We started by reading the challenge guidelines and reformulating the task as a supervised regression problem: predicting sea-ice velocity components from wind and environmental variables.
Tracking Sea-Ice Movement with AI
- Featured
- AI
Introduction
What was this project about?
This Climate Crisis Hackathon challenge explored how machine learning could help predict Arctic sea-ice movement from wind fields and environmental data.
My Role
I contributed to the machine learning pipeline, including problem reformulation, data visualization, preprocessing, feature exploration, model training, model comparison, and test-set prediction generation.
Technical Details
Machine Learning: Python, Scikit-learn. Dataset: Daily Arctic sea-ice drift observations, combining ice velocity with wind, sea-ice concentration, sea-ice thickness, position, date, and distance-to-coast features.
Timeline
48h
Links
TL;DR
Problem
Arctic sea ice moves under the influence of wind and ocean currents, sometimes by several kilometers per day.
- Sea-ice motion is difficult to model because it depends on several environmental factors.
- Reliable sea-ice velocity information is often obtained from drifting buoys.
- The challenge was to reproduce buoy drift using wind fields and other relevant features.
Solution
As part of a multi-disciplinary team, I built and compared machine learning models to predict sea-ice velocity from environmental data.
The model used features such as wind velocity, sea-ice concentration, sea-ice thickness, position, and distance to coast to predict the u and v components of buoy drift.
information
Process
We visualized the dataset and used correlation heatmaps to understand relationships between wind, ice, position, thickness, and buoy drift.
We normalized the data with MinMaxScaler and tested PCA to make the features easier to compare and reduce dimensionality where useful.
We trained a simple linear regression model as a baseline, giving us an initial reference point with a validation MSE of about 25.
We trained and compared KNN, Ridge, Random Forest, and MLP with 5-fold cross-validation, then used GridSearch to tune key parameters such as KNN neighbors and Random Forest settings.
Impact
15.9 MSE
best validation result obtained with a Random Forest regressor using 5-fold cross-validation.
Model comparison
Random Forest outperformed the other tested models, followed by KNN and MLP.
Climate AI
The project showed how machine learning can be applied to climate-related data and environmental prediction problems.