Research On Remote Sensing Hyperspectral Image Classification

With the broader application of hyperspectral sensors,it gets easier for people to obtain various type of hyperspectral images(HSI).HSI have been widely used in many fields,such as land coverage,environmental protection,agriculture,and so on,due to the abundance in spectral and spatial information.HSI classification,as one of the biggest challenges for HSI application,has attracted more and more attention..For example,the number of training samples are relatively small,and the distribution of training samples is uneven.Because the band of hyperspectral image is relatively narrow,there is a strong correlation between its bands.The narrow band of HSI makes the bands have strong correlation and vulnerable to the impact of noises.Furthermore,it is a universal phenomenon that different materials have the same spectrum and the same material has different spectrum,which further increases the difficulty of classification of hyperspectral images.Aiming at the above problem,we study the HSI datasets,and improve the classification performance by using artificial intelligence algorithms.We propose several feature extraction and classification methods of HSI.To verify the effectiveness of the proposed methods,we did the comparative experiments on widely used datasets.The specific research contents are as follows.(1)To address the issue that there is inaccuracies in classification for the KNN algorithm,we propose two HSI classification methods using guided filter and JKNN,which is a method of combining joint representation with KNN.We take the principal components of HSI as guidance image,to preserve the structural feature extremely.By using guided filter to smooth HSI and preserve the gradient of the HSI,we first adopt the guided filter to process the HSI for reducing the intra-class feature differences,and then use the JKNN model to classify the HSI.We call the above method FGF-JKNN.Secondly,we first apply the JKNN model to classify the HSI and obtain the classification map of the HSI.Then we use guided filter to extract the spatial features of HSI for optimizing the classification result of JKNN.We call this method PGF-JKNN.In addition,because different HSI datasets have different spatial resolutions and spectral resolutions,we studied the effects of different filter radius on the classification performance,and processed the hyperspectral images with different filtering radius.The classification performance of hyperspectral images was improved greatly.(2)In view of the strong correlation of spectral information in HSI,the current methods for feature extraction are to study decorrelation.We propose a HSI classification method using long short term memory network(LSTM)and guided filter.This method regards the spectral curves of HSI as time series,and classifies the hyperspectral image by using the correlation of spectral features and LSTM.In order to solve the problem of unbalanced distribution of samples in HSI,we use weighted cost function to improve LSTM model,and obtain weights through training and learning,thus improving the classification performance of HSI.In order to increase the stability of LSTM method and reduce the intra-class differences of HSI,we introduced a guided filter to smooth hyperspectral images,which further improves the classification performance of HSI.(3)Because the spatial resolution and spectral resolution of HSI are different in the datasets,the method based on single scale spatial characteristics cannot be suitable for different hyperspectral image datasets.In order to further mine spatial features and improve the classification performance of HSI,a deep convolutional neural networks(CNN)method based on multi-scale spatial features is proposed for HSI in this paper.Specifically,we used the principal component analysis(PC A)method to reduce the dimension of HSI.Then,the reduced HSI is extracted at different scales,and the spatial and spectral features of the HSI are fused.Finally,the fused HSI is reconstructed,and classified by CNN.We call this method MSCNN.The experiments show that MSCNN method can greatly improve the classification performance of HSI.MSCNN method is robust and is suitable for all the datasets.In addition,the impact of Dropout and regularization strategies have on the classification performance of HSI are also analyzed.Besides MSCNN method,we propose several other fusion method of spectral and spatial feature for HSI classification.By comparing the spectral and spatial fusion methods of HSI,we find that CNN methods also have dimension disasterand information concentration issues.If the dimensions are the same,increasing information concentration can improve the accuracy of spectral image classification.(4)To address the issue that manual labeling in HSI is difficult,an unsupervised HSI classification method based on maximum and minimum distance embedding is proposed by combining deep learning with clustering algorithm.Firstly,multi-scale spatial features are used to enhance the discrimination of HSI.Secondly,feature extraction and dimensionality reduction are realized by deep autoEncoder based on maximum and minimum distance embedding.Thirdly,k-means clustering algorithm is used to obtain the results,and the guided filter is used to optimize the results.Finally,the Hungarian algorithm is used to evaluate the results.The experimental results show that the proposed method can improve classification performance.