Research On Hyperspectral Image Classification Method Based On Spectral-spatial Combination

Hyperspectral remote sensing can capture objects in hundreds of continuous narrow spectral bands,the hyperspectral image contains both spatial information and spectral information of the target area,and has unique advantages in the field of feature recognition and classification.It has played an increasingly important role in the field of erath observation,and has become one of the fastest growing directions in the field of remote sensing,at present,the hyperspectral remote sensing has been widely used in geological exploration,mineral identification,precision agriculture,military monitoring and other related fields.Hyperspectral image classification is a research hotspot of current hyperspectal processing technology,and the purpose is to extract effective classification features from the large amount of information contained in hyperspectral image,and construct a classifier to give certain feature categories to each pixel in the image.In order to achieve high-precision classification,the current classification methods still have some problems in how to mine the deep features of hyperspectral image and how to deal with complex background information.Therefore,based on the summary of the research status of hyperspectral image classification,this paper proposes two spectral-spatial classification methods,the main contributions of this paper can be conclude as follows:(1)Considering the problems of the hyperspectral image background pixels are complex and cannot participate in training,this paper introduces the target detection method into the classification,and proposed a classification method based on target constraints and spectral-spatial iteration.Based on the detection theory,this method can extracts multiple types of target features and constructs constraines to suppress background information.At same time,in order to eliminate the over-classification problem caused by the spectral features,the method uses the feedback fusion of spectral-spatial to strengthen the spatial information,so as to improve the classification accuracy gradually.(2)The semantic features of the image can express the meaning of the image without losing the essential features,however,most of the semantic feature extraction methods are only used the underlying information(color,texture,etc.)of the image,and lack of application to the hyperspectral image.For this issue,this paper proposes a hyperspectral classification method based on hash semantic embedding for convolutional neural networks.On the one hand,extracting semantic features through hash mapping of inter-class and intra-class constraints,and enhancing classification features;On the other hand,constructing a deep convolutional network to extract depth classification features,and further optimization of spectral-spatial features through the deconvolution layer.Experimental results show that proposed methods can effectively improves the accuracy of classification and has higher stability.