Mapping building change detection in very high-resolution remote sensing imagery with a deep learning method: Fid-Bau Portal

Mapping building change detection in very high-resolution remote sensing imagery with a deep learning method

Jiang, Zhuoyi / Zhang, Hua / Yang, Zijun / Guo, Jun / Wang, Zhipan

Building change detection is a crucial task for urban area renewal and earthquake disaster emergency management. However, the conventional assessment method mainly relies on field investigation and visual interpretation of images, which entails drawbacks such as low efficiency,and high cost, and information abstraction. These shortcomings make it challenging to meet the requirements of urban spatial change detection and disaster response and recovery. Therefore, this paper proposes a building change detection method based on convolutional neural network (CNN) to detect changes in buildings in multi-temporal remote sensing images, which was verified using two remote sensing images before and after the Türkiye earthquake on February 6, 2023. In this method, we improved DenseNet 121 to serve as a key component of the deep learning model, utilizing dense connections between layers to improve feature propagation efficiency and enhance the capture of graphic details. Simultaneously, we designed a novel loss function based on the Tanimoto coefficient to measure the difference between the predicted segmentation map and the live segmentation map. Experimental results show that the proposed building change detection method is effective in detecting building collapse in earthquake disasters, with a quantitative accuracy of F1 score exceeding 70%. Additionally, the model design approach for building change detection presented in this paper can provide valuable insights for other element detection tasks based on remote sensing images, such as road detection.