- The paper demonstrates a scalable AI-based pipeline for extracting lowest floor elevations from street-view imagery to enhance property-level flood risk analysis.
- It integrates direct computer vision extraction with ML imputation using terrain and hydrologic predictors, achieving R² values up to 0.974 in selected regions.
- The approach enables actionable risk assessment by merging extracted/imputed data with high-resolution flood models, guiding emergency management and insurance decisions.
Property-Level Flood Risk Assessment via AI-Enabled Street-View LFE Extraction and ML Imputation
Introduction and Motivation
Accurately assessing flood risk at the property level critically hinges on precise determination of the lowest floor elevation (LFE) for each structure. Since LFE directly anchors depth-damage functions and largely governs structural vulnerability, its absence introduces significant uncertainty into regional loss modeling. Traditional survey-based LFE data collection (e.g., through FEMA Elevation Certificates) is rarely feasible at scale due to logistical and financial constraints, causing persistent gaps, especially across large, diverse regions like Texas. This work systematically addresses the paucity of building-specific elevation data by leveraging advances in AI-based computer vision for GSV street-view imagery, combined with ML imputation grounded in physically meaningful terrain and hydrologic predictors, to construct a scalable, performance-gated pipeline for flood risk quantification (2604.01153).
Methodological Framework
The study develops a three-stage workflow deployed across 18 Texas areas of interest (AOIs), encompassing over 12,000 residential structures. The pipeline comprises:
- Direct Extraction of LFE and HDSL from Street-View Imagery: Using Elev-Vision, a panoptic segmentation method built on OneFormer and text-prompting (CLIP+SAM), the system detects the door bottom and roadside grade in high-resolution equirectangular panoramas. The vertical offset computation incorporates georeferenced camera metadata and depth maps, yielding both absolute LFE and height difference relative to street grade (HDSL).
- Performance-Screened Machine Learning Imputation: Where direct extraction is infeasible (e.g., due to facade occlusion, vegetation, or missing imagery), Random Forest and Gradient Boosting models impute HDSL from 16 engineered features spanning geographic coordinates, terrain, hydrologic context (notably HAND and stream proximity), and local flood exposure. Exclusion of AOIs with insufficient cross-validated generalization (R2_CV threshold) ensures that no unreliable values contaminate the downstream risk model.
- Integration with Flood Surfaces and Economic Damage Estimation: The combined extracted/imputed LFE dataset is merged with high-resolution 1-in-100-year Fathom inundation rasters. Interior flood depth (FDIS) and economic loss per structure are calculated using USACE depth-damage functions, permitting loss aggregation at both property and regional scales.
Results
Imagery Coverage and Extraction Yield
Across all 18 AOIs, GSV panoramas were obtained for 73.4% of structures, with direct LFE/HDSL extraction succeeding for 49.0%. The primary bottleneck was not street-view coverage but facade visibility—front-door occlusion or absence was decisive. Extraction performance was consistent across diverse urban and rural settings, demonstrating that the underlying computer vision pipeline is robust to heterogeneous architectural styles and local topography.
Machine Learning Imputation Efficacy
ML imputation was executed for 13 AOIs where sufficient extracted HDSL data permitted meaningful model training and validation. Among these, the best models achieved R2_CV values from 0.159 (high heterogeneity, e.g., Harris County Pocket 1) up to 0.974 (homogeneous post-Harvey subdivisions, e.g., Brazoria County Pocket 2), with RMSE as low as 0.035 m. Critical features included HAND, local absolute elevation, proximity to stream networks, and flood exposure metrics. Notably, five AOIs were excluded from imputation due to systematically poor cross-validated fit or high overfitting, underscoring the importance of model quality control and transparent boundary conditions.
Impact on Flood Risk and Loss Characterization
Integration with Fathom flood surfaces generated structure-specific interior flood depths and economic loss estimates. Results revealed spatial clustering of high-exposure parcels, strongly correlated with grading practices and developed era. In representative AOIs (e.g., Brazoria County Pocket 1), the LFE/HDSL distributions display clear bimodality reflecting both pre- and post-flood mitigation construction, with anticipated economic loss concentrated in lower-elevation clusters. Imputation extended the fraction of structures with quantifiable flood vulnerability well beyond what image-only extraction would have permitted, materially improving the regional risk profile.
Practical and Theoretical Implications
The pipeline represents a paradigm shift in regional flood risk analytics—from exposure-centric, area-averaged metrics toward building-level interior damage estimation, enabling sharper identification of at-risk assets and prioritization of mitigation investments. Strict imputation gatekeeping (by AOI, based on validated out-of-sample accuracy) raises the standard for reliability in regional-scale ML workflows; AOI-level performance metrics become indispensable for decision transparency and reproducibility.
Practically, jurisdictions lacking comprehensive Elevation Certificate coverage can adopt this workflow using open data sources and public APIs, incrementally improving datasets through staged imagery acquisition and dataset updates. The blended extraction-imputation approach provides actionable loss estimates for emergency management, insurance underwriting, and floodplain regulation—particularly critical in Texas given its historic, multi-billion-dollar flood exposure.
Theoretically, the work advances the state of the art by:
- Demonstrating the generalizability and scalability of deep learning-based elevation extraction to tens of thousands of parcels in diverse settings.
- Establishing performance-screened ML imputation as the preferred alternative to spatial interpolation for complex structural attributes like HDSL.
- Providing a replicable model for exclusion-based imputation, which transparently communicates geographic and methodological limits rather than overfitting or hallucinating data where prediction is unjustified.
Limitations and Prospects for Future Research
The study's regional screening orientation is explicit: it does not supplant engineering-grade LFE surveys, and its downstream loss estimates inherit compounded uncertainty from imagery quality, model imputation, flood surfaces, and damage functions. Damage model calibration remains tailored to slab-on-grade residential construction (per USACE DDF); future work should develop function variants for pier-and-beam or mobile homes. Multi-angle/multi-modal imagery integration, transfer learning across AOIs, and validation against claims records are recommended avenues for expanding extraction/imputation coverage, refining uncertainty quantification, and aligning model outputs with observed post-event damage.
Conclusion
This paper establishes that AI-enabled street-view segmentation and ML imputation can address the chronic absence of regionally complete LFE data, scaling from pilot studies to operational workflows for property-level flood risk and loss analysis. By integrating robust, AOI-gated extraction and imputation with physically referenced inundation and damage modeling, the framework advances both methodological rigor and practical utility in flood risk governance. Future extensions focusing on multi-modal imagery assimilation, imputation transferability, and post-event validation will further strengthen the pathway from regional screening to actionable resilience planning.
Citation:
"Property-Level Flood Risk Assessment Using AI-Enabled Street-View Lowest Floor Elevation Extraction and ML Imputation Across Texas" (2604.01153)