A Study Of Spatial-spectral Hyperspectral Image Classification Based On Regional Structure

Recent advances in hyperspectral remote sensor technology allow the simultaneous acquisition of hundreds of spectral wavelengths for each image pixel.Hyperspectral Image Classification(HIC)means using detailed spectral information to discriminate materials of interest.In the literature,many of the methods have concentrated on exploring the role of the spectral signatures of hyperspectral data employing exclusively the spectrum of a pixel to determine its class belonging.However,there are two dilemmas for such pixel-wise approach: the relatively small training set versus the high-dimensional spectra and spectral variabilities.Hyperspectral image is more like an image which has certain local spatial structures rather than a collection of independent spectral pixels who do not interact with one another.Homogeneous region is an adaptive spatial shape structure.We believe pixels in a homogenous region is likely to belonging to same material.Based on this assumption,Region structure is applied to spatial-spectral classification.The main works and innovations are as follows:A region regularized spatial-spectral nearest neighbor classifier is designed.Region distance can adaptive measure distance between pixels which may be far from each other but belonging to same homogenous region.As Euler distance lacks region discrimination,region distance is added to fit the joint probability distribution.The experimental result of five hyperspectral datasets shows that the region regularized spatial-spectral nearest neighbor classifier shows better classification accuracy.A region regularized probabilistic collaborative representation classifier is designed.Region information is incorporated by including a spatial weight to control the coefficients.Region regularized probabilistic collaborative representation classifier combines the spatial and regional regularization together through the weight sun of a spatial regularization term and regional regularization term.The representation coefficients of pixels in the same superpixel are supposed to be similar,while those coefficients that do not belong to the same area are penalized,so that the entire classification result has good spatial consistency and border preserving ability.The classifier shows better classification on five hyperspectral datasets.A deep homogenous region structure feature extraction model is proposed.In the face of,most of the existing homogenous region that extracted are only some small and discrete area.Deep region appearance features is extracted by a pre-trained semantic segmentation.The semantic segmentation model is transfer to hyperspectral classification task.The well pretrained Mobile Net is beneficial to achieve the regional distribution prediction for HSI after a pretreatment of dimensionality reduction.However,due to the differences between HSI and the natural image data sets,it is difficult to reserve the high-resolution spectral information,which is its unique advantage for target detection.Therefore,in this paper,the learned deep feature integrated with the spectral feature is utilized for the final HSI.The classifier shows better classification on three hyperspectral datasets.