地球信息科学学报 ›› 2021, Vol. 23 ›› Issue (5): 928-937.doi: 10.12082/dqxxkx.2021.200339

• 遥感科学与应用技术 • 上一篇    下一篇

基于决策融合的SPOT-6影像土地覆被分类研究

李宏达1,2,3,4(), 高小红1,2,3,4,*(), 汤敏1,2,3,4   

  1. 1.青海师范大学地理科学学院,西宁 810008
    2.高原科学与可持续发展研究院,西宁 810008
    3.青海省自然地理与环境过程重点实验室,西宁 810008
    4.青藏高原地表过程与生态保育教育部重点实验室,西宁 810008
  • 收稿日期:2020-06-30 修回日期:2020-09-19 出版日期:2021-05-25 发布日期:2021-07-25
  • 通讯作者: 高小红
  • 作者简介:李宏达(1995— ),男,湖北荆门人,硕士生,研究方向为遥感应用与地理空间数据分析。E-mail:2395789679@qq.com
  • 基金资助:
    青海省科技厅自然科学基金项目(2016-ZJ-907)

Land Cover Classification for SPOT-6 Image from Decision Fusion Method

LI Hongda1,2,3,4(), GAO Xiaohong1,2,3,4,*(), TANG Min1,2,3,4   

  1. 1. College of Geographical Sciences, Qinghai Normal University, Xining 810008, China
    2. Academy of Plateau Science and Sustainability, Xining 810008, China
    3. Qinghai Province Key Laboratory of Physical Geography and Environmental Process, Xining 810008, China
    4. Ministry of Education Key Laboratory of Tibetan Plateau Land Surface Processes and Ecological Conservation, Xining 810008, China
  • Received:2020-06-30 Revised:2020-09-19 Online:2021-05-25 Published:2021-07-25
  • Contact: GAO Xiaohong
  • Supported by:
    Natural Science Fund Project of QingHai Science and Technology Department(2016-ZJ-907)

摘要:

多分类器决策融合方法在提高遥感影像分类的准确性和可靠性方面已表现出了巨大潜力,但这一过程中对所有像元多次分类会产生巨大的时间代价,为改善这一问题,本文提出了主分类器的概念。在青海湟水流域确定2个试验区,对7种常用的分类器进行评估,排除精度较低的3种分类器后,选择支持向量机(Support Vector Machine, SVM)、多层感知机(Multilayer Perceptron, MLP)、随机森林(Random Forest, RF)和梯度提升树(Gradient Boosting Decision Tree, GBDT)4种不同的分类器,建立决策规则共同对SPOT-6影像分类。为提高分类效率,以精度最高的GBDT作为主分类器对影像分类后,仅对结果中可信度不高的像元使用多分类器共同决策。研究结果表明,2个区域内主分类器独立完成分类的像元分别占38.10%和65.26%,错分率为1.57%和2.18%;多分类器共同决策的区域,相比GBDT的分类结果,总体精度分别高出2.49%和3.66%。整体上看,决策融合使2个区域的总体分类精度分别提高了1.18%和1.09%,能够有效减少分类结果中的“椒盐噪声”,精度更加均衡。相比现有的决策融合方法,主分类器的使用在保证分类精度的同时有利于分类效率的提高及分类结果保持良好的一致性。

关键词: 决策融合, 多分类器, 主分类器, 机器学习, GBDT, 土地覆被分类, SPOT-6, 湟水流域

Abstract:

The decision fusion of multi-classifiers has shown great potential in improving the accuracy and reliability of remote sensing image classification. However, multi-classification of all pixels is usually time-consuming. In order to solve this problem, this paper puts forward the concept of master classifier based on existing studies. Firstly, two experimental areas were selected in the Huangshui river basin of Qinghai province: region A representing urban area with serious human activities and complex spectra, and region B representing rural and mountainous area with relatively simple spectra. To obtain a high classification accuracy, seven different commonly used classifiers were selected for decision fusion. Using the testing samples, the classifiers with low accuracy were excluded based on the average accuracy of two regions. Excluded classifies were Naive Bayesian (NB), K-NearestNeighbor (KNN), and Decision Tree (DT). The other four classifiers including Gradient Boosting Decision Tree (GBDT), Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) were kept to establish the decision rules for SPOT-6 images classification. Particularly, the GBDT with the highest accuracy among all classifiers was regarded as the master classifier. After the classification by GBDT, the pixels with low confidence were classified again using the other three classifiers, and the decision fusion was made together with the result of GBDT to select the best classification result. The results show that, 38.10% and 65.26% of the pixels in the two regions were classified by the master classifier alone, respectively, and the misclassification rate was 1.57% and 2.18%, respectively. For the regions of decision making using multiple classifiers, the overall classification accuracy was respectively 2.49% and 3.66% higher than that using GBDT. On the whole, the decision fusion improved the overall classification accuracy of the two regions by 1.18% and 1.09% respectively, effectively reduced the "salt and pepper " noise in the results, and achieved a more homogeneous classification accuracy. Compared with the existing decision fusion researches, the use of the master classifier not only ensures the accuracy of classification, but also improves the classification efficiency and maintains a good consistency of the classification results.

Key words: decision fusion, multiple classifiers, the master classifier, machine learning, GBDT, land cover classification, SPOT-6, Huangshui river basin