Estimation of Fireproof Structure Class and Construction Year for Disaster Risk Assessment

Estimation of Fireproof Structure Class and Construction Year for Disaster Risk Assessment ，Kazushi Okamoto，Atsushi Shibata，Kei Harada，Koki Karube Hibiki Ayabe The University of Electro-Communications Dec 9, 2025 ACM MM Asia 2025 Workshop 1 / 16

Background and Motivation Administrative Use Case Background: Municipal administrations need construction year and fireproof structure class for disaster risk assessment. The fireproof class is a legal rating derived from building structure and property type that indicates how long a building can withstand fire. Challenge: However, building metadata containing this information is often missing or difficult to obtain. Potential Approach: Fireproof class can be inferred from facade images, easing data collection and supporting city-scale risk mapping. Dec 9, 2025 ACM MM Asia 2025 Workshop 2 / 16

Related Work [Hafidz+ ,2024]: Building typology classification using CNNs for city-scale seismic vulnerability assessment in Indonesia. [Chen+, 2022]: Multi-task CNN–based estimation of structural attributes from Street View images for flood risk assessment in the U.S. [Ogawa+, 2023]: Prediction of Japanese building structure and construction year using CNNs and Vision Transformers. Gap: While these studies successfully infer building structures and related attributes, none of the approaches address Japan-specific fireproof class that are essential for domestic firerisk evaluation. We derive Japan's fireproof class from predicted building structure and property type, ensuring interpretability while using only facade imagery. Dec 9, 2025 ACM MM Asia 2025 Workshop 3 / 16

Defining Our Research Questions (RQs) RQ1: Can disaster-relevant building attributes be accurately estimated from residential facade images using a multi-task learning (MTL)? RQ2: How effectively can the fireproof class be inferred from intermediate attribute predictions in a hierarchical modeling approach? Dec 9, 2025 ACM MM Asia 2025 Workshop 4 / 16

Task Definition (Task1, 2, 3 & 4) Task4: Fireproof Structure Class Building Structure Property Type Fireproof Class Concrete Any M (Fireproof) Steel Communal M (Fireproof) Steel Wooden Dec 9, 2025 Non-communal T (Semi-fireproof) Any H (Non-fireproof) ACM MM Asia 2025 Workshop 5 / 16

Multi-Task Model Design Uncertainty-weighted loss [Kendall+,2018]: dynamically weight and combine regression and classification losses. The construction year regression (Task1) acts as an anchor role, helping to stabilize multitask training and balance gradients across tasks. Dec 9, 2025 ACM MM Asia 2025 Workshop 6 / 16

Dataset and Preprocessing Dataset: Large-scale LIFULL HOME'S data (approx. 2.77M training images).[LIFULL,2015] Filtering: Rigorous steps including deduplication and filtering with CLIP to ensure highquality "Entire residential property" images. Representative Sample Facade Year: 2004 Communal / Concrete Year: 1997 Year: 1966 → Fireproof class: M (Fireproof) Non-communal / Steel Non-communal / Wooden → Fireproof class: T (Semi-fireproof) → Fireproof class: H (Non-fireproof) Dec 9, 2025 ACM MM Asia 2025 Workshop 7 / 16

Results I: Construction Year Prediction (Task1) Learning Rate (LR) MAE ↓ RMSE ↓ MedAE ↓ 4.97 6.88 3.59 6.02 7.78 4.90 Considering the MAE (10.689) and RMSE (12.121) reported for CNN-based models by Li et al. [Li+, 2018], our approach achieves favorable estimation accuracy. Dec 9, 2025 ACM MM Asia 2025 Workshop 8 / 16

Results II: Intermediate Classification (Tasks2&3) LR Structure Acc ↑ Structure Macro F1 ↑ Type Accuracy ↑ Type Macro F1 ↑ 1e-5 0.9275 0.8714 0.8321 0.7945 1e-6 0.9182 0.8582 0.8147 0.7721 Dec 9, 2025 ACM MM Asia 2025 Workshop 9 / 16

