andre-kotze/satellite-superresolution

Upscaling satellite imagery using GAN by 4 times

Satellite SuperResolution using GAN / 4x Improvement

Model Training and Architecture

This a slightly modified version of SRGAN model as it does not have the BatchNormalization layer as mentioned in the original paper. The training happens in two stages. The data preprocessing and metrics have been modified to support satellite reflectance data values (0-1) instead of traditional 0-255 range.

The model supports only 2x, 3x and 4x upscaling of TIF Images however it uses feature maps of VGG network in the final stage

Results

AGRI

URBAN

1. PSNR Training

First very deep ResNet architecture using the concept of GANs to form a perceptual loss function as measured by PSNR and structural similarity (SSIM) with our 16 blocks deep ResNet (SRResNet) optimized for MSE.

2. GAN Training

Traditional SRGAN model is used with L1 Loss for training. The backend used is VGG-19

Data

Sample data is available here. Standard TIF imagery (reflectance) is used for model building. The zip file should be extracted into data/ directory

Installation

Separate conda environment with Python >= 3.7 and tensorflow-gpu is recommended. python setup.py install

Running Training

1. Configure File: options\train\SRResNet_SRGAN\train_MSRResNet_x4.yml
2. Configure File: options\train\SRResNet_SRGAN\train_MSRGAN_x4.yml

Running Inference

1. Configure File: options\test\SRResNet_SRGAN\test_MSRGAN_x4.yml

Starting the process

1. Running the SRRSNET Model python basicsr/train.py -opt options/train/SRResNet_SRGAN/train_MSRResNet_x4.yml
2. Running the SRGAN Model python basicsr/train.py --options options/train/SRResNet_SRGAN/train_MSRGAN_x4.yml
3. Inference from SRGAN Model python basicsr/test.py --options options/test/SRResNet_SRGAN/test_MSRGAN_x4.yml

❤️ References

Inspired from https://github.com/xinntao/BasicSR