xmpuspus/solar-map-ph-clf-v4

SolarMap.PH classifier (clf_v4)

Frozen openai/clip-vit-large-patch14 image encoder plus a scikit-learn LogisticRegression head, Platt-calibrated on a held-out 20% source-disjoint split. Trained to detect rooftop solar PV installations from 600x600 px Esri World Imagery tiles over Greater Metro Manila, Philippines.

Reproducibility hash: sha256:56900722a8427be404883d5f5bee8850f1f5ed837e150df4c4b1cc991d062609 (prefix 56900722a8427be4).

The classifier is bit-exact reproducible from the committed dataset_v4.npz embeddings cache in the project repo. Verify locally with:

python3 scripts/verify_clf.py detection/train/clf_v4.joblib
# or
make hash-verify

Intended use

Public-interest research on the gap between informal rooftop solar and the formal net-metering registry in the Philippines. The model is calibrated on Esri World Imagery over Greater Metro Manila; out-of-domain performance is not characterized.

Appropriate use cases:

• Civic-tech research, journalism, and public-policy reporting on rooftop solar adoption.
• Academic encoder-ablation comparisons.
• Region extensions to other Philippine geographies (Cebu, Davao, Iloilo, Cagayan de Oro) with explicit re-validation.

Out-of-scope:

• Per-household enforcement, permit verification, tax records, or any individual-identifying decision.
• Engineering or financial advice for specific installations.
• Adversarial or off-distribution imagery (different satellite source, drone, or composite imagery).

Model architecture

• Encoder: openai/clip-vit-large-patch14, frozen, image projection head, 768-dim embedding.
• Classifier head: sklearn.linear_model.LogisticRegression, class-balanced weights, C=1.0, 2000 max iterations.
• Calibration: Platt sigmoid P = sigmoid(1.2916 * decision_function(x) + 0.1627) fit on the held-out 20% source-disjoint subset.
• Production threshold: 0.85.

Training data

• 294 OSM-tagged positives (generator:source=solar AND location=roof) within the Meralco franchise bounding box, deduplicated to unique buildings.
• 4 case-study positives, 4 promoted random-negative positives surfaced during active learning, 111 v3-confident positives (active-learning round 2).
• 197 negatives: 46 hand-labeled not_solar tiles + 150 random tiles drawn from the NCR bounding box + 1 v3 false positive.
• 2 noisy cases dropped.
• 4x augmentation (rotation, flip, color jitter) per source.
• Group-aware 5-fold cross-validation.

Evaluation

Metric	Value	Notes
F1 at t=0.85 (calibrated holdout)	0.870	TP=47, FP=2, FN=12, TN=37
Precision at t=0.85	0.959
Recall at t=0.85	0.797
F1 at t=0.85 (5-fold group-aware CV)	0.867	TP=271, FP=27, FN=29, TN=173

Encoder ablation on the same head and same Platt holdout:

• openai/clip-vit-large-patch14: F1 = 0.870, P = 0.959
• facebook/dinov2-large: F1 = 0.830 (-4 pt), P = 0.936
• satlas Aerial_SwinB_SI: F1 = 0.727 (-14 pt), P = 0.900

Limitations and biases

• Imagery vintage. Esri PH imagery is typically 1-3 years stale. Recent installs (2024-2026) may be missed.
• Domain shift. Calibrated on Esri World Imagery over Metro Manila. Performance on other satellite sources or other regions is not characterized.
• False positives. Highly reflective non-PV surfaces (skylights, polished metal, water tanks, blue painted metal roofs) score above 0.85 occasionally. Calibration precision is 95.9% on holdout; expect 1 in 25 high-confidence detections to be incorrect.
• Mount-type confusion. The model detects panels in imagery regardless of mount type. The largest detection in the production scan (2370 kWp) turned out to be a ground-mount utility-scale solar farm rather than a rooftop install (see docs/screenshots/qa-2026-05/spot-check/findings.md). Mount-type classification is a queued post-launch feature.
• SAM over-segmentation. Per-roof panel polygons can be fragmented across multiple OSM building footprints, especially on large contiguous installations. The aggregate kWp is conservatively correct; per-building kWp values for the smallest entries (<10 kWp on small OSM ways) may under-allocate the visible array.
• Class imbalance at deployment. Even with 95.9% calibrated precision, scanning 16,544 tiles at low base-rate (about 1-2% prevalence) produces a non-trivial false-positive count. The candidate tier (0.70-0.85) is a review surface, not a verified inventory.

How to use

import joblib
import torch
from PIL import Image
from transformers import CLIPProcessor, CLIPModel

# 1. Verify the hash before loading. joblib.load executes pickle.
import hashlib
EXPECTED = "56900722a8427be4"
with open("clf_v4.joblib", "rb") as f:
    digest = hashlib.sha256(f.read()).hexdigest()
assert digest.startswith(EXPECTED), f"hash mismatch: {digest[:16]}"

# 2. Load CLIP encoder.
device = "cuda" if torch.cuda.is_available() else "cpu"
proc = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
clip = CLIPModel.from_pretrained("openai/clip-vit-large-patch14").to(device).eval()

# 3. Load classifier + calibration.
clf = joblib.load("clf_v4.joblib")
calib = joblib.load("clf_v4_calibrated.joblib")  # contains Platt A/B

def score(img_path):
    img = Image.open(img_path).convert("RGB")
    inputs = proc(images=img, return_tensors="pt").to(device)
    with torch.no_grad():
        feats = clip.get_image_features(**inputs).cpu().numpy()
    raw = clf.decision_function(feats)[0]
    return 1 / (1 + 2.718281828 ** -(calib["platt_A"] * raw + calib["platt_B"]))

Citation

SolarMap.PH (2026-Q2). Open-source rooftop solar detection from public satellite imagery. https://github.com/xmpuspus/solar-map-ph. CC-BY-4.0.

See CITATION.cff in the project repo for the canonical citation format.

License

• Model weights: MIT (this Hugging Face artifact)
• Dataset embeddings (dataset_v4.npz in repo): CC-BY-4.0
• Code in the project repo: MIT
• Per-building GeoJSON published with the project: CC-BY-4.0

Contact

Author and self-designated Data Protection Officer: Xavier Puspus, xpuspus@gmail.com. Open issues at https://github.com/xmpuspus/solar-map-ph/issues.

Related work

See docs/research/related-work.md in the project repo. Closest sibling project in the Philippine rooftop-solar mapping space is ICSC's SPECTRUM (July 2025): nationwide coverage, complementary scope. DeepSolar (Stanford, 2018) is the methodological precedent.