Research On Hyperspectral Image Classification Based On Convolutional Neural Network

Hyperspectral images(HSIs)have hundreds of spectral bands.Such abundant spectral information enables them to discriminate more land cover materials under various conditions,facilitating a wide range of applications,including environmental monitoring,resource assessment,urban development monitoring,etc.Classification is one of the important tasks for these applications.Feature extraction is of significance for HSI classifification.Handcraft-feature-based methods can only extract low-level features,and rely on the prior knowledge of designer,which is time-consuming and labor-intensive.In contrast,deep learning algorithms can automatically learn hierarchical features with discriminative information.Among deep learning algorithms,convolutional neural network(CNN)generally outperforms others in feature extraction,mainly because its local connections and shared weights characteristics enable it to maintain the original structure while learning spatial features,and greatly reduce the number of network parameters.These unique characteristics make the CNN-based methods valuable in spectralspatial classification of HSI.However,CNN requires fixed-size windows as input,which makes it difficult to represent land covers at multiscale.Secondly,the classification maps obtained by CNN-based methods often result in “salt and pepper” noises.Moreover,the existing CNNbased spectral-spatial classification methods have not simultaneously made the best use of the rich spectral information and multi-scale contextual information of HSI,especially for representing the scale variability of complex objects.To sovle the above problems,this article mainly made the following research:(1)Land covers in HSIs usually appear at different scales,while CNN requires fixed-size windows as input,which makes it difficult to describe these land covers.Besides,the classification maps abtained by CNN often result in “salt and pepper” noises.To resolve these problems,we propose a multiscale CNN combining with a region-based max-voting scheme for HSI classification.First,a multiscale CNN is proposed to extract multiscale features for HSI classification,and then a region-based max voting scheme is applied to the classification map to solve the “salt and pepper” noises.Experiments on two classical data sets demonstrate that the proposed method is effective for HSI classification.(2)The exisiting CNN-based methods for classification of HSI still have not simultaneously made the best use of the rich spectral information of his,especially for representing the scale variability of complex objects.Particularly,the spectral bands are usually equally treated when extracting spectral and spatial information,ignoring the variance and correlation difference between bands.To tackle such drawbacks,we develop a new CNN framework to combine the spectral and spatial information by considering the correlations between different bands.Firstly,to cluster the continuous bands of the HSI into several band groups based on their similarity measurements.Secondly,apply two strategies,localized spectral feature extraction(LSF)and hierarchical atrous spatial pyramid pooling(Hi ASPP)to extract the spectral and spatial features in parallel from the determined band groups.Finally,concatenate the extracted spectral and spatial features and feed into a fully connected layer for data classification.Experiments demonstrated that the proposed architecture outperforms several state-of-the-art methods.(3)The convolution filters of the traditional CNN treat the input content equally and only modele local features.Generally,spectral and spatial features extracted from the input have different contributions to classification.To investigate this opportunity for better HSI classification,we proposed a spectral-spatial self-attention network with two subnetworks,designed for spectral and spatial feature extraction,respectively.Specifically,the spatial subnetwork introduces the proposed spatial self-attention module to capture the spatial feature correlations between the center pixel and its surroundings.Meanwhile,the spectral subnetwork introduces the proposed spectral self-attention module to exploit the long-range correlations over local spectral features.The “score weighted” fusion method is then used to fuse the extracted spatial and spectral features for classification.Experiments show that the proposed spatial attention model and spectral self-attention module can significantly improve the discriminative features related to the pixels to be classified,improving the performance of classification.