Summary of RQ1: Feasibility and Interpretation RQ1: Can disaster-relevant building attributes be accurately estimated from residential facade images using an MTL? Answer: The MTL model can estimate disaster-relevant attributes from facade images with practical accuracy construction year MAE 4.97 years building structure acc 92.75% property type acc 83.21% indicating reliable extraction of core features for risk assessment. Dec 9, 2025 ACM MM Asia 2025 Workshop 10 / 16

Results III: Fireproof Class Derivation (Task4) LR Acc ↑ Macro F1 ↑ Weighted F1 ↑ F1 (H / M / T) ↑ 1e-5 0.8916 0.6459 0.8900 0.7880 / 0.9304 / 0.2194 1e-6 0.8802 0.6227 0.8789 0.7692 / 0.9223 / 0.1767 Performance for the T (Semi-fireproof) class is low, primarily due to dataset imbalance. Dec 9, 2025 ACM MM Asia 2025 Workshop 11 / 16

Qualitative Analysis: Interpretation and Error Propagation Negative samples Pred: Steel / Communal →M Pred: Steel / Non-communal →T Pred: Steel / Non-communal →T True: Steel / Communal →M True: Wooden / Communal →H True: Wooden / Non-communal →H By splitting the tasks, we can understand where errors occur in the intermediate attributes, ensuring that explainability is maintained. Dec 9, 2025 ACM MM Asia 2025 Workshop 12 / 16

Summary of RQ2: Robustness of Hierarchy RQ2: How effectively can the fireproof class be inferred from intermediate attribute predictions in a hierarchical modeling approach? Answer: Hierarchical rule-based mapping achieves high overall accuracy (89.16%), but the Tclass F1 is low (~0.22) and requires improvement, recommend data reweighting and Tfocused model refinements. Finding: Even when intermediate attribute predictions are incorrect, the final fireproof class remains correct in about 12% of cases, indicating partial error resilience. Dec 9, 2025 ACM MM Asia 2025 Workshop 13 / 16

Conclusion Summary This study presents an MTL model for estimating disaster-relevant building attributes from facade images, achieving practical accuracy. The model addresses challenges in building metadata by inferring attributes directly from exterior imagery, enhancing disaster risk assessment. Contribution Introduces a novel approach for estimating fireproof structure classes and construction years, tailored to Japan's specific standards. Demonstrates the effectiveness of hierarchical rule mapping in inferring fireproof classes, highlighting areas for improvement in dataset balance and model refinement. Dec 9, 2025 ACM MM Asia 2025 Workshop 14 / 16

Future Work 1. Improving Class Balance: Address dataset imbalance for the T-class through reweighting and data augmentation techniques to enhance prediction accuracy across all fireproof categories. 2. Interpretability Enhancement: Apply explainability methods such as Grad-CAM and LIME to visualize which facade regions contribute to attribute predictions, strengthening model transparency for stakeholders. 3. Regional Risk Assessment: Integrate predictions with Google Street View imagery at scale to generate city-wide and region-specific disaster risk maps for urban planning and emergency preparedness. 4. Global Disaster Risk Framework: Extend the approach beyond Japan by adapting the methodology to diverse building standards and construction practices across different countries and regions. Dec 9, 2025 ACM MM Asia 2025 Workshop 15 / 16

References [Hafidz+, 2024] Hafidz R. Firmansyah et al. 2024. Building typology classification using CNNs for city-scale rapid seismic vulnerability assessment. Eng. Appl. Artif. Intell. 131. [Chen+, 2022] F.-C. Chen et al. 2022. Deep learning-based building attribute estimation from Google Street View images for flood risk assessment. J. Comput. Civ. Eng. 36, 6. [Ogawa+, 2023]Y. Ogawa et al. 2023. Deep Learning Approach for Classifying the Built Year and Structure of Individual Buildings by Automatically Linking Street View Images and GIS Building Data. IEEE J. of Selected Topics in Applied Earth Observ. and Remote Sens. 16. [Kendall+, 2018] A. Kendall et al. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. [LIFULL, 2015] LIFULL Co. 2015. LIFULL HOME’S Dataset. doi:10.32130/idr.6.0 [Li+, 2018] Y. Li et al. 2018. Estimating building age from Google Street View images using deep learning. Dec 9, 2025 ACM MM Asia 2025 Workshop 16 / 16