Research On Hyperspectral Multi-label Classification Algorithm Based On Label Related Features

Hyperspectral image classification is an important research field in hyperspectral image analysis.On the basis of considering the spectral and spatial information of hyperspectral images,many excellent algorithms have been proposed and applied to the classification.A large number of mixed pixels in remote sensing hyperspectral images weaken the separability.Current hyperspectral classifications are single-label classifications.From the perspective of label,it is not appropriate to use single label to mark multiple ground objects in a mixed pixel.From the perspective of classification,mixing will increase the intra-class differences and decrease the inter-class differences in spectral feature space,which will lead to the final classification result worse.Aiming at the existence of mixed pixels,this dissertation applies the idea of multi-label learning to hyperspectral classification,so that one pixel can have a group of appropriate labels.The research on hyperspectral multi-label classification in this dissertation mainly includes three aspects:(1)Three hyperspectral multi-label datasets were produced.At present,there is no hyperspectral multi-label classification data set,but there are common hyperspectral unmixing data sets,which contain endmembers and the corresponding abundance information.This is completely consistent with the physical mechanism of hyperspectral mixed pixels.Therefore,based on three hyperspectral unmixing datasets of Samson,Jasper Ridge and Urban,three corresponding hyperspectral multi-label datasets were produced in this dissertation.(2)A hyperspectral multi-label classification algorithm based on label specific features is proposed.Pixels of hyperspectral images often contain a variety of ground objects and spectral curves are weak in separability.To solve this problem,this dissertation divides the training instances into positive and negative instance sets according to the original label for k-means clustering respectively,and then uses the clustering centers and instance set to obtain the label specific features through SAM-ED mapping,which are used to complete the classification task.(3)A new feature construction algorithm based on sample richness and inter-class relationships is proposed.Hyperspectral classification data sets often show inter-class sample imbalance and intra-class spectral diversity,and the k-means clustering for positive and negative instances respectively does not consider inter-class relationships.To solve this problem,SMOTE oversampling is adopted to enrich the spectral curves,and SIA clustering which considers the relationships between positive and negative instances is used to extract the label discriminative features.