Small ground-mounted PV has no instrument: measured, 2026-08-16¶

PARTLY ADDRESSED (2026-08-16): the accounting fix shipped, the instrument did not

Option 2 below ("widen roofclf's label rather than add a detector") was implemented the same day, on the owner's instruction, ahead of the cropland quadrats this page recommends first. earthpv roof-classifier --parcel-label now counts mapped PV within 20 m of a footprint, attributed to its nearest building, below the 400 m² segmentation floor. Option 3 still stands: no standalone detector was built, and the AND-gate measured below was not shipped.

Two results from implementing it, both of which change how the fix should be read:

Most of what the parcel label adds is not ground-mount. Across the 27 quadrats it adds 146,766 m², of which 29,764 m² (20%) is OSM placement=ground and 117,003 m² (80%) is mapped ROOFTOP PV whose polygon extends past an imagery-derived VIDA footprint. The second is a real undercount in the roof-only label, and correcting it is self-consistent (the same undersizing shrinks the calibration denominator and the national flagged roof area), but it is not the gap this page is about.
The yard feature block still does not earn its place, even against the widened label. Median fold AUC 0.8712 with it against 0.8734 without. It ships off.

Scored nationally, validated and published 2026-08-17, the widening moves Best estimate 18,218.4 to 19,745.9 MWp. That increase should not be read as a ground-mount figure: of the PV area the parcel label now prices on flagged buildings, 1.03% is ground-tagged and 4.84% is rooftop overhang, and roughly two thirds of the sub-400 m² move comes from the larger flagged population rather than from the widened area term at all. The small ground-mount gap this page is about is therefore still open: the published atlas now counts yard arrays it can see beside a flagged building, and still has no instrument for a field-sited array with no building to attach to.

The capacity range below (0.4 to 3.5 GWp) is unchanged by any of this and still rests on four settlement-drawn quadrats over 16 km². Cropland quadrats remain the binding input.

OPEN, now with measurements (as of 2026-08-16)

The 2026-08-13 version of this page established the gap on structural grounds and sized it against two external estimates, but explicitly measured nothing ("No code was written, no data was pulled beyond what's already in this project's own quadrat labels"). This version measures it. Nothing has been shipped and no atlas number has moved; the point of the exercise was to decide whether an instrument is buildable and what it would be worth, and three of the four items in the old "what a fix would need" sketch turn out to be answered already by data on disk. The reproduction script is scripts/small_ground_mount_assessment.py.

One-line verdict: the candidate universe problem is solved (the VIDA building index already brackets the population), the detection problem is not (under 1% precision at 30% recall, and 2% for a yard-SPPI/roofclf AND-gate, added 2026-08-16), and the size of the prize cannot be pinned down better than 0.4 to 3.5 GWp until somebody maps cropland quadrats, because 23% of the national building population sits in a density stratum currently represented by four quadrats and 16 km².

The external evidence (unchanged)¶

TransitionZero + PRIED's "Shedding light on Pakistan's distributed solar revolution" (Oct 2025) reports two independently-derived national estimates for Pakistan's distributed solar. The one number that matters here: both converge on roughly 5 GW of solar installed specifically on agricultural land, mostly tube-well irrigation pumps. TransitionZero's satellite-based estimate says 5.64 GW ground-mounted nationally; PRIED's nationally representative household survey (5,320 respondents, stratified random) says 5.04 GW for the agriculture sector. This project's published atlas reports 2.2 GW of ground-mount in total, every size and type combined, and zero below 400 m².

What counts as a small ground-mount array, and why the definition is load-bearing¶

Two independent tests disagree with each other on roughly half the population, so every number below requires both: OSM placement == "ground" and less than 30% of the polygon overlapping a VIDA footprint.

	geometry says ground	geometry says rooftop
OSM tag says ground	300	443
OSM tag says rooftop	592	14,968

Requiring both is the conservative read: it discards 1,035 ambiguous installations and keeps 300, of which 262 are under 400 m², totalling 23,369 m² across the 27 quadrats. Median 64 m², quartiles 38 and 153. This is 17x more labelled examples than the five-quadrat hand count of 15 that the previous version of this page worked from.

