SAR Image Colorization Using Pix2Pix GAN

This repository contains the implementation of SAR Image Colorization using a Conditional GAN (Pix2Pix), developed as part of a research project at Lovely Professional University. The project aims to colorize grayscale Sentinel-1 SAR images using paired Sentinel-2 optical images.

Overview

SAR (Synthetic Aperture Radar) images are grayscale and captured regardless of weather conditions, making them ideal for remote sensing. However, their lack of color limits visual interpretation. This project leverages Pix2Pix GANs with U-Net-based generator and PatchGAN discriminator to colorize SAR images.

Dataset

• Dataset Name: SEN1-2
• Image Size: 256 x 256
• Total Images: 64,000 (50% SAR, 50% Optical)
• Classes: Agricultural, Urban, Barren, Grassland
• Paired Dataset: Each grayscale SAR image has a corresponding color optical image.

Methodology

Generator

• U-Net architecture with skip connections.
• Downsamples the input using convolution, upsamples using transposed convolution.

Discriminator

• PatchGAN: Operates on N x N patches instead of full images.
• Focuses on local features, improves fine detail generation.

GAN Type

• Pix2Pix: Conditional GAN for image-to-image translation.

Loss Functions

• L1 Loss: Pixel-wise absolute error.
• Perceptual Loss: Measures feature map difference using pretrained VGG-16.
• Adversarial Loss: Binary cross entropy loss to fool the discriminator.
• MSE Loss: Used in Denoising Autoencoder for additional refinement.

Training Setup

• Framework: PyTorch
• GPU: NVIDIA DGX A100 (200GB)
• Optimizer: Adam
  • Generator: LR = 0.0002, β1 = 0.005, β2 = 0.99
  • Autoencoder: LR = 0.001
• Input Shape (Generator): (1, 256, 256, 3)
• Input Shape (Discriminator): (1, 256, 256, 3)

---

Results

Model	SSIM	PSNR (dB)	MSE
Pix2Pix	0.159	11.32	-
cGAN + SSIM + L1	0.353	16.28	-
Cycle GAN	0.252	13.23	-
EPC-GAN	0.188	12.07	0.0047
Pix2Pix + Perceptual Loss	0.97	27.42	0.0021

🛠️ Installation

git clone https://github.com/yaseeng-md/SAR_Image_Colorization_Using_Deep_Learning_Algorithms.git
cd SAR_Image_Colorization_Using_Deep_Learning_Algorithms
pip install -r requirements.txt