CCAiM – Cloud Classification AI Model

project links

CCAiM hugging face collection 🤗

Project Status

Currently in the dataset collection phase.

Project Goal

CCAiM aims to develop an AI-powered model for classifying clouds based on ground-level photographs. The project uses image recognition techniques to identify cloud types according to the WMO International Cloud Atlas classification.

Target Cloud Classes

The model will be trained to recognize the following cloud types:

1. Cirrus (Ci) – Thin, wispy clouds high in the sky
2. Cirrostratus (Cs) – Transparent, whitish veil clouds
3. Cirrocumulus (Cc) – Small, white patchy clouds
4. Altostratus (As) – Gray/blue layer clouds preceding storms
5. Altocumulus (Ac) – White/gray layered clouds with shading
6. Stratus (St) – Uniform gray cloud blanket
7. Stratocumulus (Sc) – Low lumpy clouds with blue sky gaps
8. Nimbostratus (Ns) – Dark precipitation clouds
9. Cumulus (Cu) – Fluffy white clouds with flat bases
10. Cumulonimbus (Cb) – Towering thunderstorm clouds

🛠 Planned Features

dreams

• Dataset expansion – Using contributor photographs
• Multiple model versions:
• V1: Initial first stable model (minimal viable dataset)
• Image classification API for integrations
• Interactive web demo
• Model evaluation tools

now can

• collect dataset
• create first model V0.0.1

License Information

• All data exist in hugging face
• Code in this repository is licensed under the MIT License (see LICENSE).
• Photographs located in folders named clouds_<dataset_number> are licensed under CC0 1.0 Universal (Public Domain), meaning they can be used freely for any purpose, including commercial use, without attribution.
• If other datasets are added in the future, their license terms will be specified in a separate license file inside their respective folder.

models

All models exist in hugging face

Two model lines

The project trains two separate model lines on the same data, split and metrics:

• scratch line (src/train.py, weights CCAiM_V0_0_X.pth) — the compact CCAiMModel CNN trained from scratch. This is the project baseline: it measures the contribution of dataset growth, so its results are only compared against earlier scratch versions.
• ResNet18 line (src/train_resnet.py, weights CCAiM_R18_V0_0_X.pth) — transfer learning from ImageNet-pretrained ResNet18 (frozen backbone first, then optional fine-tuning of the top block). This is the practical line for maximum accuracy and the future API / web demo.

⚠️ Note: the accuracy gain of the ResNet18 line comes from ImageNet pretraining — it is a one-time head start, not a result of dataset growth. Project progress is still measured by the scratch line and by dataset expansion.

Both lines use the same fixed random seed and train/val split (src/common.py), the same class-weighted loss, and report confusion matrix, per-class recall and macro-F1 on validation, so they are directly comparable.

Run the lines (from src/):

# scratch line
python3 train.py

# ResNet18 line (two phases: head, then fine-tuning of layer4)
python3 train_resnet.py
# head-only training, backbone stays frozen
python3 train_resnet.py --no-finetune

Inference works with both lines — the architecture is detected from the checkpoint / filename:

# scratch model (default, unchanged)
python3 predict.py path/to/image.jpg

# any specific model, e.g. the ResNet18 line
python3 predict.py path/to/image.jpg CCAiM_R18_V0_0_1.pth

Validation metrics (confusion matrix, per-class recall, macro-F1) for any saved model, on the same val split both lines train against:

python3 evaluate.py CCAiM_V0_0_5.pth
python3 evaluate.py CCAiM_R18_V0_0_5.pth

models list

• v_0.0.1 # first model learned by 23 photo.
• v_0.0.2 # second model learned by 42 photo
• v_0.0.3 # model learned by 88 photo
• v_0.0.4 # model learned by 165 photo

How to Contribute

• The most helpful contribution at the moment: If you find a discrepancy between the cloud class specified in the JSON label and the actual image content, correcting it will greatly improve dataset quality.