The capacity constant is module, not land, and it is a 3.6x error¶

The old sketch's item 4 guessed that a small-ground-mount instrument would "likely inherit" DEFAULT_KWP_PER_M2_LAND. It must not. Split by OSM kind:

kind	placement	count	median area
`generator`	ground	473	63 m²
`plant`	ground	270	1,031 m²
`generator`	rooftop	15,560	45 m²

DEFAULT_KWP_PER_M2_LAND = 0.05 was calibrated on power=plant site perimeters (QASP, Sukkur), where only the ground-cover ratio is module. A small ground array is a power=generator way traced around the panel table itself, so it is module area and converts at DEFAULT_KWP_PER_M2_MODULE = 0.18. The sanity check agrees: the median 64 m² strict array reads 11.5 kWp at the module constant, squarely inside the 5 to 15 kWp a solar tube-well pump actually draws, and an implausible 3.2 kWp at the land constant.

The candidate universe is already built: these arrays sit next to buildings¶

The old sketch's item 1 asked for "a footprint source that isn't VIDA buildings" (cropland parcels, or a sliding tile classifier). That is not needed. Distance from each installation to the nearest VIDA footprint:

within	quadrat installations (n=262)	quadrat PV area	national OSM (n=201)	random point, same cells
5 m	86.6%	85.9%	80.6%	21.3%
10 m	93.9%	93.8%	91.0%	26.7%
30 m	98.5%	97.7%	98.5%	39.9%
100 m	100.0%	100.0%	99.0%	64.5%

The national column is the important one, because it is independent of where the quadrats happen to be drawn: every OSM ground-tagged generator under 400 m² in Pakistan, across 43 cells, against 200 random points per cell as a null. The anchor holds at 2.5x lift at 30 m and 3.4x at 10 m.

And it is cheap. Sampling 48 national cells stratified over all eight building-density octiles, the share of land within 30 m of a VIDA building is 9.6% nationally (1.4% in the sparsest octile, 22.8% in the densest), so a building-anchored search scans a tenth of the country and keeps 98.5% of the population. That is the whole of the old item 1 and item 2, answered with the index roofclf already scores nationally.

Item 3 ("more ground-mount negatives than the 15 positives") is answered the same way: in a Rule-1-complete quadrat every building's yard is a verified negative, so the 27 quadrats yield 118,756 labelled parcels carrying 253 positives with no new mapping at all (253 rather than 262 because a few arrays share a host building and a few sit beyond the 20 m yard depth).

Can the three existing instruments see them?¶

Per pixel, against a matched control: yes, and SPPI leads¶

For each installation, the composite pixel it falls in, against 60 control pixels drawn from the same quadrat and matched on both distance to the nearest building and the fraction of the pixel covered by a footprint. The matching matters: without it, positives sit beside bright roofs that no control could contain, NDBI scores 0.856 AUC on that confound alone, and every figure is inflated by about 0.06.

Within-quadrat AUC, 300 positives against 17,991 matched controls:

feature	all	< 50 m²	50 to 150	150 to 400
SPPI	0.833	0.763	0.832	0.933
NDBI	0.779	0.743	0.756	0.841
fraction head	0.660	0.606	0.625	0.765
segmentation	0.534	0.498	0.499	0.567
all features, LOQO logistic	0.851	0.792	0.847	0.965

Segmentation is at chance, as expected: everything under chips.MIN_PV_AREA is burned as ignore. The fraction head is the instrument with the right shape for this problem (sub-pixel, footprint-free, and chips._burn_fraction does burn ground placements below 400 m²), and it is nonetheless the weaker of the two spectral options. SPPI, which needs no training at all, is the best single feature here.

At the deployable unit, with the real base rate: no¶

The pixel table's 60:1 matched design flatters every instrument. The unit an instrument would actually score is a building, and the base rate is 253 in 118,756. Features are the roof block (what roofclf reads today), a yard block (zonal statistics over the surrounding land out to 20 m, partitioned by a distance-transform Voronoi so no pixel is credited to two buildings), or both. Leave-one-quadrat-out, on all 27 quadrats:

features	AUC	P@R30	P@R50
roof only (= `roofclf` today)	0.670	0.004	0.004
yard only	0.672	0.006	0.005
roof + yard	0.713	0.007	0.005
roof + yard + probability rasters	0.722	0.008	0.005

and on the 99,804 buildings carrying no rooftop PV of their own, where the yard array is the only PV present and so the only thing a score can be responding to:

features	AUC	P@R30	P@R50
roof only	0.678	0.001	0.001
yard only	0.718	0.005	0.002
roof + yard	0.747	0.004	0.002

and, separately, on the six quadrats under 600 bldg/km², which is the stratum holding 77% of the national building population and the closest thing here to the agricultural case:

features	AUC	P@R30	positives
roof only	0.578	0.005	14
yard only	0.622	0.003	14
roof + yard	0.617	0.005	14

