Cloud and Shadow Segmentation using Deep Learning

This project implements a two-stage deep learning pipeline to segment clouds and shadows from satellite imagery using TOA reflectance and metadata. The models are trained on LISS-IV data.

---

📂 Project Structure

├── inference_code.ipynb        # Main inference pipeline
├── Requirements.txt         # Dependencies
├── Training_Files/              # Labelling of cloud mask and training scripts for ResUNet and U-Net
├── masks/                  # Predicted masks on test dataset
├── Report.pdf             # Methodology and analysis
├── NRCC251050_Inference_Training_Results.csv  # Training log (losses, metrics)
├── Earth_Sun_Distance.xlsx             # Earth-Sun distance values for TOA reflectance computation               └── Training_labeled_data/         # Dataset used to train models

---

🧠 Models Used

Task	Model	Input Channels	Output Classes
Cloud Segmentation	ResUNet (ResNet34 encoder)	TOA Bands (B2, B3, B4)	2 (Clear, Cloud)
Shadow Segmentation	Custom U-Net	TOA Bands + Cloud Mask + Sun Elevation	2 (Clear, Shadow)

📥 Download Trained Models

You can download the trained models (ResUNet for cloud segmentation and U-Net for shadow segmentation) from the link below:

🔗 Download Trained Models from Google Drive

---

🛠️ Requirements

Create a new python environment(python=3.10). Navigate to the downloaded folder where requirements.txt is present.

Install dependencies using:

pip install -r requirements.txt

🚀 Inference Guide

1. Place your test folder containing B2, B3, B4 bands and metadata file.
2. Update paths inside inference_code.ipynb:
   • Band paths (e.g., B2.tif, B3.tif, B4.tif)
   • Metadata text file path
   • Cloud model path (.pth)
   • Shadow model path (.pth)
3. Run all cells in the notebook to get:
   • Georeferenced mask.tif with values:
     0 = clear, 1 = cloud, 2 = shadow
   • Optional: cloud and shadow shapefiles