Research On Feature Selection And Feature Extraction Of Hyperspectral Remote Sensing Images Based On Convolutional Neural Network

Hyperspectral imaging provides abundant spatial distribution information and spectral reflection information of land-cover objects at the same time,which greatly enhances the remote sensing earth observation capabilities.Thus,hyperspectral image(HSI)has been widely used in many fields,e.g.,precision agriculture,geological exploration,urban planning,and military defense,and it has become a scientific and technological frontier area developed by different countries around the world.Along with the rapid improvement in the spatial and spectral resolution of hyperspectral imaging,both opportunities and challenges are brought to hyperspectral image classification.On the one hand,high-dimensional hyperspectral images always contain hundreds of bands and different bands have high redundancy,easily leading to the ”curse of dimensionality” issue for the HSI classification task.On the other hand,the spatial-spectral structure of hyperspectral images is too complex,which makes it difficult for traditional artificial design-based feature extraction methods to adaptively learn the highly discriminative spatialspectral features for accurate classification of HSI.Thus,how to achieve low-dimensional feature selection of low redundancy and extract spectral-spatial features of high discrimination are the key to improvement of classification accuracy,which are important scientific problems that should be solved in the field of hyperspectral remote sensing.Based on the in-depth summary and analysis of the research status,this thesis focuses on the difficult problems caused by the high spectral dimension and complex spatial-spectral structure.In brief,this thesis combines theoretical knowledge of deep learning,and proposes novel feature selection and feature extraction methods for HSI based on convolutional neural networks.As a result,the proposed methods realize the optimal selection of low-dimensional band subsets with spatial-spectral structure preservation and improve the discrimination capability of spatial-spectral features.Moreover,a hyperspectral image spatial-spectral feature learning software is developed here.The specific research contents are as follows:(1)Focusing on the ”curse of dimensionality” caused by the high dimension and redundant information of HSI,an unsupervised knowledge distillation convolutional neural network feature selection method is proposed.This method learns more effective image feature representation by introducing teacher network with more complex structure and better performance.Then,the teacher network guides the training of the student network with simple structure and lower complexity,and the error is easier to back-propagate.As a result,the method realizes the optimal low-dimensional band subsets selection while the spatial-spectral structure information can be preserved.Experiments are carried out on two hyperspectral datasets and compared with several classical feature selection algorithms,the results show that the proposed method selects a better subset of low-dimensional feature bands,especially when the number of selected bands is small.(2)Focusing on the challenge of complex spatial-spectral structure and the difficulty of adaptively learning highly discriminative spatial-spectral features,a contrastive learning-guided convolutional neural network spatial-spectral feature extraction method is proposed.First,the method obtains discriminative spectral features with intra-class compactness and inter-class dispersion by the spectral contrastive learning model.Then,the spectral features guide the multi-layer spatial feature fusion process to extract more discriminative spatial-spectral features.In addition,the pair-wise learning strategy is used instead of the pixel-wise learning strategy,which greatly improves the ability to learn detailed information.Experiments are carried out on three hyperspectral datasets and compared with several state-of-the-art feature extraction methods.The proposed method has significant advantages in objective evaluation metrics,which proves that the proposed method can extract more discriminative spatial-spectral features to obtain better classification results.(3)Based on the feature selection and feature extraction method proposed in this thesis,an integrated spatial-spectral feature learning software for the hyperspectral image is developed.The software is implemented in the Python3.9 development environment with the Py Qt tool.Specifically,it includes six functional modules: data loading,data preprocessing,feature selection,feature extraction,classification and parameter setting.Furthermore,the effectiveness of the proposed methods is verified by applying the software to the Yellow River Delta wetland HSI dataset.