VE3 developed a cloud-native building footprint extraction pipeline built around a Mask R-CNN instance segmentation architecture, trained on a stratified sample of aerial imagery tiles drawn from across the national coverage area. The training corpus was deliberately designed to capture the full diversity of English and Welsh building typologies, including terraced housing, semi-detached suburban stock, detached rural dwellings, industrial units, agricultural buildings, and complex multi-wing institutional structures. Pre-processing involved reprojecting all input imagery and vector reference data into British National Grid (EPSG:27700) and harmonising tile boundaries to eliminate seam artefacts at processing block edges.

The pipeline integrated directly with the client's existing Esri-based GIS environment through a custom PostGIS interface, meaning all outputs were immediately queryable within their operational tooling without intermediate file conversion steps. GPU-accelerated inference was deployed on AWS EC2 instances using a Kubernetes orchestration layer that autoscaled compute resources based on the processing queue depth, enabling sustained throughput of over 100,000 structures per day during peak processing windows. Each extracted footprint was accompanied by a confidence score and a set of QA flags indicating cases where the model's certainty fell below predefined thresholds, allowing the client's GIS team to prioritise manual review effort on the highest-risk outputs rather than reviewing the entire dataset.

Topology correction was applied as a post-processing step, using GEOS-based geometry validation routines to close gaps, resolve overlapping polygon boundaries, and align building outlines to the underlying OS MasterMap reference geometry. This was critical for ensuring downstream usability of the extracted footprints in planning and infrastructure applications, where topological integrity is a hard requirement.

One of the most significant challenges was handling the prevalence of terraced and semi-detached properties in urban areas, where shared party walls mean that individual building footprints cannot be separated using height or spectral information alone. VE3's approach incorporated adjacency pattern learning into the Mask R-CNN architecture, training the model on manually verified instance masks that encoded the correct separation lines between adjoining structures. Greenhouses, garden rooms, and outbuildings presented a secondary challenge; these were addressed through a combination of minimum area thresholds and geometric compactness filters applied in post-processing, calibrated against OSMM descriptive attributes where available.

For large complex structures such as schools, hospitals, and warehouses — which frequently exhibit highly irregular plan forms — the pipeline applied a hierarchical segmentation strategy, first identifying the primary building envelope and then resolving internal courtyards and covered linkways as separate geometries where appropriate.

‍