Adaptive Spatial-Spectral Weighting Base On Convolutional Neural Network Research For Hyperspectral Image Classification

In recent years,the classification of hyperspectral remote sensing image（HSI）has gradually become an important topic in the field of remote sensing.Because of its wide application in precision agriculture,military recognition,environmental monitoring and other fields,it is great significance to improve the classification performance of remote sensing image and promote the development of hyperspectral remote sensing image classification technology.The feature extraction is the most critical step in the classification task.However,in the traditional classification algorithm,it often faces many problems,such as low abstraction,poor generalization ability,large amount of calculation and so on.In recent years,hyperspectral remote sensing image classification algorithm based on convolution neural network has made great achievements in deep feature extraction,which effectively alleviates the problems faced by traditional algorithms.Meanwhile,there is still broad research space in spectral feature extraction and spatial feature extraction.On the one hand,not all bands have the same amount of information and classification contribution in the high-dimensional band data of hyperspectral image.In many practical applications,efficient feature extraction and classification algorithms are needed.On the other hand,the spatial information of hyperspectral image in different positions also has different contribution to the classification results in feature extraction.In order to solve the above two problems,this paper carried out the research and design of spectral band weighting and spatial weighting model based on convolutional neural network.The main work is as follows:（1）Aiming at the unequal contribution of spectral band information in hyperspectral image classification task,a lightweight compact band weighting module（CBW）based on convolutional neural network is proposed.By capturing and gathering the correlation between adjacent bands,CBW module recalibrates the spectral band information,enhances the high contribution band information and suppresses the low contribution band information,so as to improve the accuracy of hyperspectral image classification.Through the optimization design,the number of adjustable parameters required by the module is only 20,which is far less than the number of parameters of the existing related modules.It greatly improves the calculation speed,and has a more obvious classification accuracy improvement effect compared with other band weighting algorithms.（2）In view of the high spatial resolution of hyperspectral remote sensing images and the existing spatial weighting algorithms ignore feature extraction in different scale sensing fields.The module uses the strategy of multi-scale feature fusion to fuse the spatial correlation features of different scales.In the process of extracting the spatial scale features,multi-level fusion strategy is used to extract the multi-scale and multi depth spatial features,so as to improve the positive impact of spatial weighting on feature extraction.Through the experimental verification on three datasets,the module has better spatial weighting effect and can significantly improve the classification accuracy of the classification network.（3）From the perspective of simultaneous enhancement,combining the above lightweight band weighting structure and multi-scale spatial weighting function,an adaptive space spectrum weighting module is proposed.The module has two functions:band weighting and space weighting,which can be used to adjust the characteristics from the spectral and spatial perspectives,and has the function of adaptive adjustment in the whole process.The experimental results on three datasets show that the adaptive spatial spectrum weighting module is superior to other related weighting algorithms in accuracy improvement.In this paper,a lightweight compact band weighting module based on convolutional neural network,a multi-scale spatial weighting module based on convolutional neural network and an adaptive spatial weighting module are proposed from three aspects of band weighting,spatial weighting and spatial spectrum weighting respectively.The performance of the algorithm is proved by experiments,it provides a new idea for the research of hyperspectral remote sensing image classification algorithm.