Pakistan's PV Atlas

Best estimate: 18,827 MWp, with a 90% range of 16,022–24,358 MWp. That range covers three specific, measured sources of uncertainty -- but not everything that could move this number. Below: what's inside the range, what isn't, how many ground-truth areas it rests on, and how it compares to two unrelated data sources.

What's inside the range:

• The two "panel area to power" conversion numbers. One converts rooftop panel area to kWp, the other converts open-ground solar-farm land to kWp. Both are measured against real, confirmed power plants rather than assumed, but each carries its own uncertainty.
• How well the detection model finds panels of different sizes. Its measured precision and recall were checked installation-size by installation-size, and that check itself has a margin of error.
• How much the small-panel corrections shift depending on which neighborhoods were ground-truthed. Two corrections -- how much of a flagged roof is actually covered in panels, and what share of real installations get flagged at all -- are fit on the same set of ground-truthed neighborhoods ("quadrats"). This source of uncertainty is measured by refitting both on random subsets of those quadrats and seeing how far the answer moves.

What's outside the range -- and can't be added back in with more arithmetic:

• The ground-truth areas were hand-picked, not randomly sampled, so this isn't a formal statistical margin of error. That wasn't a shortcut: a genuine random sample needs every sampled location to have recent-enough reference imagery to confirm or rule out a small installation, and random locations outside the calibrated areas have so far landed on imagery too old to tell "no panels" apart from "panels installed after this photo was taken." Hand-picking was the fallback that let ground-truth areas be placed where recent-enough imagery actually exists.
• Ground-truth "complete" means complete as of when that area was mapped, not as of the satellite image used for detection. That cuts both ways, but in the same direction: it makes the model's measured accuracy look slightly worse than it is (recent real installations get scored as false alarms) and its measured miss rate look slightly better than it is (installations built after mapping can't be missed if they were never counted as ground truth to begin with). Both effects point the same way -- this page's figure is more likely an undercount than an overcount.
• Most of the Best estimate leans on one correction that's only lightly tested where it's applied most. That correction is fit using ground-truth areas as sparse as 124 buildings/km² at the sparsest of them, but about 13.5% of the buildings it's actually applied to nationally are still sparser than that (measured 2026-08-20; this was a much larger gap -- 872 buildings/km² sparsest, 84% of buildings sparser -- as measured 2026-08-16, before a later refit happened to admit a sparser ground-truth area into the fit). The range above only resamples the areas the correction was fit on -- it says nothing about how well that correction holds up in the sparser areas outside that fit.

Treat this as an early-stage estimate from an active research pipeline, not a finished census. What's genuinely new here -- a reproducible way to estimate distributed solar from free satellite imagery and open-source AI, in a country where official statistics are sparse -- holds regardless of whether any single number on this page turns out exactly right. Expect these figures to keep moving as more ground-truth areas get added.

The ground-truth areas: hand-picked to cover a mix of landscapes, not a random sample. All 24 quadrats behind the small-panel instruments were chosen by a researcher to span different kinds of places -- planned housing developments, dense informal urban neighborhoods, industrial estates, arid/bare land -- rather than drawn at random from a national list. Only 24 of them have had a full manual check thorough enough to trust their "no panels here" verdicts (the teal markers on the map). However many quadrats exist, hand-picked ones can't produce a formal national margin of error on their own -- that needs a genuine random sample of the country's buildings, which doesn't exist yet. More quadrats do help: each new one added so far has turned up a new way the method can go wrong.

Two independent, non-satellite data sources land in the same ballpark. Pakistan's NEPRA net-metering register -- a government administrative record with no connection to this project -- puts registered rooftop solar at 5.3–6.3 GW nationally; that's a floor, since it only counts customers who completed formal registration paperwork. Separately, Chinese customs export data puts cumulative solar-panel imports into Pakistan at roughly 50 GW by mid-2025 -- a much looser ceiling that covers the whole market, utility-scale plants included. This page's headline figure falls inside that bracket. Two unrelated, non-satellite sources agreeing on the same order of magnitude is real corroboration -- though it can't confirm any single number on this page precisely.