Research On Hyperspectral Remote Sensing Image Ensemble Learning Classification And Post-classification Processing

The hyperspectral remote sensing image has higher spectral resolution which provides a kind of effective measure for ground recognition. But the huge data and the high dimensions bring on that the traditional remote sensing image classification techniques can not be used, so the classification means which are fit for the hyperspectral remote sensing image need to be researched. Otherwise, because the applications of classification image are widening along with the development of GIS, the post-classification processing techniques containing dealing with noise, vectorization and making thematic map need to be studied. In allusion to the features of hyperspectral remote sensing and classification image the ensemble learning technique and the mathematical morphology theory were investigated so that they can be applied for classification and post-classification processing. Some achievements have been made as follows:(1) Hyperspectral remote sensing image ensemble learning classification methods were researched. The weak classifier based on CART was build and the real adaboost, the improved weighted vote bagging and the PSO selective ensemble learning classifier were brought forward. This experiment indicated that this three ensemble learning classifiers excelled the CART, real adaboost excelled the improved bagging and the PSO selective ensemble learning classifier was the best in classified precision, training speed and classified speed.(2) In allusion to the feature of the noise of classification image, a technique based on binary morphology theory which deals with the classification image noise was studied. The binary image of every class was obtained using classification image layering and the morphology structure element adopted the unit structure element so that the classification image noise problem was resolved basing on binary morphology theory and making use of improving dilate and erode operation. The experiments indicated that the technique of dealing with classification image noise of this paper pressed was close to the real classification distribution and its function of dealing with noise was enhanced significantly.(3) The classification image vectorization method based on binary image was achieved by five steps that were clearing up classification image noises, layering, extracting the classification edge, thinning and vectorization. The Canny was introduced for extracting the classification edge and the hit-miss transformation was studied for thinning the classification image. Lastly, the classification image vectorization processing was finished utilizing the following every point arithmetic and the taking out curve arithmetic in turns. The experiment proved that this vectorization method was better than ENVI.(4) The characteristic element of ground was analyzed particularly and the recodification about classification was studied so as to optimize the category. And then the classification image thematic maps were made using the ENVI combining the thematic map theory and the features of classification image.