Classifying weather conditions from photos using efficientnet-b0 and grad-cam visualization.
dataset: multi-class weather dataset — 1125 images across 4 classes (cloudy, rain, shine, sunrise)
| metric | value |
|---|---|
| validation accuracy | 91% |
| model | efficientnet-b0 |
| training epochs | 18 (10 head + 8 finetune) |
precision recall f1-score
Cloudy 0.88 0.82 0.85
Rain 0.97 0.97 0.97
Shine 0.85 0.85 0.85
Sunrise 0.94 0.99 0.96
accuracy 0.91
two-phase transfer learning:
- freeze all efficientnet weights, train only the new classifier head (lr=1e-3, 10 epochs)
- unfreeze the last two feature blocks and fine-tune everything together (lr=1e-5, 8 epochs)
this approach converges faster and avoids destroying pretrained features early on.
- checked all 1125 images for corruption with
PIL.Image.verify()— 0 bad files found - 80/20 train/val split
- augmentation on train only: horizontal flip, ±15° rotation, color jitter (brightness + contrast)
- normalized with imagenet mean/std since we're using a pretrained backbone
Grad-CAM lets us see which regions of the image actually drive the model's prediction. two examples worth highlighting:
sunrise — the model locks onto the bright horizon glow and warm color gradient near the sun. it's not just detecting orange pixels; it focuses specifically on the light diffusion pattern at the edge of the sky, which is what separates sunrise from shine.
cloudy — attention spreads across the texture of the cloud cover rather than any single spot. the model picks up on flat, diffuse light and the absence of shadows — exactly what distinguishes overcast skies from the other classes.
this tells us the model learned genuinely meaningful visual features rather than shortcut patterns or background noise.
