Pakistan's solar, by installation size

This is the evidence atlas's own Best-estimate total, decomposed by installation size and placement instead of by 0.1° grid cell. Every MWp here is computed by the identical formula the main atlas uses for the same six components (hand-mapped OpenStreetMap installations the model never found, segmentation's own detections above and below the density-matched domain, and roofclf's per-building estimate for small rooftops — both model instruments recall-corrected, each standing in for the installations of its size class it is measured to miss) — nothing is recomputed by a different method, and the grand total matches the evidence atlas's headline figure for the same run.

It does not carry its own credible interval. The main atlas's 90% range applies to the same total; this page exists to show its shape by size and placement, not to re-derive its uncertainty. Read the two pages together, not as competing numbers.

Ground-mount segmentation, every size, no domain restriction

roofclf has no building footprint to score for a free-standing array, so ground-mount is segmentation's own instrument at every scale, from the smallest detected candidate up through utility plants like Quaid-e-Azam Solar Park.

Rooftop, ≥400 m² roofclf in-domain, segmentation elsewhere

Inside the density-matched domain, roofclf's own building-level estimate replaces segmentation's; outside it, segmentation's recall-corrected rooftop total is the only evidence-backed figure. A candidate is counted by exactly one of the two, never both, by cell.

Rooftop, <400 m² roofclf, cross-checked with SPPI

Below the segmentation model's practical floor, every rooftop's PV is estimated per building. Inside the domain this is roofclf alone, calibrated against hand-mapped quadrats for both how much of a flagged roof is panel and what share of real installations get flagged at all; outside it (no calibration quadrat in that density range) it falls back to roofclf-AND-SPPI agreement, a strict, marked extrapolation carrying neither correction — the hatched portion of the small rooftop bars.

Hand-mapped, unmatched OpenStreetMap installations the model missed

Sized by their own real OSM geometry, split rooftop/ground-mount by the mapped placement tag, added wherever the model's own detections don't already carry that installation's geometry (see the evidence atlas's own "How double counting is avoided" section).

Same imagery, same calibration, same domain restriction as the evidence atlas for this AOI and run. Size bins are fixed-width display buckets, independent of the calibration table's own lookup bins (those govern the precision/recall correction per candidate, not how this chart groups results). Full derivations and caveats live in this project's methodology notes (docs/methods/density.md).