PVscope

PVscope puts one 640 m window of Sentinel-2 imagery next to the solar panels mapped by hand inside it, so you can see why most rooftop PV is hard to see from space and what the rooftop classifier has to work with. The window lies in the Peshawar West calibration area, where every visible panel has been mapped, and shows the same pixels twice: stacked by band in space, and moved into spectral space.

Interactive. Drag to orbit, scroll to zoom, click a pixel to pin it. Open full screen.

Reading it

• Image cube. The top layer is the mapped panels and roofs, drawn at 1 m. Below it sits the share of each 10 m pixel the panels cover, then the ten bands from blue (492 nm) down to shortwave infrared (2,202 nm). The six bands that are 20 m at the sensor show as 2 by 2 blocks. Click a pixel to drill a column through every layer, or select an array in the list to outline it on all of them. Peel lifts layers off so any band can be seen from above.
• Spectral space. The same pixels, moved to a space whose axes are three bands of your choice, measured as contrast to the roofs around them. The amber line is where panels push a pixel as their cover grows from none to all of it; the blue ellipses show how far PV-free roofs already scatter. A pixel only stands out once it leaves them.
• Inspector. The selected pixel's spectrum against nearby roofs, its PV cover on the 10 m and the 20 m grid, and whether a score combining the bands clears two standard deviations of roof-to-roof variation.

What it shows

Most mapped arrays are smaller than a single pixel: the histogram at the top of the page counts every array in Pakistan's calibration areas against the 100 m² of a 10 m pixel and the 400 m² of a 20 m one. On Peshawar West's concrete roofs, panels are darker than the roof in every band, but the visible bands barely register it, and combining all ten bands on the 20 m grid reveals more than the four sharper 10 m bands do.

That is the picture for one pixel. roofclf reads a whole footprint instead, averaging every pixel under it, and the limits that approach runs into are measured in the experiments register: roof-to-roof variation is twenty to forty times the sensor's noise (the spectral SNR budget), and most footprints resolve to a single pixel (most footprints are one pixel). Arrays a fifth of a pixel across, the norm in France, are where this stops working altogether.

Limits

Every number on the page is fitted and tested on Peshawar West's own pixels, so it describes those roofs, not Pakistan. The background each pixel is compared with is built from roofs the mapping shows to be PV-free, which a deployed model cannot know: panels cluster, so a real neighbourhood often carries them too. Read the thresholds as a best case for a single pixel.

Rebuilding it

pixi run python scripts/build_pvscope.py   # writes results/pakistan_pvscope.html
pixi run docs-figures                      # copies it to docs/assets/interactive/pvscope.html

--quadrat <stem> points it at another calibration area, and it finds the composite cell that covers the area unless --composite names one. It needs the composites, the VIDA buildings and the area's mapped solar under data/, so it runs locally, not in CI.