用于遥感图像变化检测的深度监督网络

Deep supervised network for change detection of remote sensing image

为了解决大多数全卷积网络出现的特征提取单一、遥感图像变化检测(CD)能力差的问题,借助Unet++网络构建用于遥感图像CD的深度监督网络(DSNet).设计多尺度残差模块替换传统卷积层,通过融合孪生网络双分支上的空间与光谱特性,获取遥感图像在不同层次间的语义信息,有效解决了特征提取单一的问题.在模型解码端设计横向输出层,实现节点从低级向高级特征聚合的深度监督过程.将具备信息差异化的不同特征融合结果传输至基于归一化的注意力模块(NAM)中.在不引入额外参数的前提下增强了变化区域的信息权重.实验结果表明,所提模型在遥感图像CD任务中的召回率和精度分别为90.39%和92.04%,模型的参数量和计算量为6.38 M和60 G.与不同网络模型的对比表明,该方法具有检测精度高、速度快和轻量化等优点.

A deep supervised network(DSNet)for remote sensing image change detection(CD)was built with the Unet++to address the inadequate feature extraction and poor ability of remote sensing image CD in most fullly convolutional networks.A multi-scale residual module was designed to replace the conventional convolution layer.The problem of single feature extraction was resolved by combining the spatial and spectral properties of twin networks to obtain the semantic information of remote sensing images at various levels.The horizontal output layer of model was built to implement the deep supervision process of node feature aggregation from shallow to deep features at the decoding end of the model.The fusion results of different features with information differentiation were transmitted to the normalization-based attention module(NAM).The information weight of the changed region was enhanced without introducing additional parameters.Experimental results showed that the recall rate and accuracy of the proposed model in the remote sensing image CD task were 90.39%and 92.04%,respectively,and the parameter quantity and calculation quantity of the model were 6.38 M and 60 G.Comparise with different network models indicates that the proposed method has the advantages of high detection accuracy,fast speed and lightweight.