Building Classification

The output is a categorical raster at a spatial resolution of ~10 m assigning each settlement pixel to a building-type class. Built-up areas are delineated using DLR’s World Settlement Footprint (WSF) – a global mask of human-made structures derived from multi-sensor satellite imagery – or an equivalent Built-up Extent layer. Settlement space is partitioned into Local Neighbourhoods (LNs) using a combination of OpenStreetMap (OSM) networks and road geometries from Meta labelled “roads missing from OSM,” constrained by the built-up mask. For each LN, zonal indicators are computed including spectral statistics and vegetation indices from Sentinel-2, night-time lights, built-up height proxies (e.g., WSF-3D or equivalent), imperviousness, terrain slope, proximity to key points of interest, and shape metrics.

A hierarchical decision tree first applies simple rules (e.g., rural vs urban; informal vs formal), then a machine-learning model (e.g., Random Forest) separates pixels into more fine-grained residential types. Each built-up pixel inherits the class of its LN. Variants of this capability compute indicators per building footprint or use VHR imagery with deep learning for per-building outputs where data and VHR imagery licensing permit.

Indicative Cost Range Details

Costs vary with (i) size of the Area of Interest (AOI) and settlement share, (ii) input stack (the baseline version uses only open data, whereas the VHR variants require licensing of commercial imagery), (iii) reference data for training/validation, (iv) update cadence (one-off vs periodic), and (v) delivery/integration needs (tiling, symbology, dashboards/APIs). As a guide, city-scale pilots (≤ ~500 km²) using the baseline HR stack typically start in the low five-figure EUR range. Larger regions, national-level products, or per-building VHR variants are quoted on a case-by-case basis. In those scenarios, commercial imagery licensing and expanded QA/validation are usually the primary cost drivers.