Rapid GeoAI assessment tool for a metropolitan local planning authority, Northern England
.png){kind=small}
Rapid GeoAI assessment tool for a metropolitan local planning authority, Northern England
A metropolitan local authority in northern England was undertaking a strategic review of its housing stock in preparation for a Local Plan review that required up-to-date, spatially accurate data on the type, size, and density of existing residential properties across the borough.
The authority's existing property data was held across several disconnected systems council tax records, planning application archives, and a legacy GIS containing building footprints of varying accuracy none of which provided the comprehensive, attributerich spatial dataset the planning team needed. In particular, the authority required reliable estimates of dwelling floor area by property type (detached, semi-detached, terraced, bungalow) across the entire residential stock, as this underpinned housing need calculations and viability assessments for proposed development sites.
The authority did not have in-house geospatial AI capability and required a supplier that could both develop the analytical pipeline and deliver a validated output dataset suitable for direct import into their GIS environment, alongside a methodology that their team could understand, interrogate, and explain to elected members and the Planning Inspectorate if required.
VE3 developed a multi-source geospatial analysis pipeline that fused OS MasterMap building footprints with Environment Agency LiDAR data (1m DTM and DSM tiles covering the full borough) to derive building height profiles and enable automated classification of residential dwelling types. The methodology followed an established nDSM derivation approach subtracting the DTM from the DSM to obtain a normalised surface model representing true building heights above ground which was then aggregated per OS MasterMap building footprint to extract stable height statistics including median roofpoint elevation, 95th percentile height, and roof variance.
Classification of dwelling type (bungalow, one-storey house, two-storey house, and multi-storey residential) was performed using a Random Forest classifier trained on a manually verified sample of 3,500 properties drawn from across the borough, with training labels assigned through a combination of height threshold rules and manual review of aerial photography for edge cases. The classifier incorporated both LiDAR-derived height features and geometric attributes from OSMM including footprint area, perimeter-to-area ratio, elongation, and adjacency topology enabling it to distinguish between, for example, a large single-storey bungalow and a small two-storey terraced property that might exhibit similar maximum heights.
Floor area estimation was derived by multiplying the ground-floor OSMM footprint area by the estimated floor count, with adjustments applied for properties where LiDAR profiling indicated significant roof pitch complexity (predominantly older Victorian stock with steep gable roofs) that would affect habitable floor area relative to the nominal footprint. Garages, outbuildings, and conservatories were filtered from the RCA calculations using rule-based classifiers applied to OSMM feature codes combined with LiDAR height range thresholds.
A significant proportion of the borough's housing stock particularly in its eastern residential areas consisted of back-to-back and courtyard terraces with shared or co-incident rear boundaries and minimal separation between building groups. Conventional footprint-based approaches struggled to correctly delineate individual properties in these areas. VE3 resolved this by using OSMM's existing building polygon topology as the authoritative boundary reference, applying the classification and height extraction logic at the OSMM polygon level rather than attempting to re-segment from aerial imagery, which preserved the legal and administrative property boundaries that the planning team required.
The borough also contained a substantial number of post-war prefabricated properties and system-built housing with non-standard roof forms that caused misclassification in initial model runs. These were addressed through a targeted retraining exercise using labelled examples of the specific non-standard typologies identified during QA review, which reduced the error rate on this sub-population from 18% in the initial model to 6.4% in the final production run.
The final delivered dataset covered all 58,400 residential properties in the borough, with each record attributed with dwelling type classification, estimated floor count, ground-floor footprint area, and estimated total dwelling floor area. The dataset was validated against a stratified random sample of 1,200 properties using a combination of aerial photography review, planning records, and field survey checks, achieving an overall dwelling type classification accuracy of 93.1% and a floor area estimation accuracy of 88.7% within a ±10% tolerance band. The output was delivered as a GeoPackage file compatible with the authority's QGIS environment and accompanied by a methodology report and data dictionary.
.png)
This case study is directly analogous to the VOA Dwelling Area Estimates scope. The client requirement automated floor area estimation across an entire local authority's residential stock using OS MasterMap and LiDAR — mirrors the VOA trial specification almost exactly. The borough scale of 58,000 properties falls within the 50–60k range cited in VE3's infrastructure scalability documentation, meaning this engagement constitutes a practical proof of concept for the proposed trial LA scope. The methodology for handling non-standard property types, the multi-source data fusion approach, and the QA framework applied here will be directly carried forward into the VOA delivery.
.png)
