基于改进DeepLabv3+孪生网络的遥感影像变化检测方法

Remote Sensing Image Change Detection based on Improved DeepLabv3+Siamese Network

传统遥感影像变化检测方法依赖人工构建特征,算法设计复杂且精度不高;而将2幅不同时相影像叠加后输入神经网络的遥感影像变化检测方法会造成不同时相的特征相互影响,难以保持原始影像的高维特征,且模型鲁棒性较差。因此,本文提出一种基于改进DeepLabv3+孪生网络的遥感影像变化检测方法,以经典DeepLabv3+网络的编解码结构为基础对网络进行改进:(1)在编码阶段利用共享权值的孪生网络提取特征,通过2个输入端分别接收2幅遥感影像,以保留不同时相影像的高维特征;(2)在特征融合中用密集空洞空间金字塔池化模型代替空洞空间金字塔池化模型,通过密集连接的方式结合每个空洞卷积的输出,以提高对不同尺度目标分割的精度;(3)在解码阶段中针对不同层级特征图信息差异较大,难以融合的问题,引入基于注意力机制的特征对齐模型,引导不同层级的特征对齐并强化学习重要特征,以提升模型的鲁棒性。应用开源数据集CDD验证本文方法的有效性,并与UNet-EF、FC-Siam-conc、Siam-DeepLabv3+和N-Siam-DeepLabv3+网络对比试验。试验结果表明,本文方法在精确率、召回率、F1值和总体精度上达到87.3%、90.2%、88.4%、96.4%,均高于UNet-EF、FC-Siamconc、Siam-DeepLabv3+网络和N-Siam-DeepLabv3+网络,检测结果较为完整,对边界的检测也更为平滑,且对尺度变化具有更高的鲁棒性。

Traditional remote sensing image change detection method relies on artificial construction of features,and the algorithm design is complex and has a low accuracy.Moreover,the remote sensing image change detection technique,which superimposes two different phase images and then inputs them into the neural network,will cause the interaction of characteristics of different phases.It is difficult to maintain the high-dimensional features of the original image,and the model is less robust.Therefore,this paper proposes a remote sensing image change detection method that improves the DeepLabv3+Siamese network based on the encoding and decoding structure of the classic DeepLabv3+network:1)In the encoding stage,the features are extracted by the Siamese network sharing weights,and remote sensing images are received through two input terminals respectively,so as to preserve the high-dimensional features of different time-phase images;2)The dense atrous space pyramid pooling model replaces the atrous space pyramid pooling model in feature fusion.In addition,the method that combines the output of each atrous convolution through dense connections improves the segmentation of objects of different scales;3)In the decoding stage,multiple levels of feature map information contain variance that causes integration problems.As a result,a feature alignment model based on the attention mechanism is introduced to guide the feature alignment of different levels,and then strengthen the learning of critical features to enhance model robustness.The open-source dataset CDD is used to verify the efficacy of the method in this paper,compared with UNet-EF,FC-Siam-conc,Siam-DeepLabv3+and N-Siam-DeepLabv3+networks.The test results demonstrate that the presented approach in the study achieves 87.3%、90.2%、88.4%、96.4%in precision rate,recall rate,F1 value,and overall accuracy,respectively,which are higher than those of the UNet-EF,FC-Siam-conc,Siam-DeepLabv3+and N-Siam-DeepLabv3+networks.The detection results turn out to be more comprehensive,and the boundary detection is smoother and more robust to scale changes.