PV ATLAS

Research Paper

The effects of spatial resolution on rooftop photovoltaic detection in satellite imagery

Husnain KhurshidNational University of Sciences and Technology, Islamabad

Preprint · in preparation

Download PDF

Abstract

Rooftop photovoltaic detection is reported at ground sample distances spanning two orders of magnitude, but no study isolates resolution as a variable: comparisons change the sensor, the scene and the geography together, so no number from one paper is commensurable with another. We hold the scene fixed and vary only resolution, degrading 0.1 m aerial orthophotos of 645 rooftops through a point-spread-function sensor model to 0.3, 0.8, 2.5 and 10 m, and evaluating four detector families at every rung under identical spatially blocked cross-validation. We then acquire real Sentinel-2 over the same rooftops to test whether the simulation is faithful.

Four results follow. First, and most consequential for the field, random cross-validation inflates chip-level R² by a factor of 22 on this data (0.699 ± 0.047 random, 0.032 ± 0.440 spatial) because overlapping chips place the same rooftop in training and test; the evaluation protocol dominates every other design choice we measure. Second, 97.5% of arrays are smaller than one Sentinel-2 pixel, and three-band imagery at 10 m has no predictive power for panel fraction whether the three bands are simulated or real (R² ≈ 0 in both cases). Third, spectral breadth substitutes for spatial resolution at 10 m: moving from three bands to twelve buys +0.275 R², more than any change in resolution we measure. Fourth, spatial context becomes more valuable as resolution coarsens, not less — a small convolutional network matches per-pixel colour classifiers above 0.8 m but exceeds them by 0.05–0.06 F1 at 2.5 and 10 m.

We report detection floors as curves against array area rather than as a single threshold, alongside a matched chance baseline that the conventional per-array detection criterion fails to beat.

What We Found

01

Random cross-validation inflates R² by 22×

0.699 under random folds, 0.032 under spatial folds — same data, same model. Overlapping chips put the same rooftop in train and test. The evaluation protocol dominates every other design choice measured here.

02

At 10 m, band count beats spatial resolution

Going from 3 bands to 12 buys +0.275 R². No change in spatial resolution we measured buys as much.

03

97.5% of rooftop arrays are sub-pixel at 10 m

Median array 14.9 m²; 69% fall below 25 m² — the size of a 5 kW domestic system.

04

Three-band imagery at 10 m has no predictive power

R² ≈ 0 whether the three bands are simulated drone RGB or real Sentinel-2 visible.

Figures

Array size distribution, and the same arrays measured in pixels at each ground sample distance. At 10 m, 97.5% fall below one pixel.
Figure 1. Array size distribution, and the same arrays measured in pixels at each ground sample distance. At 10 m, 97.5% fall below one pixel.
The resolution ladder: 0.1 m orthophotos degraded through a point-spread-function sensor model to 0.3, 0.8, 2.5 and 10 m.
Figure 2. The resolution ladder: 0.1 m orthophotos degraded through a point-spread-function sensor model to 0.3, 0.8, 2.5 and 10 m.
Simulated 10 m against real Sentinel-2 over the same rooftops — a test of whether the sensor model is faithful.
Figure 3. Simulated 10 m against real Sentinel-2 over the same rooftops — a test of whether the sensor model is faithful.
F1 against resolution by detector family. Spatial context becomes more valuable as resolution coarsens, not less.
Figure 4. F1 against resolution by detector family. Spatial context becomes more valuable as resolution coarsens, not less.
Detection floors reported as curves against array area, with a matched chance baseline.
Figure 5. Detection floors reported as curves against array area, with a matched chance baseline.

Data

  • Jiang et al. (ESSD 2021), PV01

    645 chips of 0.1 m UAV orthophoto over Jiangsu, China; three roof types; 5,238 annotated arrays. CC-BY-4.0, via Zenodo.

  • Sentinel-2 L2A

    Real acquisitions over the same rooftops, via Element84 Earth Search over the public sentinel-cogs bucket — no credentials required.

Limitations

Research prototype. These findings rest on one dataset over a single 7×17 km area on a single date; the direction should generalise, the magnitudes should not be quoted as universal. The method axis currently covers per-pixel colour baselines and a small CNN — no promptable-segmentation rows yet.

Cite this work

@article{khurshid2026resolution,
  title  = {The effects of spatial resolution on rooftop photovoltaic
            detection in satellite imagery},
  author = {Khurshid, Husnain},
  year   = {2026},
  note   = {Preprint}
}