Research status at home and abroad

The main purpose of the study is to collect, analyze and evaluate information about the classification of hyperspectral remote sensing images in order to reach a new conclusion or to find out the effectiveness of something. Hyperspectral remote sensing image classification and optical remote sensing image recognition are important research materials in the development of remote sensing technology. From the viewpoint of image data, hyperspectral image classification spectrum data can be divided into classification methods based on spatial data, classification methods based on space-spectrum combined data, and so on. This topic analyzes the status of current research at home and abroad, and coordinates methods using hyperspectral remote sensing image classification and image identification and as well as the methods combined with deep learning in the fields, and summarizes the existing problems and development trends. From the perspective of model training, it can be classified as supervised training classification method, semi-supervised training classification method, unsupervised training classification method, and so on.[18]. Traditional classification methods for multispectral images, such as Support Vector Machine (SVM) and Random Forest (RF), have been gradually applied to hyperspectral data. The spectral features of the class for each training are defined by the possibility of statistics or feature space, and the unknown pixels that are classified are compared with the statistically known class and the most similar class labels are assigned to them Method. [19]. In remote sensing image processing, features have been used as effective methods to improve the use of sample features. Xu et al. [20] weighted and added the spatial features and spectral features of the hyperspectral remote sensing images extracted by the attribute feature method to achieve feature fusion, and used a support vector machine for classification. Qi et al. [21] Proposed a method for small object detection in single infrared images based on Boolean graph visual expression, which realizes target enhancement and background suppression by using Boolean graph weighted fusion of color and directional feature channels, thereby achieving object detection.

To be intelligent, you may have to learn something. Machine learning means that a program learns on its own. This can be a robot or any software. I'll say the logic of what works for an ordinary computer program, the program works accordingly. But machine learning programs are a bit different. They understand themselves when they need to work. Here the programs first train with some data. In the beginning, they don't give the right output. They give the right output at one point to learn.

Deep Learning (DL) is one of the technologies and research areas of machine learning. Artificial intelligence is realized in computer systems by establishing artificial Neural Networks (ANNS) with a hierarchical structure. Since hierarchical ANN can extract and filter input information layer by layer, deep learning has the capability of representation learning and can realize end-to-end supervised learning and unsupervised learning. Deep learning algorithms represented by deep neural networks with hot development of deep learning hyperspectral image have been well applied to the work of classification, and a large number of deep learning structures have been achieved to improve the learning of accuracy in hyperspectral image classification. Meanwhile, the Convolution Neural Network (CNN), Stacked Auto-Encoder (SAE）and Deep Belief Networks in 2006 (DBN) have made significant contributions to the category of hyperspectral remote sensing image. Reference [20] proposed a framework for hyperspectral data analysis and verified the performance of Restricted Boltzmann Machine (RBM) and DBN through classification based on spectral information. Reference [22] designed a CNN-based framework for feature extraction and classification of hyperspectral images. Reference [23] Hyperspectral designed a CNN-based structure for photo feature extraction and classification. Classification methods have been studied respectively based on spectral data, special data, and space spectral combination data to deal with the shortage of training samples. A virtual sample construction method has been proposed to compensate for the lack of labeled samples and the imbalance of samples.

Hyperspectral image in image classification is the primary domain in which deep neural networks (DNN) play the most important role of HSI image analysis. The deep forest is a deep learning model proposed by Zhou et al. [24] Image classification accepts the given input row images and produces output from input image classification. CNN is a typical deep neural network model. CNN-based hyperspectral image classification is mainly 4 categories as shown in Figure 1.2. By extracting the neighborhood information around the pixel points and inputting it into the convolution layer for feature extraction and classification of spatial information. Only the spectral vector corresponding to a single pixel is stretched and input into the one-dimensional convolution layer for feature extraction and classification.

After feature fusion of spectral information and spatial information, input it into the convolution layer for feature extraction and classification. In addition, the three-dimensional convolution kernel is directly used to take the three-dimensional blocks of the empty spectrum features and input them into the convolution layer for feature extraction and classification. A complete spectrum is achieved at each point without the need for any prior knowledge of the operator samples. Hyperspectral imaging can also take advantage of spatial relationships between different spectra in the neighborhood of pixels, allowing for more detailed spectral-spatial models for further division and classification of the image. However, due to the specificity and complexity of hyperspectral data, at present, the classification method of hyperspectral remote sensing image based on the deep neural network has some difficulties, such as high requirement on the number of training samples, difficulty in parameter adjustment, and model overfitting is other disadvantages for cost and complexity. Fast computers, sensitive detectors, and large data storage capacities are needed for analyzing hyperspectral data so on[21]. Therefore, the problem of hyperspectral image classification should start from the point of view of data, design the feature extraction and classification structure with good performance and improve the classification efficiency while minimizing the need for representations and samples as much as possible.

Figure 1.2 Classification framework for specific images based on CNN.

It is a decision-making method of decision tree methods with much less hyper-parameters than deep neurol networks and its model complexity can be automatically determined in a data-dependent way. Current dip-learning models are mostly created through the neural networks, multiple layers of parameterized different able non-linear modules that can be learnt by backpropagation. Reference[25] proposed a deep connected deep random forest structure for hyperspectral image classification.

Principal Component Analysis (PCA) is shown to reduce the dimensionality of the original hyperspectral data. Then the spectral vector corresponding to a single pixel is input into the deep forest for training and classification through the cascaded deep forest and the original hyperspectral data were first dedimensionalized using principal component analysis (Principal Component Analysis, PCA), retaining the first 10 main components as high-spectral images after PCA de-dimensionality. The corresponding spectral vector of a single pixel is then fed into the deep forest, which is trained and classified by cascading the depth forest. Each forest contains four random forests and a completely random tree forest (Complete-Random Tree Forest, CRTF), and the class probability output from each forest is cascaded and fed into the next forest. The number of layers of a forest is determined by k-fold cross-validation, and the number of layers does not increase when the classification accuracy of a layer of forest does not increase significantly on the validation set. After the average probability of the output class of the last forest. The class corresponding to the highest value is selected as the final output class. Experimental results compared with deep neural networks and traditional machine learning algorithms show that the proposed classification structure has higher classification accuracy than deep neural networks and has faster training speeds. However, the classification framework still has some room for improvement, such as the use of spatial environmental information to achieve the effect of support classification. Now may not deliver the expected results. We Considered that this study will be able to come up with better and more predictable of deep forest. 

Figure 1.3 Classification framework of deep forest hyperspectral image based on spectral information

A deep multi-granularity cascade forest (DgcForest) for deep hyperspectral image feature extraction is proposed in this paper. Although DgcForest improves on the original deep forest and makes up for the lack of the original deep forest in the HSI classification. However, the spatial feature utilization rate of DgcForest for hyperspectral remote sensing data is not high.

Reference proposes an improved deep forest algorithm, which is named deep multigranular cascade forest (DgcForest). As shown in Figure 1.4, the spatial information of  hyperspectral images can be extracted by using the multi-granularity scanning framework, and the surrounding neighborhood information can be extracted as well as the spectral information. After multi-level feature learning, input to single-layer forest for classification[24]. In the DgcForest algorithm, the framework connects multi- multigradient scans at multiple levels and only uses one level to classify the forest layer. Compared with several deep neural networks, it is proved that the proposed classification framework has high classification performance.

Figure 1.4 Classification structure of a deep forest hyperspectral image based on spatial data