Research On Hyperspectral Remote Sensing Image Classification Algorithms Based On Convolutional Neural Network

Owing to the development of imaging spectrometers and the improvement of spectral and spatial technology,remote sensing detection has entered the hyperspectral stage.Hyperspectral imaging organically combines subdivisional spectroscopy technology and two-dimension imaging technology.The hyperspectral remote sensing image obtained is no longer a traditional two-dimensional image,but a threedimensional data cube containing abundant spectral dimensional features and luxuriant spatial information of ground objects.The data cube has the characteristics of high spectral resolution,high spatial resolution and realizes the true “unification of image and spectral”,so it is often used to classify ground objects.Hyperspectral remote sensing image has the strong practical significance and is widely used in military defense,resource exploration,medical diagnosis,precision agriculture and so on.Traditional hyperspectral remote sensing image classification methods are gradually becoming mature,but these methods need to rely on manual extracted features and prior knowledge and cannot effectively interpret high-dimensional spectral information and complex spatial structure,while weak generalization ability and limited representation ability are adverse to the high requirements of classification task.In recent years,deep learning can automatically capture the high-level semantic information of images by simulating the hierarchical structure of human visual system,which has been introduced into the field of hyperspectral remote sensing image classification by multitudinous researchers and has made great progress.Among many deep learning models,convolutional neural network(CNN)has become a mainstream algorithm in hyperspectral remote sensing image classification due to its weight sharing and local connection properties.Considering the special properties of hyperspectral remote sensing image,this paper combines spectral dimension information and spatial dimension information to establish a representation framework conforming to data characteristics and captures more discriminative and expressive spectral-spatial joint features to achieve accurate identification of ground objects category attributes.This thesis proposes three different convolutional neural networks for hyperspectral remote sensing image classification.In order to tackle the problems that spectral-spatial features cannot be extracted fully,the intimate interdependency between spatial features and spectral features cannot be captured and there is a lot of redundant information in the training process for various existing classification methods,a novel 2D-3D CNN with spectral-spatial multiscale feature fusion for hyperspectral remote sensing image classification is proposed.In order to solve the problems that the loss of spectral-spatial features with the deepening of the model,the category boundaries are easily disturbed by noisy pixels and the data imbalance between different sample categories in various existing classification methods,a discriminative spectral-spatial-semantic feature network based on shuffle and frequency attention mechanisms for hyperspectral remote sensing image classification is proposed.In order to deal with the problem that many existing classification methods only capture first-order spectral-spatial features and rarely consider secondorder spectral-spatial representation,a hybrid-order spectral-spatial feature network for hyperspectral remote sensing image classification is proposed.The main research contents and innovation points of this thesis are introduced as follows:(1)A novel 2D-3D CNN with spectral-spatial multiscale feature fusion for hyperspectral remote sensing image classification is constructed.In the network,the spectral feature extraction stream is devised to capture effectively multiscale spectral features,while emphasizing important spectral bands and suppressing useless spectral bands.The spatial feature extraction stream is designed to extract multi-level spatial features,while focusing on vital informational regions and obtaining the relationship of spatial pixels.The spectral-spatial-semantic multiscale feature fusion module is built to not only capture the correlation of spectral features and spatial features,but also effectively aggregate multiscale spectral features,multi-level spatial features and high-level semantic features,outputting more complete and abstract multiscale fusion features for classification.The classification scheme is proposed to control adaptively fusion weights and avoid the overfitting problem,further improve the classification accuracy.The network is compared with other methods of many types on three hyperspectral remote sensing image benchmark datasets(Indian Pines,Pavia University and Salinas-A Scene)and the experimental results show that the effectiveness and superiority of the network.(2)A discriminative spectral-spatial-semantic feature network based on shuffle and frequency attention mechanisms for hyperspectral remote sensing image classification is constructed.In the network,the spectral-spatial shuffle attention module is devised to capture local and global spectral and spatial independent features,while aggregating effectively the large short-range correlation of spectral-spatial features and modelling the large long-range interdependency of spectral-spatial features.The context-aware high-level spectral-spatial feature extraction module is built to fully integrate the spectral-spatial features of different scales and different subnetworks repeatedly to obtain multiscale context awareness spectral-spatial features,while maintaining the scale invariance of each subnetwork and the high-resolution characterization of spectral-spatial features in the whole working process.The spectral-spatial frequency attention module is designed to compress the spectral channels and introduces multiple frequency components to enrich the diversity of spectral-spatial features.The cross-connected semantic feature extraction module is proposed to make full use of spectral-spatial frequency attention features to weight spectral-spatial shuffle attention features,outputting more complete and abstract semantic features for classification.Meanwhile excavating the global information of spectral-spatial frequency attention features to effectively suppress noise pixels and restore class boundaries,further improve the classification accuracy.The network is compared with other methods of many types on four hyperspectral remote sensing image benchmark datasets(Salinas-A Scene,Kennedy Space Center,Indian Pines and Salinas)and the experimental results show that the classification ability of the network has reached an advanced level.(3)A hybrid-order spectral-spatial feature network for hyperspectral remote sensing image classification is constructed.The precedent feature extraction module is devised to effectively aggregate the spectral-spatial of different scales,different convolutional layers and different branches and focus on adaptively recalibrating the channel-wise and spatial-wise feature responses to achieve first-order spectral-spatial feature distillation.The feature rethinking module is designed to heighten the representation ability of hyperspectral remote sensing image by capturing the importance of cross-dimension,while learning more discriminative second-order spectral-spatial representations by exploiting the second-order statistics of hyperspectral remote sensing image,thereby improving the classification performance.The network is compared with other methods of many types on four hyperspectral remote sensing image benchmark datasets(Pavia University,Kennedy Space Center,Indian Pines and Salinas)and the experimental results show that the advancement and superiority of the network.