pawel-83748/weather-image-classifier

Weather image classifier (convnext-tiny-224 fine-tuned)

Classifies an outdoor photo into one of 4 weather classes: Cloudy, Rain, Shine, Sunrise.

Model details

• Base model: facebook/convnext-tiny-224 (ImageNet pre-trained), fully fine-tuned.
• Training data: Kaggle "Multi-class Weather Dataset" (1125 images), stratified split 787 train / 169 validation / 169 test (seed 42).
• Training: 6 epochs, batch size 32, AdamW, cosine schedule, fp16, light augmentation (random resized crop, horizontal flip, color jitter). Best epoch chosen by validation accuracy.

Results (test set, 169 images)

• Accuracy: 0.9882
• Macro-F1: 0.9888

Class	Precision	Recall	F1	Support
Cloudy	1.000	1.000	1.000	45
Rain	1.000	1.000	1.000	33
Shine	0.950	1.000	0.974	38
Sunrise	1.000	0.962	0.981	53

Confusion matrix and per-class table are in the eval/ folder of this repo.

Usage

from transformers import pipeline
clf = pipeline("image-classification", model="<this-repo-id>")
clf("photo.jpg", top_k=3)

Limitations

• Small dataset (1125 images, 4 classes): not robust to night scenes, indoor photos, or unseen weather types.
• Classes are visually overlapping (e.g. Cloudy vs. Rain, Shine vs. Sunrise); expect confusion there.
• Educational project, not intended for safety-critical use.