Research On Small-sized Labeled Samples Of Hyperspectral Image Classification Based On Convolutional Neural Network

Remote sensing hyperspectral data contains abundant spatial-spectral information,which has the typical characteristics of "atlas in one" and the characteristics of large number of bands,fewer label samples and large spatial variability.Convolutional neural network(CNN)is a deep learning model with strong feature extraction and expression ability,which shows remarkable characteristics superior to other classification models in remote sensing hyperspectral image classification.However,because of the high-dimensional and redundant characteristics of remote sensing hyperspectral data,the classification accuracy of convolutional neural network is often affected,which makes the model rely heavily on label data,and makes the training time too long,which leads to the low efficiency of the model.In addition,many classification models based on convolution neural network only consider the feature extraction of spectral dimension,ignoring the enhancement of spatial information on classification performance.In order to solve the above problems,a dimensionality-varied convolutional neural network spectral-spatial classification model for small samples of remote sensing hyperspectral data is proposed in this paper.Dimensionality-varied convolution neural network is an improved model based on convolution neural network.The optimization of three-dimensional and two-dimensional convolution feature extraction structure is the key to improve the classification performance of small samples of hyperspectral images.Dimensionality-varied convolution neural network can be divided into spectral-spatial information fusion,dimension reduction of feature map,spectral-spatial feature extraction and feature classification according to the change of dimension of feature map,which ensures that the small samples of data of network has strong feature extraction ability.In the process of feature extraction,dimensionality-varied convolution neural network reduces a lot of computation and saves storage space by transforming the dimension of three-dimensional feature graph in data stream.This dimensionality-varied structure can simplify the network structure and reduce the computational complexity.At the same time,it can fully extract the spectral-spatial features and improve the accuracy of convolutional neural network for small samples of hyperspectral image classification.The experiment is divided into performance analysis experiment,classification performance comparison experiment and parallel dimensionality-varied convolution neural network acceleration experiment.The data sets used are Indian Pines and Pavia University remote sensing hyperspectral data sets.The experimental results show that the dimensionality-varied convolution neural network has high classification accuracy for small samples of remote sensing hyperspectral images.The selection of parameters has an important influence on the classification performance of dimensionality-varied convolution neural network,and reasonable parameters can make the classification model achieve better performance;the overall classification accuracy of dimensionality-varied convolution neural network on Indian Pines and Pavia University datasets is 87.87% and 98.18%,respectively,which has obvious performance advantages compared with other classification algorithms;parallel dimensionality-varied convolution neural network can greatly reduce training time.The experimental results show that the complexity of the model is reduced by changing the dimension of the feature map and extracting the spectral-spatial features with high accuracy in the process of feature extraction,and the efficiency of the model can be significantly improved by parallel classification.This dimensionality-varied model can greatly improve the classification performance of small samples of hyperspectral images,and can be further extended to other deep learning classification models related to hyperspectral images.