Research difficulties and challenges

Let's compare the progress of the current study of hyperspectral image classification with that of the home and abroad, current research difficulties and challenges of the classification of hyperspectral images combined with the features of the hyperspectral image and the classification process and the current research difficulties and challenges of hyperspectral image classification are summarized as follows

（1）Hyperspectral images require an in-depth understanding of the amount of information, surface information, and spectral behavior. After the development of remote sensing technology in the second half of the 20th century, great changes have taken place in theory, technology and application. Hyperspectral images not only have rich spectral information, but also contain a lot of usable information in spatial scenes. And for the use of spatial information should fully consider the space redundancy feature and background interference effect, deal with spatial characteristics in the appropriate pretreatment and depth of feature extraction, and supervised feature extraction method is difficult to global characteristics, excavation so should select suitable unsupervised feature extraction method to make use of spatial information.

(2) The classification of pixels in hyperspectral imagery is often made more challenging by the availability of only small numbers of samples within training sets. The limited availability of training samples poses a great challenge to the classification of hyperspectral images. Indeed, it is often the case that the number of training samples per class is smaller, sometimes considerably smaller, than the dimensionality of the problem. The HSI classification has been an important research topic at present and the fundamental challenges are (i) curse of dimensionality and (ii) insufficient samples pool during training. Given a set of observations with known class labels, the primary goal of classifying hyperspectral images is to assign a class label to each pixel. Therefore, in order to fully mine features in limited training samples, it is necessary to effectively integrate spatial and spectral information and make use of the complementarity of the two information sources to improve the classification effect.

3) Unsupervised algorithms usually assume that a priori can obtain a limited number of labeled samples, and then use unlabeled samples to expand the training set. But in order for this strategy to work, a few requirements need to be met. First, new unlabeled samples should be obtained without considerable cost and effort. Secondly, as the number of unlabeled samples increases, the classifier may not be able to properly utilize all available training samples due to the limitation of computational performance [26]. In addition, if unlabeled samples are not selected correctly, these samples may have a negative effect on the classifier, resulting in significant differences or even reduced classification accuracy using the initial labeled sample set. Therefore, it is very important to identify unlabeled samples with large amount of information in a computationally efficient way, so that the classification performance can be significantly improved without using a large number of labeled samples.

4) Algorithms are the steps to do something and an algorithm is a mathematical method of solving a problem. The algorithm will take some data or parameters as input It will then give some output as a return at the end of the required competition. Nonlinear systems have been widely used in signal processing, automatic control, wireless communication, and artificial intelligence, which has promoted the rapid development of nonlinear system-related theories. The occurrence of linear and non-linear spectral mixing and noise interference in the measurement process. The continuous development of hyperspectral instruments allows for dimensions of data. This requires rapid calculation modes and performs effective classification of hyperspectral data speeds of different scales and situations.

  Organizational structure of the article

In this thesis paper, we study the classification method of hyperspectral images based on feature extraction and deep forest. The first chapter is the introduction which mainly elaborates the research background, significance and the problems to be solved in this paper. The second chapter is related theory and algorithm model, mainly introduces the basic algorithm involved in this paper, and integrates the two aspects of feature extraction and image classification. The third chapter is the research work of the extraction method of hyperspectral images based on visual salience. It mainly analyzes the construction of Boolean mapping and the extraction method of hyperspectral images based on visual salience introduces the algorithm flow of the proposed framework in detail and verifies the performance, and compares the algorithm on the data set. Chapter four is based on attention mechanism for the depth of the forest of hyperspectral image classification method research, the main analysis based on the characteristics of contribution to the attention mechanism model, the depth of the agency is put forward in detail based on attention mechanism forest hyperspectral image classification method framework, and in the data set on the performance of verification and comparison of the algorithm. The fifth chapter is the summary and prospect, which mainly summarizes the work content of the whole paper and prospects the current research field. References and acknowledgments are attached.