地球信息科学学报 ›› 2018, Vol. 20 ›› Issue (12): 1756-1767.doi: 10.12082/dqxxkx.2018.180353

• 遥感科学与应用技术 • 上一篇    下一篇

基于Adaboost的高分遥感影像自动变化检测方法

陈伟锋1,2,3(), 毛政元1,2,3,*(), 徐伟铭1,2,3, 许锐1,2,3,4   

  1. 1. 福州大学福建省空间信息工程研究中心,福州 350002
    2. 福州大学空间数据挖掘与信息共享教育部重点实验室,福州 350002
    3. 福州大学地理空间信息技术国家地方联合工程研究中心,福州 350002
    4. 福建工程学院信息科学与工程学院,福州 350118
  • 收稿日期:2018-07-30 出版日期:2018-12-25 发布日期:2018-12-20
  • 通讯作者: 毛政元 E-mail:woundclock@foxmail.com;zymao@fzu.edu.cn
  • 作者简介:

    作者简介:陈伟锋(1992-),男,硕士生,主要从事遥感影像信息提取、数字图像处理研究。E-mail: woundclock@foxmail.com

  • 基金资助:
    国家自然科学基金项目(41701491);福建省自然科学基金面上项目(2018J01619)

Automatic Change Detection Approach for High-Resolution Remotely Sensed Images Based on Adaboost Algorithm

CHEN Weifeng1,2,3(), MAO Zhengyuan1,2,3,*(), XU Weiming1,2,3, XU Rui1,2,3,4   

  1. 1. Provincial Spatial Information Engineering Research Center, Fuzhou University, Fuzhou 350002, China
    2. Key Laboratory of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350002, China
    3. National Engineering Research Centre of Geospatial Space Information Technology, Fuzhou University, Fuzhou 350002, China
    4. School of Information Science and Engineering, Fujian University of Technology, Fuzhou 350188, China
  • Received:2018-07-30 Online:2018-12-25 Published:2018-12-20
  • Contact: MAO Zhengyuan E-mail:woundclock@foxmail.com;zymao@fzu.edu.cn
  • Supported by:
    National Natural Science Foundation of China, No.41701491;Project of Science and Technology of Fujian Province, No.2018J01619.

摘要:

基于监督分类的高分辨率遥感影像变化检测需要大量人工标注,且单个监督分类器难以适应高分影像中复杂多样的地表变化信息提取,检测结果中“椒盐噪声”严重、变化图斑破碎。因此,本文提出一种基于Adaboost集成算法、自动标注训练样本的变化检测方法。首先利用非监督分类方法完成变化初检,接着在初检结果中进行“非等距”区间采样自动获取均匀分布的训练样本;然后以Adaboost算法为集成框架,选择决策树桩、Logistic回归和kNN作为弱分类器,构建一种混合分类器集成系统,充分挖掘和利用高分影像中的空间信息以提升分类精度和分类器泛化能力,最后利用SLIC分割算法和空间邻域信息对像元级检测结果进行空间约束滤波,进一步提升变化检测精度。为验证本文方法的有效性,选取SPOT-5和WorldView-2影像为实验数据,结果表明本文方法能有效降低训练样本人工标注成本、提高变化检测精度。

关键词: 变化检测, 高分影像, 分类器集成, Adaboost, 自动标注, 空间约束

Abstract:

Human annotation is a massive labor cost for the training sample selection process when applying any kind of supervised learning algorithm for change detection based on high-resolution remotely sensed satellite images. It is limited and unreasonable to use just one single sort of classifier generated from a supervised algorithm to extract change information of variety from the time-series images both in completeness and accuracy, let alone the inevitable salt-and-pepper noise and tiny patches falsely detected which turn out to be ubiquitous in and out of geographical entities. To tackle with problems mentioned above, a change detection approach based on a new automatic training sample annotation strategy and an improved Adaboost ensemble learning algorithm was proposed. At first, the unsupervised change detection algorithm CVA was applied to generate a low-level change detection result as referencing labels for further annotation, then the low-level result was divided into several parts with different intervals to ensure the automatic acquisition of the evenly distributed training samples with confidence. Furthermore, decision stump, logistic regression and kNN were employed as the weak classifiers to construct a hybrid multi-classifiers ensemble system with the help of the improved Adaboost algorithm, which would effectively promote the classification accuracy and generalization capacity of weak classifiers by sufficiently mining and making use of the spatial information with potential values. Finally, the SLIC segmentation algorithm was implemented in the difference image, and the segmentation border information was combined with spatial contextual information to build up a dual-filter for spatial constraint aiming at decreasing the omission rate and the false alarm rate of the detection results. To verify the validity of the proposed method, we conducted experiments using two datasets of multispectral images collected by SPOT-5 and WorldView-2. Experimental results indicated that the proposed method would significantly lower the labor costs of training sample annotation and demonstrated superiority compared with four other methods in accuracy.

Key words: change detection, high resolution remote sensing image, ensemble learning, Adaboost, automatic annotation, spatial constraint