What the Best estimate is built from

Reading this breakdown

roofclf alone accounts for roughly four-fifths of the Best estimate - its own ≥ 400 m² replacement inside the density-calibrated domain and its sub-400 m² "central" estimate together dwarf every other source. Direct OSM mapping and TerraMind segmentation (ground-mount everywhere, plus rooftop outside the calibrated domain) split roughly the remaining fifth. SPPI's role in Best is deliberately small: it only appears in the out-of-domain extrapolation, a strict, explicitly-flagged judgement band with no calibration quadrat of its own - never blended into the in-domain roofclf estimate. SPPI's larger role is in Verified, the stricter floor, where roofclf-AND-SPPI agreement is almost as large as all of hand-mapped OSM combined.

The two roofclf slices in Best are recall-corrected; the floor's is not. Both Best-estimate roofclf components divide by the measured share of real installed panel area that the classifier actually flags (0.81 below 400 m², 0.98 at or above it), so they estimate the whole population of each size class rather than only the roofs the classifier caught — the same correction the segmentation slices have always carried. The roofclf-AND-SPPI slices are deliberately left uncorrected: a floor built by requiring two independent detectors to agree stops being a floor the moment it extrapolates to installations neither of them saw.