基于Adaboost的高分遥感影像自动变化检测方法
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
收稿日期: 2018-07-30
网络出版日期: 2018-12-20
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
Received date: 2018-07-30
Online published: 2018-12-20
Supported by
National Natural Science Foundation of China, No.41701491;Project of Science and Technology of Fujian Province, No.2018J01619.
Copyright
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。
陈伟锋 , 毛政元 , 徐伟铭 , 许锐 . 基于Adaboost的高分遥感影像自动变化检测方法[J]. 地球信息科学学报, 2018 , 20(12) : 1756 -1767 . DOI: 10.12082/dqxxkx.2018.180353
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.
Fig. 1 Flowchart of the proposed change detection method图1 本文变化检测方法流程 |
Fig. 2 Original images for change detection图2 变化检测的原始影像 |
Tab. 1 Error matrix of change detection表1 变化误差矩阵 |
实际变化 | 实际未变化 | 总和 | |
---|---|---|---|
检测变化 | TP | FP | P |
检测未变化 | FN | TN | N |
总和 | P′ | N′ | T |
Fig. 3 Difference images图3 差值影像 |
Fig. 4 The number of error pixels obtained with different α1图4 不同α1下样本选择分类精度 |
Fig. 5 Visualization images of different sample selection strategies with their optimal α1图5 最佳α1下不同样本自动选择方法可视化结果 |
Fig. 6 The number of error pixels obtained with different numbers of weak classifiers图6 不同弱分类器集成数量的变化检测精度 |
Fig. 7 The results of SLIC segmentation图7 SLIC分割结果 |
Tab. 2 The change detection accuracy assessment result of dataset 1表2 数据集1变化检测结果精度评价 |
Precision | Kappa | TPR | FPR | |
---|---|---|---|---|
CVA | 0.5038 | 0.5720 | 0.2314 | 0.4962 |
LR | 0.6685 | 0.7036 | 0.2058 | 0.3315 |
ID3 | 0.6313 | 0.6589 | 0.2510 | 0.3687 |
kNN | 0.7207 | 0.7188 | 0.2412 | 0.2793 |
Ada-LR | 0.7090 | 0.7530 | 0.1534 | 0.2910 |
Ada-DS | 0.7376 | 0.7279 | 0.2431 | 0.2624 |
HCS | 0.8371 | 0.8472 | 0.1197 | 0.1629 |
HCS-SC | 0.8756 | 0.8691 | 0.1190 | 0.1244 |
Tab. 3 The change detection accuracy assessment result of dataset 2表3 数据集2变化检测结果精度评价 |
Precision | Kappa | TPR | FPR | |
---|---|---|---|---|
CVA | 0.5725 | 0.5428 | 0.3481 | 0.4275 |
LR | 0.6224 | 0.6024 | 0.2968 | 0.3776 |
ID3 | 0.7046 | 0.5809 | 0.4260 | 0.2954 |
kNN | 0.6923 | 0.5841 | 0.4108 | 0.3077 |
Ada-LR | 0.7611 | 0.7050 | 0.2711 | 0.2389 |
Ada-DS | 0.8044 | 0.6794 | 0.3505 | 0.1956 |
HCS | 0.9013 | 0.8320 | 0.1889 | 0.0987 |
HCS-SC | 0.9197 | 0.8383 | 0.1942 | 0.0803 |
Fig. 8 The change detection results and the reference image of dataset 1图8 数据集1检测结果及标准影像 |
Fig. 9 The change detection results and the reference image of dataset 2图9 数据集2检测结果及标准参考影像 |
The authors have declared that no competing interests exist.
[1] |
|
[2] |
|
[3] |
[
|
[4] |
|
[5] |
[
|
[6] |
|
[7] |
|
[8] |
[
|
[9] |
|
[10] |
[
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
[
|
[19] |
|
[20] |
|
[21] |
|
/
〈 | 〉 |