基于数据均衡的遥感变化检测训练集生成方法

A Training Set Generation Method for Remote Sensing Change Detection Based on Data Balance Strategy

深度学习已成为遥感变化检测的主流方法。然而,深度学习需要大量的标记样本来训练网络模型,而生成训练样本是一件费时费力的工作。为降低训练网络模型所需的样本量,提出一种基于数据均衡的遥感变化检测训练集生成方法。该方法主要包括两部分:第一部分提出一种基于数据均衡策略的训练集生成方案,获得初始的训练数据集;第二部分给出一种顾及尺度多样性的数据增强方法,来扩展初始训练集。采用2组常用的变化检测数据集——武汉建筑物变化检测和谷歌地球变化检测数据集,来验证所提方法的有效性。实验结果表明,所提方法能够在保证精度的前提下,显著降低所需训练集的数量。

Deep learning has become the mainstream method of remote sensing change detection.However,training deep learning model requires a large number of labeled samples.It’s time-consuming and laborious to generate training samples.In order to reduce the number of the samples required by training network models,a training set generation method for remote sensing change detection is proposed based on data balance strategy.The method includes two main parts:in the first part,a training set generation method considering data balance is proposed to produce an initial training data set;in the second part,a data augmentation method that considers the scale diversity is presented to expand the obtained initial training set.Two commonly-used change detection datasets,Wuhan building change detection and Google earth change detection dataset,are employed to verify the effectiveness of the proposed method.The experimental results show that the proposed method can significantly reduce the number of training sets required under the premise of ensuring the accuracy.