In the field of remote sensing applications, semantic change detection (SCD) simultaneously identifies changed areas and their change types by jointly conducting bitemporal image classification and change detection. It facilitates change reasoning and provides more application value than binary change detection (BCD), which offers only a binary map of the changed/unchanged areas. In this study, we propose a multitask Siamese network, named the semantic feature-constrained change detection (SFCCD) network, for building change detection in bitemporal high-spatial-resolution (HSR) images. SFCCD conducts feature extraction, semantic segmentation and change detection simultaneously, where change detection and semantic segmentation are the main and auxiliary tasks, respectively. For the segmentation task, ResNet50 is used to conduct image feature extraction, and the extracted semantic features are provided to execute the change detection task via a series of jump connections. For the change detection task, a global channel attention (GCA) module and a multiscale feature fusion (MSFF) module are designed, where high-level features offer training guidance to the low-level feature maps, and multiscale features are fused with multiple convolutions that possess different receptive fields. In bitemporal HSR images with different view angles, high-rise buildings have different directional height displacements, which generally cause serious false alarms for common change detection methods. However, known public building change detection datasets often lack buildings with height displacement. We thus create the Nanjing Dataset (NJDS) and design the aforementioned network structures and modules to target this issue. Experiments for method validation and comparison are conducted on the NJDS and two additional public datasets, i.e., the WHU Building Dataset (WBDS) and Google Dataset (GDS). Ablation experiments on the NJDS show that the joint utilization of the GCA and MSFF modules performs better than several classic modules, including atrous spatial pyramid pooling (ASPP), efficient spatial pyramid (ESP), channel attention block (CAB) and global attention upsampling (GAU) modules, in dealing with building height displacement. Furthermore, SFCCD achieves higher accuracy in terms of the OA, recall, F1-score and mIoU measures than several state-of-the-art change detection methods, including deeply supervised image fusion network (DSIFN), the dual-task constrained deep Siamese convolutional network (DTCDSCN), and multitask U-Net (MTU-Net).

在遥感应用领域，语义变化检测（SCD）通过联合进行双时相图像分类和变化检测，同时识别变化区域及其变化类型。与仅提供已更改/未更改区域的二进制映射的二进制更改检测 (BCD) 相比，它有助于更​​改推理并提供更多应用价值。在这项研究中，我们提出了一个多任务连体网络，称为语义特征约束变化检测 (SFCCD) 网络，用于在双时态高空间分辨率 (HSR) 图像中构建变化检测。SFCCD同时进行特征提取、语义分割和变化检测，其中变化检测和语义分割分别是主要任务和辅助任务。对于分割任务，ResNet50用于进行图像特征提取，提取的语义特征通过一系列跳转连接提供给执行变化检测任务。对于变化检测任务，设计了全局通道注意力（GCA）模块和多尺度特征融合（MSFF）模块，其中高层特征为低层特征图提供训练指导，多尺度特征与多个卷积融合拥有不同的感受野。在不同视角的双时相 HSR 图像中，高层建筑具有不同的方向高度位移，这对于常见的变化检测方法通常会造成严重的误报。然而，已知的公共建筑物变化检测数据集通常缺乏具有高度位移的建筑物。因此，我们创建了南京数据集（NJDS）并设计了上述网络结构和模块来解决这个问题。在 NJDS 和另外两个公共数据集，即 WHU 建筑数据集 (WBDS) 和 Google 数据集 (GDS) 上进行了方法验证和比较实验。NJDS 上的消融实验表明，GCA 和 MSFF 模块的联合利用性能优于几个经典模块，包括空洞空间金字塔池 (ASPP)、高效空间金字塔 (ESP)、通道注意块 (CAB) 和全局注意上采样。 GAU) 模块，用于处理建筑物高度位移。此外，SFCCD 在 OA、召回、F1 分数和 mIoU 测量方面比几种最先进的变化检测方法（包括深度监督图像融合网络（DSIFN））具有更高的准确性，