Convolutional Neural Learning Representation Framework Based Hyperspectral Image Classification

Remote sensing hyperspectral images record both the spectral information and the spatial information of the ground targets at the same time.In recent years,the classification of hyperspectral images has become a hot topic in the field of remote sensing data processing.With the development of the Earth observation technology,the spectral and spatial resolution of hyperspectral images have become increasingly high.However,due to the complex spatial structure and high-dimensional spectral signature,traditional methods unable to effectively classify hyperspectral images.Recently,deep learning has become the mainstream algorithm for hyperspectral classification because of its powerful feature extraction ability.In addition,convolutional neural networks have attracted more and more attention due to its weight sharing,local connection and hierarchical feature extraction properties.In this paper,considering the special properties of remote sensing hyperspectral images,we propose several convolutional neural learning representation frameworks to extract the discriminative spectral-spatial features,achieving accurate hyperspectral classification.The main research contents are summarized as follows:1)A multipath residual network is proposed for spatial-spectral hyperspectral image classification,which is wider than other existing deep learning based hyperspectral image classification models.Considering that very deep residual network behaves like ensemble of relatively shallow networks,the proposed multipath residual network adopts multiple parallel residual functions in each residual block,so as to increase the network width,rather than depth.As a result,the proposed network is made up of shorter-medium paths,achieving efficient flow of gradient during the training phase.Finally,the advantages of the proposed method in accurate classification and efficient parameter utilization are proved in comparison with several state-of-the-art hyperspectral image classification methods.2)The convolution neural network can automatically extract hierarchical feature representations from the original data,which has been widely used in the community of hyperspectral image classification.However,most convolutional neural network based classification methods neglect making full use of the complementary and related information in each convolution layer,and only use the features extracted from the last convolution layer for classification.In this work,we propose a fully dense multiscale fusion network,which can make full use of the hierarchical features from all convolutional layer for hyperspectral classification.In the proposed network,shortcut connections are introduced between any two layers in a feed-forward manner,so that the learned features of each layer can be directly accessed by all subsequent layers.This fully dense connectivity pattern achieves comprehensive feature reuse,which is conducive to discriminative feature learning.Finally,multi-scale spatial-spectral features extracted by all convolutional layers are fused for classification.Experimental results on four widely used hyperspectral scenes show that our method is superior to other compared feature fusion based methods.3)A novel deep mixed link network is proposed for hyperspectral image classification,which further enhances the feature representation ability of the convolutional neural network.The proposed network integrates residual connections and dense connections through mixed link architecture,which makes it have the feature reusability of residual connection and the new feature exploration ability of dense connection.Compared with the dual path architecture,the proposed mixed architecture can further improve the information flow in the whole network.In addition,the residual networks,densely connected convolutional neural networks and dual path networks can be regarded as special cases of the proposed mixed link networks.Finally,experimental results on four widely used hyperspectral scenes demonstrate the robustness and effectiveness of the proposed approach.4)We propose an attention guided progressive fusion network for hyperspectral image classification.Specifically,multi-level spectral-spatial features are first extracted by the multipath residual network,and these features are then fused by two attention guided feature fusion modules.In the fusion stage,the style-based attention module can adaptively enhance the useful information and suppress the useless ones,so as to enhance the discrimination of the fused features.The proposed network can effectively improve the classification accuracy of the baseline multipath residual network,with only about 0.2% additional parameters.The classification results on three real hyperspectral scenes show that the proposed method performs better than several state-of-the-art classification approaches.In addition,compared with the feature fusion methods like addition or concatenation,our attention based feature fusion module can improve the baseline model's classification performance more robustly.5)Patch-based spatial-spectral hyperspectral classification methods usually achieve overoptimistic classification results,since traditional random split of training and test sets results in data leakage.To solve this problem,we propose a controlled random patch sampling method,which can automatically select training/test areas based on the number of labeled classes and the total number of training samples.Next,training and test image patches can only be extracted from the training areas and test areas,respectively,avoiding possible data leaks between training and test sets.Experimental results on four real hyperspectral scenes demonstrate that the proposed training-test split strategy makes the task of hyperspectral classification more challenging and realistic.