Fourteen positives in 4,987 buildings is too thin to call, which is itself the point: the stratum that matters most is the one with almost no ground truth in it.

Precision at 30% recall is 0.4 to 0.8%. For comparison roofclf's rooftop deployment threshold is set at a precision target of 0.5. A leads product at under 1% precision hands a mapper more than a hundred wrong parcels per right one, and an aggregate capacity component at that precision is a correction factor with a measurement attached rather than the other way round.

The yard block is nonetheless a real signal, and specifically a signal about the yard: it moves AUC +0.043 overall and +0.069 on the no-rooftop-PV subset (0.678 to 0.747), where roof features alone cannot help. The two best single features are log_yard_area (0.697) and yard_sppi_max (0.692).

Making SPPI and roofclf agree does not rescue it either¶

The atlas's sub-400 m² rooftop floor is an AND-gate: a building counts only where roofclf and SPPI agree. Transferring that construction here is the obvious next question, so it was measured rather than assumed (small_ground_mount_assessment.py andgate). Both scores are rank-normalised within quadrat first, since SPPI's absolute scale spans 18x across quadrats and a national absolute cut would be gating on quadrat brightness. The roofclf side is the roof-only leave-one-quadrat-out logistic from the table above, which is generous to it: that model is fit on this label, whereas the deployed classifier never saw it. Scoped to the 45,183 buildings that have at least one yard pixel, which is the only population a yard gate can score and which carries a base rate of 0.378%, 4x the all-buildings rate:

gate	flagged	true	precision	recall	lift
roofclf p90	6,983	41	0.59%	0.240	1.6x
yard-SPPI p90	4,534	56	1.24%	0.327	3.3x
AND p90	1,164	22	1.89%	0.129	5.0x
OR p90	10,353	75	0.72%	0.439	1.9x
AND p95	440	9	2.05%	0.053	5.4x
AND p98	132	3	2.27%	0.018	6.0x

The AND-gate does what an AND-gate does, trading recall for a 5 to 6x lift, and it lands at 2% precision. Sweeping both thresholds independently for the best precision at any operating point holding 20% recall gives 2.11%, at roofclf p61 and SPPI p94: the optimum reaches that precision by turning the roofclf side effectively off, and what is left is yard-SPPI alone.

That is the mechanism, and it is worth stating because it does not contradict the rooftop result. The rooftop AND-gate works because roofclf and SPPI are two readings of the same surface, so their disagreements are informative. Move the array two metres off the wall and roofclf is still reading the roof while the evidence is in the yard, so the second vote is not a second opinion, it is a weakly-correlated nuisance that mostly deletes true positives. No threshold pair fixes that, because the problem is what the feature block is pointed at.

The yard block does not help the rooftop task¶

Tested in case the change paid for itself on roofclf's existing job. It does not: rooftop AUC 0.8711 to 0.8709, and 0.8230 to 0.8209 within roof-size band. Clean negative.

Most of this population is already standing inside roofclf's flags¶

This is the finding that most changes what a fix should look like.

168 of 253 buildings with a ground-mount array in the yard also carry rooftop PV, so the pipeline already flags them, for another reason.
Of the remainder, roofclf flags 29.4% against a 9.9% background rate, discriminating yard PV at 0.697 AUC (0.638 within roof-size decile) despite never having been trained on it. At 10 m a yard array bleeds into its neighbour's footprint statistics.
Those flagged buildings hold 30% of the yard PV area on buildings with no rooftop PV.

So the buildings are largely found already. What is missing is not detection but accounting: pv_area_true_m2 is the intersection of mapped PV with the roof polygon, so an array two metres off the wall contributes zero to the coverage ratio and zero to the area recall, on a building the classifier has already flagged. A separately-built ground-mount detector added to the atlas would double-count a substantial share of this.

