In the commercial property insurance and real estate technology (PropTech) sectors, traditional site risk assessment relies on manual site visits or manual digitization of aerial imagery—processes that are slow, costly, and difficult to scale across thousands of parcels.

This project delivers a production-ready, automated GeoAI pipeline that automates physical property risk extraction directly from high-resolution satellite/aerial imagery:

• Spatial Data Ingestion & Tile Stitching: programmatically fetches multi-spectral, high-zoom satellite tiles over targeted geographic bounding boxes, stitching them into geo-referenced GeoTIFF rasters with intact CRS metadata (EPSG:3857).

• Zero-Shot Deep Learning Segmentation: Leverages Meta's Segment Anything Model (Geo-SAM / samvitb) via torchgeo to perform feature extraction. The pipeline segments structural footprints, canopy cover, and impervious surfaces without requiring domain-specific manual annotation datasets.

• Vector Topology & Spatial Risk Engine: Converts raw neural network prediction masks into vector geometries (.gpkg). Reprojecting geometries into metric Coordinate Reference Systems (EPSG:32610 / UTM Zone 10N), the engine computes structural metrics (aream2, perimeterm, and compactness) using GeoPandas and Shapely.

• Overhang Risk Quantification: Executes a spatial intersection overlay between building footprint boundaries and surrounding vegetation canopy. It calculates the exact square meters of tree canopy overlapping each roof structure (overhangaream2), generating a normalized property risk score (0–100).

• Interactive GeoJSON Dashboard: Exports clean GeoJSON risk layers and renders an interactive web map using leafmap, displaying raw aerial imagery alongside vectorized, risk-scored building footprints.