Up-to-date building maps are fundamental to urban development and analysis. However, detecting buildings from images with different spatial resolutions and ground object patterns from various imaging sensors is a challenge. Current models mostly have difficulties in extracting buildings with poor boundaries due to the various building appearances and sizes. Moreover, most published methods are trained and evaluated on subsets from the same dataset whose images are captured from one imaging sensor with similar ground object patterns, making it difficult to evaluate the transferability objectively. To address this issue, a spiking neural network in the frequency domain (SNNFD) is proposed to enhance the model transferability and the feature capability of buildings with different sizes by synthesizing frequency domain and spatial domain learning. Spiking convolution is adopted in the frequency learning module to enhance the model learning ability by mimicking the learning process of human brain. The learned frequency features are concatenated and transformed to the spatial domain, and used to generate building-extraction result images by convolution networks. SNNFD is evaluated on two datasets with different spatial resolutions (0.3–2.5 m) from different imaging sensors (Quickbird, Worldview, IKONOS, ZY-3) of different study areas (worldwide). It is compared with five recently proposed semantic segmentation frameworks (Unet, Segnet, DeepLabv3, BiSeNet, F³-Net), and obtains a minimum of 6.33 % higher accuracy with a strong transferability in detecting different sizes of building instances. Specifically, the proposed model improves the segmentation performance of small building instances by at least 6.5 % compared with the five segmentation frameworks through synthesis of spiking convolution in the frequency learning domain in the model construction. Moreover, details of building boundaries are better maintained by SNNFD, offering the possibility of detecting buildings for practical applications.

最新的建筑地图是城市发展和分析的基础。然而，从具有不同空间分辨率的图像和来自各种成像传感器的地物图案中检测建筑物是一项挑战。由于建筑物外观和大小的不同，当前模型大多难以提取边界较差的建筑物。此外，大多数已发表的方法都是在来自同一数据集的子集上进行训练和评估的，这些数据集的图像是从具有相似地物图案的一个成像传感器捕获的，因此很难客观地评估可转移性。为了解决这个问题，提出了一种频域脉冲神经网络（SNNFD），通过综合频域和空间域学习来增强模型的可迁移性和不同大小建筑物的特征能力。频率学习模块采用尖峰卷积，通过模仿人脑的学习过程来增强模型学习能力。学习到的频率特征被连接并转换到空间域，并用于通过卷积网络生成建筑物提取结果图像。SNNFD 在来自不同研究区域（全球）的不同成像传感器（Quickbird、Worldview、IKONOS、ZY-3）的具有不同空间分辨率（0.3-2.5 m）的两个数据集上进行评估。它与最近提出的五个语义分割框架（Unet、Segnet、DeepLabv3、BiSeNet、F 并用于通过卷积网络生成建筑物提取结果图像。SNNFD 在来自不同研究区域（全球）的不同成像传感器（Quickbird、Worldview、IKONOS、ZY-3）的具有不同空间分辨率（0.3-2.5 m）的两个数据集上进行评估。它与最近提出的五个语义分割框架（Unet、Segnet、DeepLabv3、BiSeNet、F 并用于通过卷积网络生成建筑物提取结果图像。SNNFD 在来自不同研究区域（全球）的不同成像传感器（Quickbird、Worldview、IKONOS、ZY-3）的具有不同空间分辨率（0.3-2.5 m）的两个数据集上进行评估。它与最近提出的五个语义分割框架（Unet、Segnet、DeepLabv3、BiSeNet、F³ -Net），并在检测不同大小的建筑实例时获得了至少 6.33% 的准确率，并且具有很强的可迁移性。具体来说，与五种分割框架相比，所提出的模型通过在模型构建中在频率学习域中合成尖峰卷积，将小型建筑实例的分割性能提高了至少 6.5%。此外，SNNFD 可以更好地维护建筑物边界的细节，为实际应用中检测建筑物提供了可能性。