基于改进STANet的遥感图像变化检测算法

Remote Sensing Image Change Detection Algorithm Based on Improved STANet

遥感图像变化检测是为了识别出双时相图像之间的显著变化。给定2个在不同时间拍摄的配准图像,光照变化和错配误差会掩盖真实物体的变化,探索不同时空像素之间的关系可以提高遥感图像变化检测方法的性能。在Spatial Temporal Attention Neural Network(STANet)中,提出了一种基于孪生的时空注意力神经网络。在其基础上进行改进:①对距离度量模块由于线性插值导致的变化特征间隙模糊问题,设计了对距离特征的上采样模块,使得变化区域间隙更加明显,虚警率更低;②针对STANet的Pyramid Spatial Temporal Attention Module(PAM)模块计算开销大的问题,引用了新的Coordinate Attention(CA)模块,在降低运算开销的基础上,更好地识别了不同空间、通道的特征;③针对STANet对Residual Network(ResNet)提取出的特征图利用不充分的问题,加入了深监督模块,利用中间层的特征计算一个权重衰减的loss,起到正则化的作用。实验表明,改进之后的网络将基线模型的F1得分从81.6提高到86.1。在公共遥感图像数据集上的实验结果表明,改进的方法优于其他几种先进的方法。

Remote sensing image change detection is to identify the significant changes between dual temporal images.Given two registration images taken at two different times,changes in lighting and mismatch errors can mask the changes in real objects.Exploring the relationship between different spatiotemporal pixels can improve the performance of remote sensing image change detection methods.In Spatial Temporal Attention Neural Network(STANet),a twin based spatiotemporal attention neural network is proposed,based on which some improvements are made.①By designing an up-sampling module for the distance measurement module,the problem of fuzzy feature gaps caused by linear interpolation is solved,making the gap in the changing area more obvious and the false alarm rate lower.②To address the high computational cost of STANet's PAM(Pyramid Spatial Temporary Attention Module)module,a new Coordinate Attention(CA)attention module is introduced to better identify the features of different spaces and channels while reducing the computational cost.③To solve the problem of insufficient utilization of the feature maps extracted from Residual Network(ResNet)by STANet,a deep supervision module is added to calculate a weight attenuation loss using the features of the middle layer,which plays a role of regularization.The experiment shows that the improved network improves the F1 score of the baseline model from 81.6 to 86.1.The experimental results on public remote sensing image datasets show that the improved method outperforms several other advanced methods.