PROBA-V Super-Resolution

About

This is my quick attempt at the PROBA-V Super Resolution Competition. Competition website: https://kelvins.esa.int/proba-v-super-resolution/.

“In this competition you are given multiple images of each of 74 Earth locations and you are asked to develop an algorithm to fuse them together into a single one. The result will be a "super-resolved" image that is checked against a high resolution image taken from the same satellite, PROBA-V.”

Custom Architecture

I developed a custom deep learning architecture specifically for this task. See report for details.

Usage

Pre-Trained Model & Results

Notebook.ipynb is the main file containing training and results.

Report.pdf is the project report describing problem analysis, approach and results.

model.h5 is the fully trained model.

submission.zip contains the results for submission (inference on the test set).

Train on Local Machine

1. Make sure you have conda installed
2. Clone this repo

git clone https://www.github.com/rizandigp/PROBA-V-Super-Resolution
cd PROBA-V-Super-Resolution

3. Download the data

wget -P probav_data https://kelvins.esa.int/media/competitions/proba-v-super-resolution/probav_data.zip
unzip -q probav_data/probav_data.zip -d probav_data

4. Prepare environment

# Set up conda environment
conda env create -f environment.yml
conda activate probav

# Get dependencies
pip install git+https://www.github.com/keras-team/keras-contrib
git clone https://github.com/lfsimoes/probav
git clone https://github.com/rizandigp/keras_superconvergence

5. Run Notebook.ipynb

Train on Google Colab

1. Upload Notebook_Colab.ipynb, dataset.py, model.py and training.py to Colab
2. Run the notebook

Results

Results on Validation Set

Results on Test Set