This study examines the applicability of the building extraction AI model released by the Geospatial Information Authority of Japan (GSI) to fire-damaged urban areas following the 2024 Noto Peninsula Earthquake. The target area is the central district of Wajima City, which suffered extensive structural losses due to post-earthquake fires. Orthomosaic images were generated from aerial photographs and automatic building extraction was performed using the pretrained convolutional neural network (CNN) model. The results demonstrated that the GSD 11cm image, consistent with the training data specifications, enabled effective detection of building distributions. By overlaying these results with the fire-burned areas, the spatial extent of the damage was intuitively visualized. In contrast, the GSD 2.3cm image produced numerous misclassifications, as the tile size was too small to capture entire building footprints. These findings reveal that the accuracy of AI-based building extraction is highly dependent on input resolution. This case study highlights both the potential and limitations of AI interpretation for rapid disaster assessment and suggests improvements, including multi-scale model development, continuous dataset updates, and cloud-based real-time sharing mechanisms.
