SAR Image Colorization for Comprehensive Insight Using Deep Learning

This repository contains the implementation of a Deep Learning-based SAR Image Colorization system developed for Smart India Hackathon 2024. The solution uses a Hybrid Model combining Attention U-Net and Conditional GANs to transform synthetic aperture radar (SAR) grayscale images into high-quality colorized images. This project aligns with the Space Technology theme to enhance the interpretation of SAR imagery for applications like disaster response, environmental monitoring, and climate change analysis.

---

Key Features

• Hybrid Model: Combines the spatial feature extraction power of Attention U-Net with the realism of Conditional GANs.
• Advanced Image Processing: Overcomes noise issues, enhances fine-grained details, and generates realistic, visually intuitive outputs.
• High Evaluation Metrics: Achieved an average PSNR of 29.29 dB, SSIM of 0.83, and an average processing time of 1.44 seconds per image.
• Real-Time Feasibility: Designed for integration into satellite systems for real-time SAR image colorization.

---

Problem Statement

• Problem Statement ID: SIH1733
• Title: SAR Image Colorization for Comprehensive Insight using Deep Learning
• Theme: Space Technology
• Category: Software

---

Technical Approach

Model Architecture

1. Attention U-Net: Ensures precise spatial focus and better feature retention during colorization.
2. Conditional GANs: Enhances color realism by leveraging adversarial training techniques.

Challenges Solved

• Mitigated intrinsic speckle noise in SAR images.
• Preserved fine-grained spatial and contextual features.
• Improved color consistency and realism.

Workflow

1. Preprocessing: Image resizing, normalization, and noise mitigation.
2. Model Training: Custom loss functions integrating L1 loss, perceptual loss, and adversarial loss.
3. Post-Processing: Visual refinement and evaluation using PSNR and SSIM metrics.

---

Benefits and Impact

Key Advantages

• 30% Faster Data Interpretation: Accelerates insights into SAR data.
• Enhanced Disaster Response: Provides improved visual data for critical decision-making.
• 20% Higher Accuracy: Supports precise monitoring of environmental changes like deforestation and urbanization.

NISAR Mission Impact

• Supports NASA-ISRO's NISAR mission by reducing SAR data interpretation time by 30%.
• Enhances public engagement with visually compelling SAR data visualizations.

---

Results and Evaluation

• PSNR: 29.29 dB
• SSIM: 0.83
• Processing Time: 1.44 seconds per image

---

Future Work

• Integration with satellite systems for real-time colorization.
• Expansion to multi-spectral analysis for broader applications.

---

Repository Links

• GitHub Repository
• YouTube Demo

---

How to Run the Project

Prerequisites

• Python 3.8 or later
• TensorFlow 2.x
• Other dependencies: See requirements.txt (to be added).

Steps

1. Clone the repository:

   git clone https://github.com/hackoverflow72/Team_HackOverflow_1733

2. Install dependencies:

   pip install -r requirements.txt

3. Prepare the dataset:
   • Organize SAR grayscale images and corresponding optical images as specified in the create_dataset function.
4. Train the model:

   python train.py

5. Evaluate results:
   • Use pre-trained models provided or evaluate your trained models on test data.

---

Contributions

Team HackOverflow thanks Smart India Hackathon 2024 organizers for the opportunity to contribute to SAR image enhancement.

---

References

• Generating High-Quality Visible Images from SAR Images Using CNNs
• Colorizing Sentinel-1 SAR Images Using Variational Autoencoder
• SAR Image Colorization Using Multidomain CycleGAN

---

For more information, visit the project GitHub page and view the demo video.