Research On Classification Of Hyperspectral Remote Sensing Image Based On Active Learning

With the rapid development of computer technology, the machine learning as a multidisciplinary cross discipline, has become one of the most popular areas in computer science and technology. In recent years, active learning as a hot research topic in the field of machine learning, has got more and more attention. Hyperspectral images are obtained by spectroradiometers with dozens or even hundreds of continuous spectral channels, so they have the advantages of the very high spectral resolution and rich band information. Because of these, hyperspectral images are widely used in many areas for researching, such as vegetation, ecological atmosphere and ocean. For hyperspectral image, classification technique has become the hot topics in the study of hyperspectral image. At present, active learning algorithm based on Support Vector Machine(SVM) becomes more and more popular in the application of hyperspectral image classification. Considering that the traditional active learning algorithms based on SVM use only single spectral information and ignore the spatial information of source image in the process of classification image, this thesis puts forward several kinds of classification methods which combine active learning and the spatial information. The research content is as follows:1. This thesis proposes a hyperspectral image classification method based on the correction on the active learning results. In this study, an initial classification result is obtained by a traditional SVM active learning algorithm for the hyperspectral image classification firstly.Then, I calculate the spectral similarity between each training sample and its neighborhood samples, and the initial classification results are corrected by the comparison results between the spectral similarity and the given threshold. Finally, the final classification results are obtained. The experimental results of hyperspectral images show that the proposed method is better than other similar algorithms by improving the classification accuracy.2. A two-stage hyperspectral image classification method based on SVM active learning is proposed. This method divides the common hyperspectral image classification process into two stages, and the spatial information of images is used to classify the neighborhood samples of each training sample in the first stage, then in the second stage, the remaining samples are classified by the combination of traditional active learning algorithm and the classification results of the first stage. Finally, the combination of these two classification results is done to get the final result. The experimental results of hyperspectral images show that comparing with other similar algorithms, the proposed method can improve the classification accuracy without changing the size of the training sample set.3. This thesis proposes a hyperspectral image classification method that combines image segmentation and active learning classification. Firstly, a segmentation map is obtained by a hyperspectral image segmentation method, then a classification map is obtained by using the hyperspectral image classification method based on SVM active learning. Finally, the majority vote fusion of the segmentation map and the classification map is done to get the final result. The experimental results of hyperspectral images show that comparing with other similar algorithms, the proposed method can not only reduce the "speckle" error, but also improve the classification accuracy.