Hybrid Fusion Feature For Hyperspectral Image Classification

Hyperspectral images play an important role in remote sensing technology.The hyperspectral image not only obtains the spectral information,but also obtains the surface image information and achieves the combination of the spectral information and the image information.As a result,hyperspectral images contain very rich information.Therefore,the development potential of hyperspectral image technology is very huge,and it has a very wide range of applications in the fields of medicine,agronomy,and food inspection.Among them,the classification of hyperspectral images is a major problem in hyperspectral image technology.Due to the characteristics of high dimensions,information redundancy,mixed pixels and uneven distribution of samples,the classification of hyperspectral images has caused difficulties in the classification of hyperspectral images.It has attracted wide attention from scholars at home and abroad.To solve these problems,this paper proposes a new method of hyperspectral classification based on broad learning of fusion features.This method uses an adaptive spatial filter to first extract the spatial spectrum features of the hyperspectral image,and then uses different sizes of convolution kernels to perform deep feature extraction on the extracted spatial spectrum features,and then cascade the extracted features.The broad learning system is used as the classification for classification.The main research contents of this paper are as follows:(1)A new feature extraction method of hyperspectral image is proposed.When substances with different spectral properties appear in the same pixel,the phenomenon of mixed pixels will appear.Therefore,it is obviously not effective to use spectral features for classification.In view of the distribution characteristics of hyperspectral images,we adopt an adaptive distance weighting filter to extract the spatial features of the original hyperspectral images.Then multiple convolution kernels of different sizes are randomly selected from the extracted spatial features.And after repeated extraction and convolution operations,a series of shallow and deep spatial spectrum features obtained are cascaded.(2)Aiming at the problem that the existing traditional classification methods can only extract shallow features,we propose a novel feature extraction method which can extract deep features.We propose directly uses a part of the image as the convolution kernel,which saves a lot of time for extracting features and has superiority in terms of time.Compared with the traditional deep convolutional network classification method,the weight of the convolution kernel of our proposed method does not need to be determined through tedious backpropagation.In addition,feature extraction with different sizes of convolution kernels can obtain different sizes of receptive fields,which improves the robustness of the classification method.(3)The extracted features are sent to the broad learning classifier for feature node mapping and enhancement node mapping,and then the feature nodes and the enhancement nodes are combined,directly connected to the label matrix.And the weight matrix is obtained through the pseudo-inverse for classification.We have obtained satisfactory results in the three publicly available hyperspectral image datasets,which verifies the efficiency and feasibility of our algorithm.