Why did we put our teeth in this challenge?

The movement of Arctic sea ice conveys important information on the effects of climate change in our Northern polar region and vice versa. Being able to analyze the flow of ice through time helps us better understand the dynamical response of the Arctic Ocean to a changing climate.

To model the drift of ice, researchers have placed buoys in the ice that regularly signal their position, so that the velocities of the ice close to these buoys can be measured. However, these buoys are scarce, and there is much land that is not covered by them. It would be beneficial to Arctic Ocean and climate research to be able to reliably fill in these gaps in knowledge by using other data sources that may indirectly contain a way to estimate these velocities for the entire Arctic Ocean.

In this Kaggle style challenge, we tackled this problem by using some exciting machine learning models! Our results indicate an improvement compared to the current state of the art models on this problem.

How did we build our solution?

Charles Brunette from McGill University, the organizer of this challenge, provided us with a great dataset from which we could start tackling this problem head-on. After extensively exploring and visualizing the data (nicely presented in our overview document), we came up with a strategy for predicting sea ice velocities based on an artificial neural network model. For the curious reader; neural networks are a machine learning technique loosely inspired by the way animal brains work, and are able to capture complex relationships in data very well.

We trained our specific version of a neural network on the dataset, and it learned to predict sea ice drifts with accuracies that significantly improve upon the previous state of the art. For the data geeks and ice drift initiated among us: our neural network has a root mean-square error (rmse) of about 4.32 cm/s on the ice speed, and a wind-ice drift angle rmse of about 31.9°.

In the future, we plan to map what kind of function the neural network has learned for mapping the wind velocities and environment variables to the ice velocities. This might provide new insights into the physics of this problem, and could improve the theoretical understanding of this problem as well.

Challenges we ran into

The hardest adversary was time. We had to realize that there was only so much we could try in one weekend, so we had to curb our excitement about the unlimited possibilities in the data and zoom in on a few approaches we could dedicate our time to for two days.

Accomplishments that we're proud of

If anything, we are really proud of pushing this exciting field forward, and making a contribution to the fight against climate change. We are both very passionate about climate change, and to see such a diversity of people coming together during this hackathon and putting in their time and efforts, is truly exhilarating and hopeful for the future.

What we learned

That sea ice drift is a very interesting but tricky problem to tackle. And that we're definitely not there yet!

What's next for AI Brussels - Sea Ice Velocity

We would love to continue the research done here, and explore the many avenues we had to leave out due to time constraints. We've already heard of an interesting new result from Charles that might improve our models even more, and we can't wait to try it out!

Built With

Share this project:

Updates