreline are divided several types, such as artificial shoreline, bedrock shoreline, arenaceous shoreline, silt soreline and biologic shoreline, characters of different shorelines are analysed firstly and five kinds of coastline interpretation methods of the corresponding five types are designed. To verify the presented methods, experiments are implemented by Landsat TM/ETM+ images with the image resolution 30m. The precisions overmatch three pixels. Our proposed methods are intuitionistic and easy to realize. Moreover, the interpretation coastlines are consecutive. In addition, a coastline interpretation accuracy assessment algorithm is proposed to evaluate the experiment results.For coastal land covers classification method, aiming at the increasing resolution of remote sensing images, two classification methods are designed. One is object-oriented coastal land covers classification method based on evidence theory, the other is object-oriented fine land covers stratified classification method based on data mining.Aiming at II-level land covers type of remote sensing investigation, introducing the idea of evidence theory, coastal land covers classification method based on evidence theory is presented. To verify the method, SPOT image with 10m resolution is used. Considering spectrum, texture, shape and neighbour features synthetically, an experiment is implemented. The accuracy beyonds 80%. The proposed method solves the uncertainty of classification by single attribute, which improve the accuracy of classification. In addition, the basic D_S evidence theory is extended to process the incompleted sample space.Aiming at III-level land covers type of remote sensing investigation, object-oriented fine land covers stratified classification method based on data mining is proposed. To verify the presented method, Quickbird image with 0.61m resolution is used. From numerous features of spectrum, texture, shape and neighbour, by using association rule technique, attributes that can distinguish each land covers type are discovered firstly. Then basing on the attributes knowledge, the classification is implemented, which the accuracy of result beyonds 80%. This method eliminates the dependence to systematizer, which is an automatic classification method.