测绘通报 ›› 2019, Vol. 0 ›› Issue (2): 103-107.doi: 10.13474/j.cnki.11-2246.2019.0053

• 技术交流 • 上一篇    下一篇

结合主动学习和词袋模型的高分二号遥感影像自动化分类

张金盈1, 姚光虎2, 林琳3, 郭怀轩2   

  1. 1. 山东省国土测绘院遥感技术部, 山东 济南 250013;
    2. 神舟航天软件(济南)有限公司卫星应用中心, 山东 济南 250013;
    3. 山东省水利科学研究院水资源与水环境省重点实验室, 山东 济南 250013
  • 收稿日期:2018-04-03 出版日期:2019-02-25 发布日期:2019-03-05
  • 作者简介:张金盈(1980-),男,硕士,高级工程师,主要研究方向为遥感影像处理、解译的理论和应用。E-mail:591248670@qq.com
  • 基金资助:

    山东省自然科学基金(ZR2015EM007)

Automatic classification of GF-2 remote sensing imagery based on active learning and bag of visual words model

ZHANG Jinying1, YAO Guanghu2, LIN Lin3, GUO Huaixuan2   

  1. 1. Shandong Geological Surveying and Mapping Institute, Jinan 250013, China;
    2. Center of Satellite Application, Shenzhou Aerospace Software(Jinan) Co., Ltd., Jinan 250013, China;
    3. Water Research Institute of Shandong Province, Shandong Provincial Key Laboratory of Water Resources and Environment, Jinan 250013, China
  • Received:2018-04-03 Online:2019-02-25 Published:2019-03-05

摘要:

高分卫星遥感影像空间分辨率的提高,使得地物的光谱和纹理变得更加丰富和复杂,这给遥感影像的自动化分类带来严重挑战。因此,本文提出了一种结合主动学习和词袋模型的高分二号遥感影像分类方法。首先,对研究区域进行多尺度分割,建立影像分割对象集;然后,采用词袋模型构建影像对象的语义特征向量;最后,充分考虑位于分类边界的不确定性样本分布,迭代选择最优样本用于训练支持向量机,用于分类遥感影像。为了验证本文方法的有效性和稳健性,以山东省某市的高分二号遥感影像为试验数据进行了试验分析。结果表明,本文提出的方法可以有效地将研究区域分为水体、地面、植被和建筑物四类,正确率达到90.6%以上。

关键词: 遥感影像, 机器学习, 分类, 主动学习

Abstract:

The improvement of high-resolution satellite images makes the spectrum and texture more rich and complex, which poses challenges for the automatic classification. Therefore, this paper combines active learning and bag of word model for image classifications. First, a multi-scale segmentation is implemented to generate image objects. Second, bag of word model is used to establish the semantic feature of image object. Finally, the uncertainty sample distribution is well considered, and the optimal samples are selected iteratively for training SVM to classify image. To verify the effectiveness and robustness, the high-resolution image in Shandong province was used as experimental data. The results show that the proposed method can effectively classified the study area into four types:water, ground, vegetation, and building, with the overall accuracy of over 90.6%.

Key words: remote sensing imagery, machine learning, classification, active learning

中图分类号: