Hyperspectral Image Classification Based On Deep Learning

Hyperspectral remote sensing images contain rich spectral and spatial information,which is helpful for the accurate classification of different ground objects.Therefore,hyperspectral remote sensing images play an important role in military and civil fields.However,the high-dimensional spectral band information also brings a lot of redundant information,which is prone to "dimension disaster" when using the traditional machine method to classify spectral features,resulting in the performance degradation of the model.In addition,the small number of labeled samples and uneven distribution of remote sensing images also bring significant challenges to image classification.Deep learning has a better capability of deep feature extraction and nonlinear modeling.To solve the above problems,this paper proposes a hyperspectral image classification method based on deep learning.The specific research works are as follows:(1)A hyperspectral image(HSI)classification method based on stack contractive autoencoder and adaptive spatial-spectral information was proposed.Firstly,a non-subsampled shear wave combined with guided filter(NG)was used to enhance the dimensionality reduction of the HSI.Then,the spatial information of the pixel was extracted by an adaptive method according to the characteristics of the pixel.Finally,the deep features of the pixel fused with spatial-spectral information were extracted by stacked contractive autoencoder(SCAE)which composed of several contractive autoencoders(CAE),and the category of the pixels was judged by logistic regression(LR)relying on the extracted deep features of the pixels.Compared the proposed method with RBF-SVM,SDAE,M3D-CNN,and DFFN on Indian Pines and Pavia University datasets.The results show that the OA of proposed method on the two data sets are 97.133% and 99.201%,respectively.The proposed method has higher classification accuracy and better model performance,and can be used for HSI classification.(2)A hyperspectral image classification method based on a multi-scale 3D convolution mix attention mechanism network was proposed.Firstly,the feature images of spectral and spatial were processed by 3D convolution with different scales.The obtained multi-scale features of images were sent into the dense network to extract more robust features.To extract more discriminant features,the fused attention mechanism module was applied to spectral and spatial features,respectively.Finally,feature-level fusion and decision-level weighted fusion are performed on the multi-scale spectral and multi-scale spatial features to obtain the final classification result.The performance of the method was evaluated on Indian Pines and Pavia University datasets,and the OA of the method are 99.032% and99.641%,respectively.The results show that the proposed method can achieve better classification accuracy and visualization effect.