一种基于CNN-GCN的高分辨率遥感影像土地覆盖分类  

Land cover classification of high resolution remote sensing images based on CNN-GCN

作　　者：田智慧[1,2] 常蓬 赫晓慧[1,2] 程淅杰 TIAN Zhihui;CHANG Peng;HE Xiaohui;CHENG Xijie(School of Earth Science and Technology,Zhengzhou University,Zhengzhou 450000,China;Joint Laboratory of Eco-Meteorology,Zhengzhou University,Chinese Academy of Meteorological Sciences,Zhengzhou 450000,China;School of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450000,China)

机构地区：[1]郑州大学地球科学与技术学院,郑州450000 [2]郑州大学中国气象科学研究院郑州大学生态气象联合实验室,郑州450000 [3]郑州大学计算机与人工智能学院,郑州450000

出　　处：《测绘科学》2023年第6期59-72,共14页Science of Surveying and Mapping

基　　金：河南省重大科技专项(201400210900);第二次青藏高原科学考察计划项目(2019QZKK0106)。

摘　　要：针对单一深度学习模型无法有效地提取局部-全局空间光谱特征的问题,该文提出了一种基于像素-超像素的异构卷积神经网络(HCNN)特征级联的分类方法,有效地提高了土地覆盖分类性能。具体而言,基于注意力机制的像素级特征提取模块和基于空间域的超像素级特征提取模块分别提取局部和全局空间-光谱特征,并将两种网络的输出特征级联融合作为最终图像表达,进行土地覆盖类别预测。并设计消融实验验证了特征提取分支设计本身的合理性和融合的必要性。本研究基于公开高分影像数据集和河南土地利用/土地覆盖(HN-LULC)数据集进行了验证,结果证明该混合卷积神经网络级联融合分类方法能够显著地提升高分辨率遥感影像中土地覆盖分类的效果,相比现有的分类模型,分类结果最佳。Aiming at the problem of ineffective extraction of local-global spatial-spectral features by a single deep learning model,this paper proposed a classification method based on a pixel-superpixel heterogeneous convolutional neural network(HCNN)feature cascade,which effectively improved land cover classification performance.Specifically,a pixel-level feature extraction module based on attention mechanism and a superpixel-level feature extraction module based on spatial domain were used to extract local and global spatial-spectral features,respectively.The output features of the two networks were concatenated and fused as the final image representation for land cover category prediction.Additionally,ablation experiments were designed to verify the rationality of the feature extraction branch design and the necessity of fusion.The proposed method was validated on the publicly available Gaofen image dataset and a self-made HN-LULC dataset,and the experimental results demonstrated that the mixed convolutional neural network cascade fusion classification method could significantly improve the effectiveness of land cover classification in high-resolution remote sensing images,outperforming existing classification models with the best classification results.

关 键 词：土地覆盖分类 深度学习 卷积神经网络 超像素分割 图卷积神经网络 高分辨率遥感影像 

分 类 号：P237[天文地球—摄影测量与遥感]