A clearer pilot near Rengo: El Mirlo#

Open the annotated gallery and offline download.

Continue with the 3D scene and reusable method, which lifts the same candidates over real terrain using explicitly assumed table mounting planes.

El Mirlo (CEN ID 2253) is approximately 5.1 km from the rounded Rengo centre (-70.858, -34.405), in Malloa. Six nearby operating registry entries were screened: Rengo Solar, Lo Miguel, Rinconada, Pueblo Hundido, El Mirlo and Rosario. El Mirlo was selected for visible individual frame lines in both directions. The developer’s project page confirms its identity and location. Its project power and the current CEN maximum-power field describe different quantities; neither is converted to a module count.

Result and scope#

The new counter finds 2,055 image candidates in a southern inspection block, across 22 resolved table components. The original counter with its defaults returns None (unresolved) on the same image and scope. Neither is a complete plant inventory. Rejected cells and tables remain in the exports.

The small same-image visual audit has 72 reference centres in three deliberately clear strips (four scoring rectangles), plus two negative regions. A one-to-one match within two delivered pixels gives:

Method

Matched

Missed

Extra detections

Local recall

Original defaults

0

72

0

0%

Orientation and lattice

51

21

0

70.8%

The references were interpreted by a single AI assistant, with approximate centre clicks and row guides. They are not independent human annotations, field ground truth, a random sample or held-out validation. The first comparison exposed seam-versus-centre misclicks in two regions; these were corrected by re-inspecting unannotated pixel enlargements. The initial and corrected files are both retained. Parameter development used the same image. Consequently, zero extra detections in these small regions is not a claim of 100% field precision. The remaining candidates have not been individually validated.

This establishes useful candidate extraction and a review workflow. Improving recall and establishing inventory accuracy require independent labels over complete tables, shadowed regions and additional sites, ideally with an as-built schedule or a sufficiently detailed UAV orthophoto.

What changed in the method#

energy_methods.pv_lattice_counting.count_lattice_pv_modules adds:

  1. Normalised colour screening without a minimum absolute brightness gate.

  2. Principal-axis estimation and per-component image rectification, allowing arbitrary table orientation rather than only nearly horizontal tables.

  3. Long-axis autocorrelation to check the repeating frame pitch independently of the observed brightness peaks.

  4. Two-row support: three visible transverse frame lines are sufficient.

  5. Per-cell spacing, original-pixel colour, local frame contrast, scope and source-resolution checks. Missing frame lines and irregular gaps are excluded; the algorithm does not fill a lattice from known dimensions or module totals.

This is a deterministic experimental method using NumPy/SciPy, not a pretrained SAM/YOLO model or an implementation of a published detector. The geometric module-detection literature, including Cardoso et al. (2024), provides background. That paper uses high-resolution UAV/3D data; its reported accuracy does not apply to this basemap pilot. Repeated cell patterns can still be confused with module frames, which is why the raw-pixel views and independent validation limits matter. Synthetic tests establish software behaviour under rotation, noise and missing modules, not real-site accuracy.

Imagery and provenance#

The Esri World Imagery source probe at (-70.8615, -34.4515) reports WV03, 2024-01-14, Vivid Advanced / Vantor, 0.31 m source resolution and 0.30 m service sampling. This is a point metadata query, not a guarantee of uniform acquisition across the image. The captured detail is 900 × 600 pixels, bounds (-70.8627, -34.4523, -70.8603, -34.4507), EPSG:4326.

Delivered sampling is approximately 0.245 × 0.296 m. This oversampling adds no optical detail. A conservative source/delivered ratio of 1.265 is applied to the three-source-pixel minimum frame pitch. No image sharpening or synthetic imagery is used. Exact export/query requests, retrieval times, SHA-256 checksums, parameters, visual annotations and the inspection polygon are committed.

The inspection polygon is manually selected, not a surveyed plant boundary. GeoJSON positions use the rendered image’s georeference and half-pixel offset; they are not surveyed module positions. Imagery remains © Esri and its licensors; see the fixture README for attribution and the static-map terms.

Reproduce and inspect#

uv sync --locked --all-packages
uv run --no-sync python examples/analysis/plot_rengo_module_counting.py
from energy_analysis.module_counting.rengo import export_rengo_study

export_rengo_study(
    "examples/analysis/data/rengo_modules",
    "rengo-audit",
)

The gallery download includes the script, exact local inputs, both annotation versions, the guide and computed CSV/GeoJSON/JSON exports. It runs offline using the repository’s installed packages, on Python 3.12 or 3.13. The gallery also reports the effect of frame-prominence values 0.012, 0.015 and 0.018 without selecting whichever count looks most desirable.