UnIT competition 2025

This solution won first prize at the UnIT 2025 competition in Prague.

The task was to develop a method that, given 5 satellite images of the same area, spaced in 5-day intervals, would predict a satellite image 25 days after the last one. Our final model was a custom Convolutional LSTM neural network.

The model has an input of five $128 \times 128$ images, each with 4 channels: red, green, blue, and near infrared (NIR). The prediction is a $128 \times 128$ image with one channel corresponding to the Enhanced Vegetation Index (EVI).

The model is accompanied by a website visualizes the input and prediction of the model.

Cloud removal

Clouds in satellite images posed a challenge, as they confused the model and reduced the usable data. Due to the unclean dataset, some images contained clouds. We applied a preprocessing algorithm that removes them using approximation methods across selected input images.

Demonstration

Input

Above we can see:

1. Five input images
2. Slider to select batch of images
3. Model selection dropdown

We can notice that the second image is contaminated with clouds.

Prediction

Above we can see:

1. Prediction on the left
2. Ground truth on the right
3. Evaluation metric for each class at the bottom

Setup

0. Create environment

Only tested with Python 3.10. For example, Python 3.13 is unsupported. For anaconda it would be:

conda create -n unit2025 python=3.10

Web visualization

1. Install dependencies

pip install -r visualization/requirements.txt

2. Run

make vis

Training the ConvLSTM

The full training data (~100GB) has to be provided to you by the competition's partner company, Datamole.

1. Install dependencies

pip install -r model/requirements.txt

2. Train

make train