注意力引导的三维卷积网络用于遥感场景变化检测

Attention Guided 3D ConvNet for Aerial Scene Change Detection

场景级变化检测策略可以容忍高分遥感影像的大量噪声,进而从语义层级更准确地描述遥感图像在前后时相的变化,为高分辨率影像变化检测提供了可能。本文提出了一种注意力引导的三维卷积神经网络用于高分遥感影像场景变化检测的方法。首先构建一个在AlexNet基础上进行简化的三维卷积网络,然后加入一个语义注意力模块来进一步提取地表覆盖变化显著的候选判别区域;最后输入分类层得到分类结果,整个框架以端对端、可训练的方式进行组织,直接由双时相场景切片通过卷积网络得到变化检测结果。为评估场景级变化检测方法性能,本文制作了一个语义级高分遥感影像场景变化检测数据集,在该数据集上的实验结果显示本文方法变化检测的准确率高于相关方法,验证了方法的有效性,初步展示了基于深度学习的场景级遥感变化检测的发展前景。

With high tolerance to the great amount of noise and precise depiction of im-age changes in high resolution remote sensing images(HRRSI),scene-level change detection strategy makes it possible to detect changes in HRRSI.In this paper,we propose an at-tention guided 3D ConvNet for HRRSI change detection.Firstly,we develop a simplified 3D AlexNet to extract convolutional features.Then we add a semantic attention mod-ule(SAM)to further extract the discriminative regions which strongly relate to land-cover changes.Finally,the refined features are fed into a classification layer to organize the whole framework in an end-to-end trainable manner.Scenes in different phases are put into the convolutional neural network(CNN)with the result of change detection.In order to eval-uate the performance of scene level change detection methods,we create a public semantic level high resolution remote sensing images change detection benchmark.Experimental results on this dataset are obviously better than other related methods,demonstrate the effectiveness of our method,and show the prospect of scene level remote sensing change detection based on deep learning.