adityashirsatrao007/solar-panel-dust-xai

Solar Panel Dust Detection with Explainable AI

Hybrid EfficientNet-B2 (frozen) + RBF-SVM model for classifying a single RGB image of a photovoltaic (PV) panel as clean or dirty, with a focus on cross-dataset generalisation and explainability (Grad-CAM, Score-CAM, Integrated Gradients, SHAP, LIME).

Model

• Feature extractor: EfficientNet-B2, ImageNet weights, frozen, global-average pooled → 1,408-d vector.
• Head: RBF-kernel SVM (C=10, gamma=auto), class_weight="balanced".
• Architecture is exactly the deployed production model (version v007).

Metric	Value
Test accuracy (merged held-out, 381 imgs)	90.29%
AUC-ROC	0.9596
5-fold CV accuracy	86.38%
External: Dusty/Clean (2,562 imgs)	98.24% (98.0% dirty recall)
External: Faulty-panel clean/dirty (383 imgs)	99.74% (100% dirty recall)

Files

• svm_classifier.pkl — trained RBF-SVM head (joblib).
• scaler.pkl — fitted StandardScaler applied to pooled features.
• pipeline_meta.json — full training metadata and per-metric scores.
• class_names.json — label order (["clean", "dirty"]).
• demo.py — self-contained inference script.

Usage

pip install tensorflow scikit-learn joblib numpy pillow

python demo.py panel.jpg            # print label, confidence, dustiness
python demo.py clean.jpg dirty.jpg  # batch

The first run downloads the EfficientNet-B2 ImageNet weights (~80 MB) from TensorFlow via keras.

How it was trained

A single-source model overfits one acquisition pipeline and collapses on new data (observed: ~6% dirty recall on external sets). The fix was to merge three public PV-dust sources into one 3,787-image corpus and retrain:

Split	Clean	Dirty	Total
Train	1,750	1,279	3,029
Val	218	159	377
Test	220	161	381
All	2,188	1,599	3,787

The frozen representation + margin-based SVM then transfers far better across acquisition sources (see external-validation metrics above).

Intended use & limitations

• For single-panel RGB inspection; labels are binary (clean/dirty), not severity.
• Confidence for every prediction is produced; a configurable confidence threshold can route low-confidence samples to human review.
• The backbone is loaded fresh from ImageNet weights at inference (not stored in this repo), matching the deployed frozen-backbone protocol.

Related

Full experimental details, the IEEE-format paper, and the five-method XAI module live in the companion project repo.