基于主动学习的油气管道沿线地物变化检测

油气管道在储运过程中,沿线区域地物变化对其安全具有较大影响,特别是道路的修建、沟壑的挖掘、滑坡等。由于油气管道分布范围广、周边环境复杂,传统的人工巡检方式存在一定的局限性,因此研究了基于卫星遥感的油气管道沿线地物变化检测。在综合考虑空间信息和算法自动化程度的基础上提出一种改进的基于多特征融合和主动学习的油气管道沿线地物变化检测算法。首先利用基于自适应阈值算法选择初始训练样本,然后利用梯度提升树、k近邻和极限随机树集成结构进行未标记样本的类别判定,并基于边缘采样的主动学习算法进行未标注样本增选。在样本增选过程中为了减少噪声对训练样本的影响并且减少冗余信息,通过两方面对增选样本进行优化,首先通过分割对象约束分类器集成变化检测结果,提高增选样本的准确性,然后利用边缘采样方法选择信息量较大的未标记样本进行标注。通过两景融合后的资源三号(ZY-3)影像进行实验,结果表明该算法可以有效检测地物变化情况,并且在提高变化检测结果精度的同时,可以有效减少训练样本的标注成本。

Ground Feature Change Detection along Pipeline Based on Active Learning

During the process of storage and transportation of oil and gas pipelines, the changes along the pipeline has a great impact on its safety, especially the construction of buildings and roads, the excavation of gully, landslide etc.. Due to the wide distribution and the complex surrounding environment of oil and gas pipelines, there are some limitations in the traditional manual inspection methods. Therefore, the change of ground objects along the oil and gas pipelines based on satellite remote sensing is studied in this paper. Considering of spatial information and spectral information, a semi-supervised changed detection based on multiple features fusion and active learning has been proposed to improve the automation of change detection methods in this paper. Firstly, the self-adaptive threshold is implemented to select the initial training samples. Gradient Boosting trees, k-nearest neighbor and extra-trees classifiers are then used to construct the ensemble system to identify the unlabeled samples and margin sampling active learning algorithm is implemented to label the unlabeled samples. In order to reduce the influence of noise and redundant information in sample selection, object information is utilized to constraint the pixel-based change detection results and margin sampling is implemented to select the samples with large amount of information. The experiment is carried out by two sets of ZY-3 remote sensing images, and the results show that this algorithm can improve the accuracy of change detection with reducing the labeling cost of training samples.