2. Machine Learning Model in Hyperspectral Image (HSI-A)-Anomaly Detection (AD)

2.1. Traditional Methods

(1)

 Abnormal spectral features are easy to distinguish in the spectral domain;

(2)

 Abnormalities usually appear in smaller areas;

Therefore, it can be detected by establishing a reference background model or suppressing the background. Traditional modeling methods are divided into three categories: methods based on statistics, representation models, and tensor decomposition. The Gaussian assumption based on the statistical model (assuming that the background obeys a certain distribution) makes the statistical model mathematically easy to handle. Therefore, the statistical model has been widely used in traditional modeling. A large number of methods are devoted to suppressing the background and establishing the background model. In 1990, Reed and Yu [5]^[4] proposed the Reed-Xiaoli (RX) method, which models the background as a Gaussian distribution, and the anomalies are sparse and can be detected from the background by the Mahalanobis distance. Based on the RX method, global GRX [5]^[4] and local LRX [6]^[5] methods are proposed, which estimate the statistical parameters of the background based on the entire image and dual windows, respectively. However, in practice, due to the complexity of HSI, it is not enough to use Gaussian distribution to simulate background distribution. Therefore, KRX [7]^[6], cluster-based (CBAD) [8]^[7] and support vector domain description (SVDD) [9]^[8] and other methods are proposed. Representation-based models usually include collaborative representation, sparse representation, and low-rank representation. Among them, the collaborative representation-based detector CRD is based on the theory that background pixels can be well represented by neighboring pixels, but abnormal pixels cannot be well represented. Therefore, even with the participation of abnormal pixels, the background can be modeled adaptively. The Mahalanobis distance method based on low-rank and sparse matrix decomposition uses the LRaSMD technology to distinguish the background from anomalies, explores the low-rank prior knowledge of the background, calculates the statistics of the background, and then uses the difference of Mahalanobis distance to detect abnormal pixels [10]^[9]. Based on the theory of tensor decomposition, since HSI is regarded as a third-order tensor cube, the tensor decomposition method can be used to better mine the spectral and spatial characteristics of the spectrum. Since 2017, the application of tensor decomposition methods to solve the HSI-AD problem has become one of the hot spots [11,12,13,14]^{[10][11][12][13]}. In traditional machine learning methods, a common problem is that it is difficult for a feature extraction method to achieve good results in all types of data. The deep learning method solves this problem well and for different types of data, deep learning models can learn features autonomously based on different data.

2.2. Deep Learning-Based Methods

In recent years, deep neural networks have shown strong advantages in the modeling and generalization of complex datasets and can obtain more data features than traditional methods. Deep learning methods have been widely used in multispectral image classification, and some results have been achieved [15^{[14][15][16][17][18]},16,17,18,19], but the HSI-AD method based on deep learning has just emerged. There are three commonly used models for anomaly detection, namely, hybrid model, semi-supervised model, and unsupervised model. The hybrid model is mainly used as a feature extractor to extract robust features, which can reduce the dimensional disaster of high-dimensional data, but because it does not improve the surface features of the potential feature layer, it is a sub-optimal choice; the semi-supervised method performs anomaly detection using pretrained model data that only contains normal samples, while the unsupervised model can better explain the characteristics of the data, and is more suitable for data with limited samples and difficult to obtain labels, such as hyperspectral data.

2.3. DBN

Deep belief nets are composed of several layers of restricted Boltzmann machines (RBM). The use of unsupervised learning is similar to the autoencoder, its purpose is to retain the characteristics of the original features as much as possible while reducing the dimensionality of the features. Among them, RBM is a kind of neural perceptron. The output layer of the previous RBM is used as the input layer of the next RBM unit, which is stacked in sequence to form the basic structure of DBN.

2.4. GAN

Generative adversarial network (GAN) is another important network framework for unsupervised anomaly detection technology. The concept of GAN was originally proposed in the zero-sum game framework of game theory and has now become the most common and effective generative model. It generates new data points with certain changes by learning a specific data distribution. GAN consists of a generator model (G) and a discriminator model (D). The generator learns its features from the input data to generate as “true” fake samples as possible to deceive the discriminator, and the discriminator combines the real samples with the samples generated by the generator to identify and judge the authenticity. G accepts random noise as input and generates false data G(z), and D estimates the probability of a real sample. Therefore, the generator wants to fool the discriminator, and the discriminator wants to know the trick of the generator to know that the generator can generate the required data distribution [20,27,28]^[19][20][21].