SAFNet: Artificial Intelligence–Based Multisource Heterogeneous Data Fusion for Comprehensive Building Attribute Prediction

Accurate assessment of structural attributes is vital for flood risk evaluation and mitigation planning. Manual surveys for such data are expensive and time-consuming, while automated methods often rely solely on image-based models or narrowly focused data fusion, limiting their effectiveness. This study introduces Structural Attribute FusionNet (SAFNet), an artificial intelligence (AI)-driven framework combining Google Street View (GSV) images with multisource heterogeneous data, such as census and building-specific attributes, to improve prediction accuracy. SAFNet utilizes convolutional neural networks (CNNs) or transformers for image feature extraction, integrated with fully connected networks for multisource data fusion. The model predicts six structural attributes critical for flood risk assessment: first floor elevation (FFE), foundation type, building type, number of stories, presence of basements, and presence of garages. SAFNet improves the mean absolute error (MAE) for FFE predictions by up to 2.17 cm and F scores for classification tasks by 2.94%–11.58%, outperforming purely image-based models (dubbed GSVNet throughout this study). Notably, SAFNet excels in underrepresented attribute classes, such as commercial buildings and structures with basements. A simpler variant of SAFNet, fusing imagery with census data, is also proposed for scenarios lacking individual building information. SAFNet is being used to predict structural attributes for 1.6 million buildings as part of the Louisiana Watershed Initiative, aimed at developing flood risk mitigation models in the flood-vulnerable state of Louisiana, demonstrating its real-world impact. The model is highly efficient and can process all the buildings in the state in <4 h per attribute. These results highlight the robustness of fusing GSV imagery with contextual data, offering a reliable framework for improving flood risk modeling and informing policy decisions.