Remote sensitive images analysis is commonly using artificial neural networks. In particular convolutional neural networks (CNN) getting more and more attention in remote sensing images classification. A class of CNN based deep neural networks, commonly used for the analysis of visual images recognition and classification, used exclusively in images made up of 2D and 3D dimensions. An RGB image is typically captured using three different wavelengths, about 610 nm for red, about 550 nm for green, and about 460 nm for blue. The hyperspectral image, on the other hand, can be captured using a wavelength range of 400-1000 nm with a resolution of 400 spectral channels. Processing within hyperspectral remote sensing data is multispectral data and is challenging in the hugeness of remote sensing data processing, as many spectral bands contain large amounts of data. With the continuous increase in application requirements and the continuous maturation of remote sensing classification and detection technology, the spectral resolution of hyperspectral images is continuously improved, and the spatial resolution is also gradually improved. Since an RGB image consists of only three spectral channels, classifying spatial information based on a widely used machine learning model implies that it is a drawing of a conventional neural network (CNN). For hyperspectral images, it may have several hundred spectral dimensions, but it is more interesting to attention to spectral dimensions than spatial information. At present, both spaceborne and airborne hyperspectral sensors have high resolution, which provides a good basis for the classification and recognition of ground materials. At the same time, the “Hughes phenomenon” and band redundancy are still two important challenges in recent years and even in the coming decades. Relevant concepts in computer vision are introduced with the task of hyperspectral image feature extraction and a deep learning framework is used to combine visual features with deep abstract features, and efficient feature learning methods are used to improve the effect of classification.

Advances in sensor technology and spectroscopy have led to high spectral resolution images called hyperspectral images. Hyperspectral remote sensing is not generated directly. Based on remote sensing technology, combined with the continuous improvement of hardware technology and imaging equipment, researchers' requirements for remote sensing applications are also growing. As shown in Figure 1.1, the band of remote sensing image is gradually evolving from wide to narrow until the band range almost reaches the nanometer level, and the hyperspectral remote sensing technology is becoming more mature[9]. Hyperspectral remote sensing technology can simultaneously use dozen’s or even hundreds of extremely narrow  Spectral bands to collect data, capture more data and provide more valuable information[10].