Classification Of Hyperspectral Image With A Few Samples

Hyperspectral images usually have nanoscale spectral resolution, it can provides a wealth of spatial information and spectral information which is widely used in fine agriculture, environmental monitoring,modern military and many other fields. Hyperspectral images not only have many advantages, but also have some disadvantages, such as high spectral dimension, high spectral correlation, much noise and rare samples, which bring many problems and challenges to the hyperspectral image classification process. The study of how to solve the above problems so to improve the classification effect, is one of the important research subjects in hyperspectral remote sensing field.In view of the disadvantages which the sample is difficult to obtain and the sample size is limited during hyperspectral image classification process, this thesis committed to solve the difficult problems due to lack of training samples from the perspective of dimension reduction,de-noising and feature extraction by the study of obtaining better characteristics of surface feature classification to reduce the burden of the classifier to improve sample quality. The study work of this thesis includes the following two aspects:1. In order to overcome the influence of strong noise on the quality of rare samples in hyperspectral images, low rank and sparse matrix decomposition methods, such as RPCA (Robust Principal Component Analysis), could be used to remove some of the strong noise. RPCA uses the kernel norm to approximate the rank of the low rank matrix, but there is an obviously deviation from the rank of the approximation method to the real. Thus proposing an extraction method based on non-convex RPCA feature, and non-convex RPCA uses a gamma norm to approximate the rank of low rank matrix which can effectively approximate the real rank. In detail, non-convex RPCA low-rank and sparse matrix decomposition of hyperspectral images are used to remove sparse strong noise, and the classification is based on characteristic of low rank components. The experimental results show that the method can effectively overcome the influence of strong noise on the quality of rare samples, and obtain a good classification effect.2. The intrinsic image decomposition technique in machine vision is a technique which can simulate the brightness and composition of the human eye. Recently, it has been successfully applied to the classification of hyperspectral images, and obtained a good classification effect in the case of rare samples. However, when the reflectivity component is calculated, the pixel brightness and spectral similarity of the partial window of the hyperspectral image are mainly considered,while the information redundancy caused by the spectral correlation between adjacent bands, the spatial information and noise of the feature profile are not fully considered. In order to remove the redundant information and noise of the essential features, at the same time,considering the global spatial information, the classification effect under rare sample is further improved. In this thesis, it is proposed an improved hyperspectral image classification method based on intrinsic image decomposition. Firstly, using the minimum noise fraction technique to extract spectral features with few SNRs in the intrinsic features to remove spectral redundancy, then using the support vector machine classification, and finally combining with spatial information to improve the small errors classification of classification result. The results show that the improved method can further enhance the classification performance of the intrinsic image decomposition technique under rare sample condition.