Hyperspectral Remote Sensing Image Classification Method Based On Deep Learning

In recent years,the automatic classification of hyperspectral images based on deep learning has become one of the research hotspots in the field of remote sensing.Hyperspectral data can be classified with high accuracy on the classification model represented by convolutional neural network(CNN).However,in the classification process of hyperspectral remote sensing images,there are problems such as weak feature selectivity,single feature level,and single classification decision,which may result in weak performance in classification.Therefore,two hyperspectral data sets,Indian Pines and University of Pavia are utilized to carry out research,and the detailed information is:(1)The 3DCNN＿AM(Three-Dimensional Convolutional Neural Network and Attention Mechanism),which is constructed by a 3DCNN based on the attention mechanism(AM),is proposed for solving the unrepresentative problem of hyperspectral remote sensing image feature extraction and classification.The attention of AM to specific information and the deep feature extraction capabilities of 3DCNN are effectively combined to perform feature extraction and classification of hyperspectral remote sensing images.The classification accuracy is greatly improved.Due to the higher dimension of spectral bands,three dimensionality reduction methods are used for classification efficiency.Then,3DCNN＿AM is utilized to extract and classify the features of the reduceddimensional images.At the same time,the loss of information caused by data dimensionality reduction can be effectively compensated.While reducing the calculation time,the classification accuracy is still improved to a certain extent.(2)A multi-scale dense network(MSDN)is developed to realize the feature fusion and classification algorithm of hyperspectral data,and solve the problem of single feature level.Through the combined use of ordinary convolution and stride convolution,a multi-scale dense network named MSDN that can generate multiple hierarchical features and merge features for classification is constructed.The results is shown that the Indian Pines has the highest classification accuracy when its four levels of features are extracted and fused,which is 99.76%;the University of Pavia has the highest classification accuracy when its six levels of features are extracted and fused,which is 99.54%;(3)A hyperspectral image classification method combining deep learning and multi-classifier integration is developed to solve the problem of single classification decision.A classification model based on the integration of 3DCNN and multi-classifiers is built to realize the classification integration decision.Three commonly used classification algorithms SVM(Support Vector Machine),KNN(K-Nearest Neighbor)and RF(Random Forest)are used here for classification prediction,and then an integrated decision is made on a single classification result according to the majority voting strategy.The results show that the KNN classification performance is the best,and the classification effect of the hard voting ensemble is stable.The multi-classifier ensemble method can integrate the advantages and disadvantages of each method to improve the classification stability,but a single factor may not reach the highest level.