solarscan-yolov8n-cls

SolarScan — solar panel condition classifier

YOLOv8n-cls fine-tuned to sort a photograph of a solar panel into one of six conditions, and map that to a maintenance action. Trained for SolarScan, a graduation project that turns a smartphone into a panel inspector.

Ships as both PyTorch (best.pt, 3 MB) and ONNX (best.onnx, 5.5 MB). The ONNX export was verified against the PyTorch model under identical preprocessing — maximum per-class probability difference 3.6e-07.

Classes and actions

Class indices are alphabetical, in the order above.

In the deployed service, a top-1 confidence below 0.5 is reported as Unknown rather than a class — without that floor an image that isn't a solar panel gets forced into one of the six.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
r = model("panel.jpg")[0]
print(r.names[r.probs.top1], float(r.probs.top1conf))

ONNX, with the preprocessing the model was trained under — resize the shortest edge to 224 (bilinear, antialiased), centre-crop 224×224, scale to 0..1, NCHW:

import numpy as np, onnxruntime as ort
from PIL import Image

NAMES = ["Bird-drop", "Clean", "Dusty", "Electrical-damage",
         "Physical-Damage", "Snow-Covered"]

im = Image.open("panel.jpg").convert("RGB")
w, h = im.size
s = 224 / min(w, h)
im = im.resize((round(w * s), round(h * s)), Image.BILINEAR)
w, h = im.size
im = im.crop(((w - 224) // 2, (h - 224) // 2, (w - 224) // 2 + 224, (h - 224) // 2 + 224))
x = (np.asarray(im, np.float32) / 255).transpose(2, 0, 1)[None]

sess = ort.InferenceSession("best.onnx")
p = sess.run(None, {sess.get_inputs()[0].name: x})[0][0]
print(NAMES[int(p.argmax())], float(p.max()))

Training

Per-epoch metrics are in results.csv.

Results

Validation split, 204 images, epoch 10:

The confusion matrix (confusion_matrix.png) has a diagonal of 187/204, which is exactly that 91.67%. The main confusion is Clean ↔ Dusty.

Read this before quoting the number

0.9167 is an upper bound, not field accuracy. Two measured reasons:

• The train/validation split has 35.8% overlap — some validation images also appear in training. Bird-drop is the clearest tell: 94% of its validation set is leaked, and its recall is correspondingly near-perfect.
• There is no held-out test split. Every figure here is validation-set.

The documented next step is an MD5-deduplicated split and a retrain. Two classes also have very small validation sets — Physical-Damage (14) and Electrical-damage (21) — so their per-class figures carry wide error bars.

Benchmarked against

Three architectures were trained and compared before one was chosen:

The figures are not directly comparable — different tasks and class counts. YOLOv8n-cls was selected on real-world smartphone performance and deployment simplicity, not on the highest validation number.

Training data

A public solar panel image dataset — 930 images across the six classes, split 726 train / 204 validation, with a 3.65:1 class imbalance (Bird-drop 252 … Physical-Damage 69).

The dataset is not my work and is not redistributed here. The six images in examples/ come from its validation split and are included only to demonstrate inference; credit belongs to the original authors.

Licence

The weights are released for portfolio and research use. No licence is asserted over the training data, which is third-party — check the original dataset's terms before any commercial use.

Deployment

In the full project the model is served behind FastAPI, consumed by a Flutter app with a PHP/MySQL backend, returning class, confidence and recommendation, with an analytics dashboard, anomaly alerts and CSV export.

A live in-browser demo of this model, running via ONNX Runtime Web, is at https://saifeleslamelgalaly.github.io/.

---

Trained by Saif Elgalaly.