Use calibration quadrats as training data, not just post-hoc correction

SHIPPED (as of 2026-08-11)

This became roofclf, now half the main workflow. The feature design proposed here was not what shipped: temporal features, spike counts and NDVI trajectory were each measured and rejected, and the deployed feature set is footprint size plus reflectance. The stratum-holdout acceptance criterion was replaced by leave-one-quadrat-out over 23 quadrats.

Labels: enhancement, model, calibration

Problem

The calibration quadrats (docs/calibration-mapping-protocol.md: ~25–35 exhaustively mapped 1–4 km² areas across 6 Pakistani landscape strata) are currently planned for one purpose only: post-hoc correction of the density model's output per stratum. That under-uses the most expensive data we will have. Exhaustive mapping means every quadrat building carries a verified has-PV / has-no-PV label -- precisely the supervision the model lacks in its measured failure regime (<500 m² installations: 6–16% detection).

Proposal

Train a per-building PV classifier on the quadrat labels, as a second model alongside (not replacing) the segmentation model:

1. Unit of prediction: the building footprint, not the pixel. For sub-pixel roofs, "does this building have PV?" is a far easier learning problem than tracing panel outlines at 10 m.
2. Features: temporal, not single-composite. Per-building Sentinel-2 time-series statistics over ~2 years -- the same per-scene machinery the glint method uses (src/earthpv/glint.py) -- e.g. spike counts, reflectance percentiles and their stability, NDVI trajectory, plus footprint area and stratum covariates. External evidence this direction pays: multi-revisit fusion beat single-composite segmentation by 5–17% IoU on the analogous substation task (arXiv:2409.17363), and Google reached 79% building mIoU from Sentinel-2 by distilling high-res-derived labels (arXiv:2310.11622) -- the quadrats play exactly that teacher role here.
3. Output: per-building PV probability → calibrated adoption rate per cell/stratum, complementing pv_area_det/pv_area_exp from segmentation (which remain authoritative for ≥1k m² arrays).

Design constraints

• Spatial holdout is mandatory: train on some quadrats, validate on held-out quadrats from the same stratum in a different province -- same-quadrat splits will overfit local roof styles and report fantasy accuracy (lesson learned the hard way in the substation project: val metrics decoupled from field quality).
• Class balance: quadrats supply verified negatives (every unlabeled building in a completed quadrat is a true negative -- that's what the completeness rule in the mapping protocol buys us).
• Date alignment: building labels are only valid for the imagery window they were mapped against; the protocol's imagery-date register field exists for this.

Synergy

Same field effort funds three things: density calibration (original purpose), this classifier's training set, and the known-negative sample the glint spike-rate estimator needs (see companion issue glint-spike-rate-density-estimator.md). The two proposed estimators share the per-building time-series pull -- build that data layer once.

Acceptance

• Per-building classifier trained on ≥4 quadrats/stratum, evaluated on ≥1 held-out quadrat/stratum in a different province, reporting per-stratum precision/recall at the building level.
• Demonstrated improvement over the current centroid-pixel fallback for the <500 m² class on the held-out quadrats.

---

🤖 Drafted with Claude Code