One recall-first model read a year of Sentinel-2 imagery over every building-populated cell of Gujarat. How much photovoltaic capacity it saw depends on how honestly you count: the same probability rasters support six defensible estimates, from a raw-detection floor to a recall-corrected estimate of the whole detectable population. Switch between them - the map, the hero number and the province ranking follow. Rooftop and ground-mount area convert at different rates, because a rooftop detection outlines the panels and a ground-mount detection outlines the site.
Same model, same rasters, different accounting. Whiskers are 90% credible intervals propagated from the calibration's measured counts. Click a bar to put that estimator on the map.
The calibrated points sit at or below their bands by construction: where the glint instrument cannot pin down candidate precision, the point estimate clips it to zero rather than guess. The interval, not the point, is the uncertainty statement.
Thresholded (p ≥ 0.3) candidate polygons intersected with VIDA building footprints, taken at face value. Consistent with the human-validation leads product, but the model is deliberately recall-first, so this both includes false positives and misses everything sub-threshold.
Detected area with every candidate weighted by its measured P(real | size, glint): the size-binned fraction already mapped in OSM, plus glint corroboration and manual review of the unmapped remainder. The conservative number to quote for rooftop PV.
Probability-weighted area: every pixel's probability × 100 m² summed over each footprint, no threshold. Integrates sub-threshold and sub-400 m² signal the polygons cannot resolve; leans high because false-positive probability mass is summed too.
Calibrated area of rooftop-placed candidates divided by the model's measured per-size recall (0% under 100 m² → 100% over 5 ha, measured against a pre-pipeline OSM reference), clamped at 0.05. A Horvitz–Thompson estimate of the whole detectable rooftop population, missed installations included.
The same precision weighting applied to every candidate -- ground-mounted farms and installations with no mapped building included. This is the scope jump: utility-scale PV dwarfs the rooftop total.
The fullest defensible number: all-PV calibrated area, recall-corrected per size bin, with a 90% credible band from posterior draws over every calibration count. The correction is modest where the area-dominant ≥ 5,000 m² bins are already almost fully recalled.