Research On Hyperspectral Image Classification Base On Dimensionality-varied Convolutional Neural Network

Hyperspectral image(HSI)is characterized by large amount of data and complex structure.Traditional classification algorithms are unable to meet the demands of practical applications because of the big data and complex structure of HSI,and the classification models based on deep learning require tremendous calculations and lots of training time.Besides,the features of bands caused by water-vapor absorption are unused in classification.In order to address the aforementioned issues,a dimensionality-varied convolutional neural network(DV-CNN)is proposed in this paper.In DV-CNN,3-D convolution kernels extract spectral-spatial features from raw HSI to make full use of HSI information.The 1-D feature vectors generated by feature extraction are fused into a 2-D feature matrix,and more features can be extracted from it for classification.The process of 2-D feature extraction is optimized to improve the accuracy of feature extraction and classification by three optimization algorithms,namely dynamic adaptively pooling algorithm,double-optimization algorithm and adaptively enhanced algorithm.DV-CNN takes raw data as the input to extract features from water-vapor absorption bands to enhance the robustness.In terms of reducing training time,a distributed network with GPU is deployed to improve the average calculation speed and classification efficiency.The experiments showed that DV-CNN obtained higher accuracy than others,which are 99.18% and 99.87% on Indian Pines and Pavia University scene data set respectively.The features extracted from water-vapor absorption bands can improve the accuracy further.Moreover,the parallel pattern of DV-CNN reduces lots of training time.The results of experiments demonstrate that the dimensionality-varied structure of DV-CNN reduces the complexity and calculations of model,and the adequate spectral-spatial feature extraction of DV-CNN from HSI improves the accuracy of classification obviously.In addition,the robustness of classification model for features extracted from water-vapor absorption bands is enhanced and the parallel design of DV-CNN can improve the efficiency of classification effectively.