基于三维卷积神经网络的遥感影像变化检测

Remote Sensing Image Change Detection Based on 3D Convolutional Neural Network

随着卫星传感器技术和深度学习技术的蓬勃发展,基于深度学习的变化检测研究已成为遥感变化检测领域中的主流方法。针对现有深度学习方法特征提取和融合不充分的问题,提出了一种基于三维卷积神经网络的遥感影像变化检测方法。在特征编码阶段,首先利用三维卷积的内部融合特性同时提取和融合双时相影像特征。在特征解码阶段,为有效利用影像特征的全尺度信息,采用全尺度跳跃连接机制将不同尺度的特征信息在时间维度进行结合,最终产生具有高精度的变化结果。实验结果表明,该方法在两个基准数据集上的精度均显著优于其他先进的深度学习变化检测方法。

With the rapid development of sensor technology and deep learning technology,change detection based on deep learning technology has become a research hotspot in the field of remote sensing.However,existing deep learning-based change detection methods suffer from limited accuracy due to inadequate feature extraction and fusion.To address that,this paper proposes a change detection method based on 3D convolutional neural network.In the encoding stage,the inner fusion property of 3D convolution is used to extract and fuse bi-temporal image features simultaneously.In the decoding stage,in order to effectively utilize the full-scale information of image features,the full-scale skip connection strategy is introduced to combine the feature information from different scales in the time dimension and then generate the change map.Extensive comparative experiments demonstrate that the proposed method performs better than the other state-of-the-art change detection methods.