Deep Learning Based Hyperspectral Image Classification And Anomaly Detection

In this paper,deep learning of popular tools for big data analysis has been applied to the analysis of hyperspectral images in recent years.Deep features of deep learning networks have replaced traditional manual features to analyze hyperspectral data.There are few samples for hyperspectral images in deep learning applications,and it is impossible to construct a depth model.The similarity of hyperspectral pixels is analyzed.The pixel matching model is proposed and a deep neural network framework is designed to solve the deep neural network training.In the process,there are a lot of problems with labeled samples.The main work of the dissertation is as follows:First,for the problem that there are few hyperspectral data samples and it is impossible to build a deeper network,a deep neural network using pixels to classify features is proposed.The method increases the training data set by pairing the pixels,extracts the features between pairs of pixels,and makes full use of the characteristics of the deep neural network data drive.During the test,the final voting test is performed by classifying the test pixel with the pixel pair of its neighboring pixels.Compared with traditional methods,the advantage of this model lies in the use of data-based depth features,which have different adaptability to different data,unlike traditional fixed artificial design features.Secondly,aiming at the problem that the traditional hyperspectral data anomaly detection has no effective extraction of spectral features,a deep network using pixel pair features for anomaly detection is proposed.First,use tagged data to construct pairs of pixels.Train a depth network to determine whether two pixels belong to the same class.After acquiring the depth characteristics of the pixel pairs of the test pixel and its neighboring pixels,migrate the model to the anomaly detection data.On the other hand,by analyzing the characteristics of the pair of pixels in the center point and its neighborhood,it can be known whether the center point is an abnormal point.Compared with the traditional method,the advantage of this model is that in addition to the use of statistical information,the characteristics of the spectrum are analyzed,and the detection of outliers is more accurate.