电子学报 ›› 2021, Vol. 49 ›› Issue (8): 1599-1614.DOI: 10.12263/DZXB.20200961

• 学术论文 • 上一篇    下一篇

神经网络训练策略对高分辨率遥感图像场景分类性能影响的评估

郑海颖1, 王峰1, 姜维1, 王志强2, 姚西文2   

  1. 1.华北水利水电大学信息工程学院,河南 郑州 450046
    2.西北工业大学自动化学院,陕西 西安 710129
  • 收稿日期:2020-08-31 修回日期:2020-12-14 出版日期:2021-08-25 发布日期:2021-08-25
  • 作者简介:郑海颖 男,1994年8月出生,山东潍坊人.华北水利水电大学信息工程学院硕士研究生,主要研究方向为遥感图像场景分类、目标检测. E-mail:zhy20130901@163.com
    王 峰 男,1970年11月生,河南汤阴人.华北水利水电大学信息工程学院副教授,硕士生导师,主要研究方向为软件工程技术、数据库技术、图形图像处理、机器学习等. E-mail:wangfeng@ncwu.edu.cn
    姜 维(通讯作者) 男,1981年12月生,河南郑州人.现为华北水利水电大学副教授,主要研究方向为场景文字检测与识别、遥感目标的检测与识别. E-mail:jiangwei@ncwu.edu.cn
  • 基金资助:
    国家自然科学基金(61601184);河南省科技攻关计划(192102210265);河南省教育厅科学技术研究重点项目(13A520713);河南省重点科技攻关计划(152102210112)

Evaluation of the Effect of Neural Network Training Tricks on the Performance of High-Resolution Remote Sensing Image Scene Classification

Hai-ying ZHENG1, Feng WANG1, Wei JIANG1, Zhi-qiang WANG2, Xi-wen YAO2   

  1. 1.School of Information Engineering,North China University of Water Resources and Electric Power,Zhengzhou,Henan 450046,China
    2.School of Automation,Northwestern Polytechnical University,Xi’an,Shaanxi 710129,China
  • Received:2020-08-31 Revised:2020-12-14 Online:2021-08-25 Published:2021-08-25

摘要:

机器学习方法在高分辨率遥感图像场景分类任务中已经得到大规模应用,但当前研究主要围绕数据特征和神经网络结构展开,极少提及神经网络训练策略对遥感图像分类性能的影响.因此,本文选取7种自然图像分类中常用的神经网络训练策略进行实验,根据其在3个规模较大的遥感图像数据集和4个广泛使用的神经网络模型上的实验表现,筛选出适用于遥感图像场景分类的神经网络训练策略.通过消融研究详细评估多个神经网络训练策略对遥感图像场景分类性能的影响,通过分析总体分类精度、混淆矩阵、Kappa系数得到有效的神经网络训练策略,并证明神经网络训练策略对遥感图像场景分类性能的有效性;根据叠加实验的结果分析,7种训练策略的组合可以在不同网络模型和数据集上表现出良好的适用性.

关键词: 机器学习, 高分辨率, 遥感场景分类, 训练策略, 神经网络

Abstract:

Machine learning have been widely used in high-resolution remote sensing image scene classification task. However, the current research mainly focuses on data features and neural network structure, and the effect of neural network training tricks on remote sensing image classification performance is rarely mentioned. Therefore, this paper selects 7 neural network training tricks commonly used in natural image classification for experiments. According to their experimental performance in 3 large remote sensing image data sets and 4 widely used neural network models, neural network training tricks suitable for remote sensing image scene classification are selected. The effect of multiple neural network training tricks on the scene classification performance of remote sensing images was evaluated in detail through ablation experiment. An effective neural network training strategy was obtained by analyzing the overall accuracy, confusion matrix and Kappa coefficient, and the effectiveness of the neural network training strategy on the scene classification performance of remote sensing images was proved. According to the results of the stacking experiment, the combination of 7 training tricks can show good applicability in different network models and data sets.

Key words: machine learning, high-resolution, remote sensing scene classification, training tricks, neural network

中图分类号: