Hyperspectral Images Dimension Reduction Based On Super-pixel And Kernel Partial Least Square

Hyperspectral image(HIS)is an imaging within a certain spectral range,generally containing hundreds of spectral bands.The spectral information and spatial information of HSI can be used to accurately identify the target.In recent years,with the development of HSI technology,it has not only been applied to remote sensing of the earth,but also made great progress in food health.However,there are many spectral bands of HSI data,resulting in the information redundancy,which will affect the subsequent data processing speed and the final classification accuracy.Therefore,it is a significant research to reduce the dimension of hyperspectral data on the premise of preserving the effective information of hyperspectral images as much as possible.Band selection and feature extraction are common methods of data dimensionality reduction in HSI.Band selection is to select the most effective set of several bands from the original band,and the data after dimension reduction in this way will not break the physical characteristics of the original data.The feature extraction method is to find a better data expression for the original data by dimension reduction after the transformation of the original data.The data after dimension reduction will change the physical characteristics of the original data.In this paper,a new method framework based on band selection is proposed for the detection of early hidden damage in kiwi fruits.On the common data sets of remote sensing image,a new data dimension reduction method is proposed based on the small sample training set that has been labeled.(1)Nondestructive testing of kiwi fruit's early invisible damage based on kernel partial least squares method.The application of hyperspectral technology in early invisible damage detection of kiwi fruit is less than the introduction of band selection based on small samples,and the classification of kiwi fruit is mostly from the aspect of image processing.In the gene sequence data,there are very few samples and a lot of features,the kernel partial least squares(K-PLS)method is very advantageous in the feature selection of the gene data,therefore,in this paper,K-PLS is used to select the bands when the number of kiwi samples is limited.In this paper,hyperspectral images of sound kiwifruit and damaged kiwifruit were collected in the spectral range of 900nm-1700 nm,and some effective bands were extracted by K-PLS.Then,principal component analysis(PCA)was used to reduce the dimension of the extracted bands.Finally,machine learning was used to classify the data.(2)Multi-kernel PCA dimension reduction method based on super pixel segmentation.In this paper,the super pixels segmentation of HIS was firstly carried out to obtain many locally homogeneous regions,then the local dimension reduction was carried out in each region,and finally the data of each part was integrated as the data set for subsequent processing.The algorithm firstly uses PCA to compress the HSI,selects the first principal component image,and then uses the entropy rate-based segmentation method(ERS)to segment the image and obtain the segmentation graph.Original data was divided into multiple regions by the segmentation image,and in each region,the data were projected and transformed into a new space with multi-kernel function,and the transformed data are further reduced dimension by PCA.Finally,all local data are integrated.The two methods proposed in this paper have achieved good results,indicating that the dimensionality reduction technique of hyperspectral data is still an important technique in practical application and theoretical research.