测绘学报 ›› 2023, Vol. 52 ›› Issue (7): 1090-1104.doi: 10.11947/j.AGCS.2023.20220565

• 高光谱遥感技术专刊 • 上一篇    下一篇

一种联合空谱特征的高光谱影像分类胶囊网络

杜培军1,2,3, 张伟1,2,3, 张鹏1,2,3, 林聪4, 郭山川1,2,3, 胡泽周4   

  1. 1. 南京大学地理与海洋科学学院, 江苏 南京 210023;
    2. 自然资源部国土卫星遥感应用重点实验室, 江苏 南京 210023;
    3. 江苏省地理信息技术重点实验室, 江苏 南京 210023;
    4. 南京市测绘勘察研究院股份有限公司, 江苏 南京 210019
  • 收稿日期:2022-09-30 修回日期:2023-05-11 发布日期:2023-07-31
  • 作者简介:杜培军(1975-),男,教授,博士生导师,研究方向为城市遥感,遥感信息智能处理与地学分析。E-mail:peijun@nju.edu.cn
  • 基金资助:
    国家重点研发计划(2022YFC3800802);国家自然科学基金(42271472)

A capsule network for hyperspectral image classification employing spatial-spectral feature

DU Peijun1,2,3, ZHANG Wei1,2,3, ZHANG Peng1,2,3, LIN Cong4, GUO Shanchuan1,2,3, HU Zezhou4   

  1. 1. School of Geography and Ocean Science, Nanjing University, Nanjing 210023, China;
    2. Key Laboratory for Land Satellite Remote Sensing Applications of Ministry of Natural Resources, Nanjing 210023, China;
    3. Jiangsu Provincial Key Laboratory of Geographic Information Science and Technology, Nanjing 210023, China;
    4. Nanjing Institute of Surveying, Mapping and Geotechnical Investigation, Co., Ltd., Nanjing 210019, China
  • Received:2022-09-30 Revised:2023-05-11 Published:2023-07-31
  • Supported by:
    The National Key Research and Development Program of China (No. 2022YFC3800802); The National Natural Science Foundation of China (No. 42271472)

摘要: 高效稳定的深度学习分类器有助于提升高光谱遥感影像的分类精度。针对卷积神经网络标量式神经元特征表达能力有限、无法有效建模特征之间空间层次结构关系的不足,设计了一种考虑数据图谱合一特性的端到端高光谱胶囊网络(H-CapsNet)。H-CapsNet主体由编码器(卷积层、PrimaryCaps层及DigitCats层)和解码器(全连接层)组成,通过在网络输入端嵌入通道和空间注意力模块,以此增强模型对空谱特征的抓取和识别,进而提升网络对特征的聚焦和表达能力。以资源一号02D卫星获取的张家港高光谱影像及公共数据集University of Pavia和University of Houston影像为例进行试验,将H-CapsNet网络与传统机器学习算法和多个深度学习网络进行对比。试验结果表明,在3景不同分辨率的高光谱影像上,H-CapsNet分类网络均取得了最优的分类效果,总体精度相较于其他方法分别提升了2.36%~7.67%、0.16%~11.8%和1.75%~15.58%。H-CapsNet网络对小像素邻域具有较好的适应性,当图像块尺寸有限时,仍可以取得相对理想的分类结果。

关键词: 胶囊网络, 深度学习, 高光谱遥感, 资源一号02D, 土地覆盖分类

Abstract: An efficient and stable deep learning classifier can improve the classification accuracy of hyperspectral remote sensing images. In order to deal with the insufficiency of the scalar neuron's limited feature expression ability and the inability to effectively model the spatial hierarchical relationship among features in convolutional neural networks, an end-to-end hyperspectral capsule network (H-CapsNet) was designed considering the characteristics of hyperspectral image. The main body of H-CapsNet is composed of encoder (Conv, PrimaryCaps and DigitCats) and decoder (fully connection layer). It mainly embeds channel and spatial attention modules at the network input to enhance the model's capture and recognition of spatial and spectral features, thereby improving the network's ability to focus and express features. Taking the hyperspectral images of Zhang-jia-gang city and two public datasets:University of Pavia and University of Houston, as examples, the performance of the proposed H-CapsNet was compared with traditional machine learning algorithms and several deep neural networks. The experimental results show that the H-CapsNet has achieved the best classification accuracy on three hyperspectral images with different resolutions, and the overall accuracy is improved by 2.36%~7.67%, 0.16%~11.8% and 1.75%~15.58% compared with other methods. In particular, the H-CapsNet has good adaptability to small pixel neighborhoods. When the image patch size is limited, it can still achieve relatively ideal classification results.

Key words: capsule network, deep learning, hyperspectral remote sensing, ZY-1 02D, land cover classification

中图分类号: