A Deep Learning Method For Hyperspectral Image Classification And Anomaly Detection And Its Software Implementation

As subdivided spectral imaging data,hyperspectral remote sensing images contain abundant radiation and spectral information.Hyperspectral image(HSI)classification and anomaly detection are the core tasks in hyperspectral data processing.They are widely applied to environmental monitoring,geological prospecting,homeland security and many other fields.Hyperspectral data has many bands which are highly related with each other,high feature dimension and redundancy.Therefore,a large number of training samples are needed to reduce the instability of classification and anomaly detection.However,it is difficult and costly to obtain training samples in the practical application process of remote sensing.In addition,there are many problems during the HSI processing,such as the low spatial resolution,big intra-class differences,the phenomenon of mixed pixel,noise,and other complex situations,which bring great challenges to hyperspectral classification and anomaly detection.How to use small training samples to achieve high-precision HSI classification and spectral anomaly detection is a key issue to be resolved.This paper first reviews the research status at home and abroad in recent years.Based on the introduction of several popular HSI classification algorithms and HSI anomaly detection algorithms,it explores the characteristics of spectral-spatial joint data and study a novel model of deep learning deeply.The main contributions of this paper are as follows:(1)For the supervised HSI classification problem,a deep learning based algorithm called multi-directional spatial message propagation convolutional neural network is proposed.Different from the traditional layer-by-layer convolutional neural network,the proposed network adopts a novel model with slice-by-slice convolution and designs a multi-directional spatial message propagation convolutional neural network.Specifically,the slice-by-slice convolution is embedded into the feature map,and the spatial information of each pixel is propagated to the surroundings to enhance the ability of learning spectral-spatial joint feature.Experimental comparison shows that the proposed method effectively combines the traditional layer-by-layer convolutional neural network and spatial message propagation convolutional neural network.Thus,it can obtain abundant and discriminative spectral and spatial features,and improve the network's small sample learning ability.The proposed algorithm has the classification performance of high precision and stability.(2)For the HSI anomaly detection problem,a pixel pair matching based dense network algorithm is proposed.The premise of applying convolutional neural network to pratical problems is to have the ground truth of a large number of training samples.However,the lack of supervision sample information is quite common in the application scenarios of HSI anomaly detection.To this end,this paper proposes a method called “pixel pair matching”,which is to match the pixels in the HSI used for training the network,and label them to construct the pixel pair with the same type and that with different types.The spectral dense convolutional neural network is constructed with one-dimensional convolution and spectral dense unit,which can fully extract the spectral characteristic of the pixel pair,thereby comparing the similarities and differences between the two spectral pixels,while avoiding the disappearance of the gradient.Finally,a dual-window spectrum anomaly discrimination mechanism is adopted to match the points to be tested with their neighboring pixels as the input of the network.According to the average similarity and the set threshold,the spectrum to be detected is judged whether it is an anomaly point.As the experimental comparison results shows,the detection graph,ROC curve and AUC of this algorithm are superior to other comparison algorithms,and better detection performance have been achieved.(3)A software system that integrates hyperspectral image classification based on deep learning and hyperspectral image anomaly detection algorithms is designed and implemented.In this system,four popular classification algorithms,five anomaly detection algorithms,and the two algorithms proposed in this paper are integrated.Besides,it consists of several core modules: reading image module,hyperspectral image visualization module,hyperspectral image classification module,hyperspectral image anomaly detection module,classification and anomaly detection evaluation module.