First-Floor-Finder (F3): Reliability-aware multi-view fusion for automated first-floor height estimation of suburban buildings

Flooding poses a substantial threat to suburban infrastructure, motivating the need for scalable approaches for flood-risk assessment. A key parameter in evaluating a building’s susceptibility to flood damage is its first-floor height (FFH), an engineering decision variable typically obtained through labour-intensive surveying. Traditional land-surveying methods are precise but labour-intensive and impractical at scale. Existing computer-vision approaches show promise, but many rely on single-view cues and high-quality imagery, making them sensitive to occlusion and sparse viewpoints in residential scenes. This paper presents First-Floor-Finder (F3) , a multi-view, multi-stage framework for automated FFH estimation that integrates 2-D imagery with 3-D LiDAR geometry. F3 detects and refines three façade cues—front door, stoop, and basement window—across multiple viewpoints, then combines the resulting candidate height measurements using a Deep Feature Fusion Network (DFFNet) that assigns adaptive weights based on per-candidate reliability. Experiments on an independently collected, occlusion-prone dataset show that F3 achieves 16.3 cm Root Mean Squared Error (RMSE), 13.4 cm Mean Absolute Error (MAE), and 94.1% availability. Ablation studies indicate that multi-cue fusion and learned weighting improve both accuracy and coverage (availability) under foreground occlusion and variable façade configurations. These results support the use of F3 as a scalable approach for suburban FFH estimation in flood-risk analytics, by formalizing FFH inference as a reliability-aware fusion of heterogeneous visual and geometric evidence.