【摘要】 为提高多时相遥感影像的海岛变化检测精度，文中采用编解码结构模型将变化检测中变和不变的二分类问题视为语义分割任务，提出一种残差卷积与多尺度特征融合的海岛多时相遥感影像变化检测方法（RMFNet）。首先，在编码器层构建4组优化的残差卷积块（RC）用于提取特征信息，每组残差卷积块通过三重跳跃连接方式提高网络的泛化能力；其次，构建基于空洞空间金字塔池化的多尺度特征融合块（MFF），融合语义信息、全局上下文信息以充分学习海岛变化与未变化的特征；然后，使用损失函数指导残差卷积块和多尺度特征融合块的训练；最后，以中国香港岛为例，基于公开的变化检测OSCD数据集进行仿真实验。结果表明：提出的RMFNet方法的Kappa值比CNN、ResNet-18、PSPNet、SegNet、UNet五种方法分别提高0.250 9、0.201 9、0.131 3、0.078 6、0.038 0，验证了该方法的有效性。更多还原

【Abstract】 In order to improve the accuracy of detecting island change in multi temporal remote sensing images, an encoding and decoding structure model is used to treat the binary classification problem of change and invariance in change detection as a semantic segmentation task, and a residual convolution and multi-scale feature fusion method for detecting change in island multi-temporal remote sensing image is proposed. Four sets of optimized residual convolutional(RC) blocks are constructed at the encoder layer for extracting feature information. Each set of RC blocks can enhance network′s generalization ability by means of triple skip connections. Multi-scale feature fusion(MFF) block based on pyramid pooling of empty space is bulit, which fuses semantic information and global context information to fully learn island changing and invariance features. The loss function is used to guide the training of RC block and MFF block. Taking Hong Kong in China as an example, simulation experiments were conducted on the basis of publicly available change detection OSCD dataset. The results show that the Kappa values of the proposed method are 0.250 9, 0.201 9, 0.131 3, 0.078 6 and 0.038 0 higher than those of CNN, ResNet-18,PSPNet, SegNet, and UNet, respectively, verifying the effectiveness of this method更多还原