Research On Hyperspectral Image Classification Method Based On Two-Branch Spatial Spectral Feature Fusion

Hyperspectral Image(HSI)is a three-dimensional stereoscopic image acquired by an aerospace vehicle equipped with a hyperspectral imager.Hyperspectral images can provide richer spectral information than traditional color images,which is very valuable for object classification and recognition.Due to this special data feature,HSI has a wide range of applications in military target detection,mineral exploration,environmental monitoring,and agricultural production.Hyperspectral remote sensing data has become an important tool for ground object recognition and classification due to its high spectral resolution,continuous spectral curve,and rich details.In recent years,deep learning techniques have made remarkable progress in hyperspectral image classification.Researchers are working on designing new network architectures,optimizing algorithms,and developing effective feature extraction methods to improve classification accuracy and efficiency.However,due to the data redundancy generated by a large number of bands,it will cause difficulties in feature extraction,and there are also problems such as "same object with different spectrum" and "different object with same spectrum".The development of hyperspectral image classification mainly depends on the research of algorithms and models.Therefore,it is a research task with scientific research value and socio-economic significance to dig out the features contained in hyperspectral images and improve the accuracy of ground cover recognition,classification and mapping.Thesis studies how to extract and fuse spatial spectral features of hyperspectral images,focusing on hyperspectral spatial feature extraction and spectral feature extraction and fusion methods to classify hyperspectral images.1.Aiming at the problem of adjacent spectral similarity and dissimilarity of hyperspectral images,thesis proposes a two-branch joint network hyperspectral image classification model based on Gaussian pyramid multiscale and wavelet transform,using Gaussian pyramid multiscale transform to get multiscale image data and then feature multiscale spatial features.In addition,the wavelet transform is used to process the spectral data to reduce the influence of its anomalous spectral data on the classification,and the feature extraction module is used to extract the spectral features.Finally,the acquired spectral features and spatial features are fused in the fully connected layer to combine spatial and spectral features to obtain the classification results.In view of the fact that the previous method has anisotropy in spatial feature extraction and cannot extract spatial features in different directions in different regions,thesis proposes a method based on multi-region multi-scale feature extraction and spectral imaging,and uses PCA(Principal Component Analysis)algorithm for dimensionality reduction to solve the problem of data redundancy.For spatial features,this method adopts multi-region segmentation Gaussian pyramid down-sampling method to generate multi-scale data,and uses an improved residual network to extract spatial features.For spectral features,thesis expands the spectral data into an N×N image,and then uses the improved residual network to extract spectral information,but the number of network layers is different from that of the spatial feature extraction model,and the image scale corresponds to the number of layers.Finally,the spatial features and spectral features are fused to obtain classification results.Experiments on several commonly used data sets in thesis show that the method proposed in thesis can improve the accuracy of the classifier,thus confirming that the dualbranch spatial spectral feature fusion method proposed in thesis will promote the development of remote sensing data classification.