☁️ CloudClassification

A Python deep learning project that classifies cloud images into the 11 standard meteorological cloud genera using a custom Xception-inspired CNN built with TensorFlow/Keras. Includes a purpose-built web scraper for dataset collection, a dataset merging utility, and separate CLI scripts for model creation, training, and single-image inference.

---

Cloud Classes

The model classifies images into all 11 WMO cloud genera:

Abbreviation	Cloud Type
Ac	Altocumulus
As	Altostratus
Cb	Cumulonimbus
Cc	Cirrocumulus
Ci	Cirrus
Cs	Cirrostratus
Ct	Contrail
Cu	Cumulus
Ns	Nimbostratus
Sc	Stratocumulus
St	Stratus

---

Dataset Pipeline

The dataset was built from scratch using a two-step scraping and merging pipeline:

1. imageScraper/imageScraper.py — Web scraper

Crawls a German meteorological photo gallery website, automatically categorising and downloading cloud images:

• Traverses all gallery pages discovered via a <select> dropdown on the index page
• For each image entry, fetches the detail page and extracts the cloud type by parsing the <b> tag with a regex
• Creates a local folder per cloud type and downloads all .jpg images into it

python imageScraper.py <gallery-url>

2. imageScraper/copy.py — Dataset merger

After scraping, cloud type folder names may contain comma-separated multi-label categories (e.g. "Cu, Cb"). copy.py resolves these by checking whether all labels share the same 2-letter prefix, and if so copies the images into the corresponding class folder in mergedDataset/. Single-label folders are copied directly. The final merged dataset contains 2,543 images across 11 classes.

---

Model Architecture

A custom Xception-style CNN built from scratch with Keras:

• Input — 180×180 RGB images
• Data augmentation — Random horizontal flip and random rotation applied during training
• Entry block — Two Conv2D layers (32 and 64 filters) with batch normalisation and ReLU
• Residual blocks — Four blocks with depthwise separable convolutions at sizes 128 → 256 → 512 → 728, each with MaxPooling and a projected residual connection
• Top block — SeparableConv2D (1024 filters), GlobalAveragePooling, Dropout (0.5)
• Output — Softmax over N classes (or sigmoid for binary); trained with categorical crossentropy and the Adam optimiser
• Checkpointing — Model saved after each epoch via ModelCheckpoint

---

Project Structure

CloudClassification/
├── createModel.py              # Build and save a fresh untrained model
├── trainModel.py               # Load and train a model on a dataset
├── classifyImage.py            # Run inference on a single image
├── image_classification_01.py  # Original all-in-one prototype script
└── imageScraper/
    ├── imageScraper.py         # Web scraper for cloud image collection
    └── copy.py                 # Merge scraped folders into clean dataset

---

Usage

1. Scrape dataset

cd imageScraper
python imageScraper.py <gallery-url>
python copy.py

2. Create a model

Builds and saves an untrained model whose output size matches the number of class folders in the dataset:

python createModel.py <modelfile.h5> <datasetPath>

3. Train the model

Loads an existing model and trains it for a given number of epochs, saving a checkpoint after each:

python trainModel.py <model.h5> <datasetPath> <epochs>

The dataset directory should contain one subfolder per class (e.g. Ac/, Cu/, Cb/, etc.) with JPEG images inside.

4. Classify an image

Runs inference on a single image, printing the confidence score for each class and the winner:

python classifyImage.py <model.h5> <image.jpg>

Example output:

Ac   2.31
As   1.07
Cb   0.84
Cu  91.45
...
Winner: Cu with 91.45

---

Requirements

pip install tensorflow numpy matplotlib requests beautifulsoup4

---

License

This project is licensed under the MIT License.