How omid-ghozatlou/hybrid-gan-and-spectral-angular-distance-for-cloud-removal works
Project type and purpose
• This is a machine-learning research pipeline focused on cloud removal in multispectral satellite images. It implements a hybrid Generative Adversarial Network (GAN) augmented with a Spectral Angular Distance (SAD) metric, plus pre- and post-processing steps (band stacking and polar coordinate transforms).
File structure
Top-level scripts and directories:
• Stacked12Bands.py – stacks 12 multispectral bands into composite images.
• images_2_Polar.m – MATLAB script converting stacked images into polar coordinates.
• PolarCoordinates_method_modified.py – Python-side polar transform and coordinate handling.
• stacked_tetas.py – assembles the polar-transformed bands back into training-ready patches.
• data_loader.py – data-loading utilities for Keras/TensorFlow.
• cyclegan_cloud.py – training driver that implements the hybrid GAN + SAD loss.
• Predict.py – inference script applying a trained model to test images.
• requirements.txt – lists Python dependencies (TensorFlow/Keras, NumPy, etc.).
README insights
• Workflow:
1. Stack raw multispectral data (Stacked12Bands.py)
2. Polar coordinate transform (MATLAB + Python scripts)
3. Assemble final training patches (stacked_tetas.py)
4. Train hybrid CycleGAN with SAD (cyclegan_cloud.py) via data_loader
5. Run inference on test data (Predict.py)
• Data organization: expects folders trainA, trainB (cloudy/clean), testA, testB.