Research On Band Selection Method Of Hyperspectral Image For Target Detection

Compared with ordinary images,hyperspectral remote sensing images contain a large amount of spectral information,and is widely used in the fields of military hidden target detection,weather forecasting and geological survey.However,due to its high dimension and large amount of data,it is difficult to process real-time information,so it is necessary to perform dimensionality reduction operations first.Band selection is a mainstream dimensionality reduction method which selects a subset from the original band set.This method can retain the original features of the image so that it’s widely used.In addition,the current main band selection methods are mainly aimed at classification tasks,and there are relatively few researches on object detection.Aiming at the above problems,the approach is to use subspace division and attention mechanism to make full use of the order and correlation of the bands to improve the performance of target detection and classification.The main research contents and work include:Based on the high correlation of spectral bands of hyperspectral images,subspace division of hyperspectral images is performed to select bands.This method takes advantage of the feature that adjacent bands have more redundant information,uses relative entropy to construct a divergence matrix to divide the image into subspaces and then uses the interclass separability factor index to select bands in each subspace.The method balances the diversity and importance of the selected bands.Compared with other band selection methods,SD-ISBS method is more reasonable to divide the subspace based on the divergence matrix for band selection,and the subsequent target detection tasks also achieves better performance.Using the deep learning method combined with the attention mechanism for band selection,the optimal band subset can be screened according to the different network training tasks.For band selection,the purpose is to focus on more important spectral channels and emphasize key feature information.Therefore,an attention module is embedded in the network to generate an attention heatmap to measure the importance of each band,so that the model can focus on more useful image features.Combined with the training of the network model,the importance of each band in the input hyperspectral image is calibrated.The important bands of the image are selected according to the weight after the training is completed,so as to realize the band selection of the hyperspectral image.The AMBS method can pay more attention to the extraction of effective features,which enables the network to achieve better performance in hyperspectral image target detection.We conducted experiments on several hyperspectral datasets: San Diego,Indian Pines,Avon and Salinas Valley,respectively.The results show improved detection and classification performance compared to existing methods,verifying the effectiveness of the proposed methods.