基于增强图注意力网络的高光谱影像分类方法

Hyperspectral image classification method based on enhanced graph attention network

高光谱遥感影像中隐含了不同地物的光谱特征,高光谱地物分类成为了遥感领域的一个研究热点。高光谱数据存在维度灾难以及训练样本标签过少等问题,进而影响了其分类精度。针对此问题,文章提出一种空-谱特征融合的增强图注意力网络高光谱影像分类方法,即从高光谱数据中获得初始的空-谱特征作为图的节点属性,并以节点的相邻关系构建图结构;将空-谱特征初步融合的高光谱图数据作为输入,并通过增强图注意力来提取节点的空-谱特征;以深度融合的空-谱特征来实现精准的高光谱地物分类。经在龙口和汉川数据集上的实验测试结果表明:这一方法能够有效提取深度融合的空-谱特征,总体分类精度分别达到99.62%和95.45%,实现了高光谱地物的精准分类。

Hyperspectral remote sensing images contain the spectral characteristics of different ground objects,making the hyperspectral ground object classification a research hotspot in the field of remote sensing.However,hyperspectral data also have some problems such as dimension disaster and too few training sample labels,which affect the classification accuracy.To solve these problems,this paper proposes a hyperspectral image classification method by an enhanced graph attention network based on spatial-spectral features fusion.The initial spatial-spectral features are obtained from hyperspectral data as the attributes of nodes,and the graph structure is constructed based on the adjacent relationship of nodes.The hyperspectral graph data fused by spatial-spectral features are used as input,and the spatial-spectral features of nodes are extracted by enhancing the attention of the graph.The accurate hyperspectral ground object classification is made by deeply fused space-spectrum features.The experimental results on data sets from Longkou and Hanchuan show that this method can effectively extract the deeply fused spatial-spectral features,make the overall classification accuracy is 99.62%and 95.45%respectively,and can get the accurate classification of hyperspectral ground object.