面向对象高分影像归纳式图神经网络分类法

Object-oriented high-resolution image classification using inductive graph neural networks

传统面向对象分类多采用遥感影像的光谱特征,忽略了影像对象之间的空间特征。本文提出了一种采用改进归纳式图神经网络的高分遥感影像面向对象分类方法,实现了光谱-空间复合节点相似度的融合系数自适应调节,以及邻域节点采样最佳数量的自动确定。首先,改进KNN图构建方法,采用标准差信息量评价法确定用于构建光谱特征和空间特征的复合节点相似度的融合系数;然后,利用反馈曲线法确定最佳的采样邻域节点数量,使用GraphSAGE节点嵌入完成特征表达;最后,依托Softmax函数预测节点类别。以GID和BDCI2017数据集为试验数据,本文的构图方法相较于改进前的构图方法在分类精度上有所提升。本文分类方法的平均Kappa系数和总体精度分别优于CART分类树算法、GCN算法、GAT算法、LANet算法、CCTNet算法和SLCNet算法0.31、0.14、0.13、0.12、0.08、0.02和42.31%、7.4%、6.73%、8.69%、6.03%、1.52%,并且在植被和建设用地提取上具有较好的稳健性。本文方法为高分遥感影像土地覆盖分类提供了有效的工具。

Traditional object-oriented classification methods mostly use spectral features of image objects and ignore the spatial features among image objects.In this paper,an object-oriented classification method for high-resolution remote sensing images using improved inductive graph neural network is proposed.The method is able to adaptively adjust the fusion coefficient of spectral-spatial composite node similarity and automatically determine the optimal sampling number of neighboring nodes.First,we improved the K-nearest neighbor(KNN)graph construction method.The standard deviation informativeness evaluation method was used to determine the fusion coefficients for constructing the composite node similarity of spectral and spatial features.Then,the optimal sampling number of neighboring nodes was determined using the feedback curve method,and feature representation was accomplished using GraphSAGE node embedding.Finally,the classifications of the nodes were predicted by Softmax function.We used GID-15 and BDCI2017 datasets as experimental data.The proposed graph construction method has improved the classification accuracy.The average Kappa coefficient of the proposed method was better than CART,GCN,GAT,LANet,CCTNet,and SLCNet by 0.31,0.14,0.13,0.12,0.08,and 0.02.The average overall accuracy,on the other hand,was better than 42.31%,7.4%,6.73%,8.69%,6.03%,and 1.52%.Meanwhile,our method had good robustness in vegetation and built-up land extraction.The method proposed in this paper provides an effective tool for land cover classification of high-resolution remote sensing images.