How big is the prize? Between 0.4 and 3.5 GWp, and the range is the finding¶

The obvious estimator does not work. Per-building rates measured in these quadrats do not transfer: the quadrats' own rooftop PV area is 14,846 m² per 1,000 buildings, which extrapolated over 75.7M national buildings would be 202 GWp against the atlas's actual 7,890 MWp for the same sub-400 m² rooftop population, a 26x overstatement. That is the documented "ranking transfers across quadrats, absolute adoption rates do not" trap (rate_ratio spans 0.2 to 5x here), and it applies to ground-mount exactly as it does to rooftop.

What does plausibly transfer is the ratio measured inside each quadrat: m² of sub-400 m² ground-mount PV per m² of rooftop PV, re-weighted onto the national building population by density stratum, the same stratification sub400_capacity.coverage_ratio_by_size_and_density already uses.

stratum (bldg/km²)	quadrats	rooftop PV	ground PV	ratio	share of national buildings
< 150	4	479 m²	380 m²	0.793	22.9%
150 to 600	2	183,477 m²	1,415 m²	0.008	54.4%
600 to 2,000	11	1,078,632 m²	14,366 m²	0.013	15.1%
>= 2,000	10	500,443 m²	6,643 m²	0.013	7.6%

Pooled, ignoring stratification: 1.29%, which applied to the atlas's 7,890 MWp sub-400 rooftop component is 102 MWp. This is the number a naive read produces and it is dominated by dense urban quadrats.
Building-weighted across strata: 18.9%, or 1,492 MWp, with a quadrat-bootstrap 90% interval of 435 to 3,526 MWp.

The entire spread comes from one cell of that table. The < 150 bldg/km2 stratum has a ground-to-roof ratio sixty times the urban one, holds 22.9% of the national building population, and rests on four quadrats covering 16 km², all of them drawn around villages. Pakistan's cultivated area is roughly 221,000 km², so the field-sited tube-well population that the external estimates actually point at is sampled at 0.007% coverage, and none of that sample is field-sited.

That is the honest position: the project's own ground truth is consistent with anything from a rounding error to two thirds of TransitionZero and PRIED's agricultural figure, and no amount of modelling on the current quadrat set narrows it.

What follows¶

Map cropland quadrats before building anything. Four to six Rule-1 quadrats placed deliberately in irrigated farmland, not around settlements, is the only step that distinguishes a 0.4 GWp gap from a 3.5 GWp one, and it is the input every other option needs anyway. It also does double duty: density.CALIBRATED_BLDG_DENSITY_KM2's floor moves only when a quadrat's own average density reads below it, and CLAUDE.md's own note says such a quadrat "has to be sized and placed to average in enough non-built land on purpose", which is exactly what a cropland quadrat is.
If an instrument is built, widen roofclf's label rather than add a detector. (Implemented 2026-08-16 -- see the banner at the top of this page for what it turned out to measure.) Add the yard block to roofclf.building_table and let pv_area_true_m2 count PV in the parcel rather than only on the roof. The candidate universe, the national scoring loop, the coverage ratio, the area recall, the density-domain restriction, the quadrat bootstrap and the dedup against OSM and segmentation are then all inherited unchanged, and there is still exactly one row per building, so the double-counting problem never arises. The yard partition is a pure raster operation per cell, so the national cost is small.
Do not build a standalone small-ground-mount detector. At 1 to 2% precision it cannot support a leads product, and as a capacity component it would be measured almost entirely outside where it is applied, which is the precise property that got the out-of-domain AND-gate dropped from the atlas on 2026-08-15.
SPPI is the feature to carry over, not the fraction head. It is zero-training, it is the strongest single spectral discriminator here (0.833 per pixel, 0.692 as yard_sppi_max per building), and it already has a role in the pipeline as the disagreeing second opinion.

Not done here¶

No model was trained, no checkpoint was produced, no atlas number moved, and no capacity component was added. The ratio estimates above are not published anywhere and should not be quoted as an earthpv figure until the cropland quadrats exist. The anchor result is measured on settlement-sited arrays plus OSM's own national sample, and OSM mapping is itself settlement-biased, so whether a field-sited tube-well array is equally close to a building is plausible but untested.