Original article was published by Simone Anton on Artificial Intelligence on Medium
The agitated path into AI, predicting a hurricane´s trajectory with ML and the power of remote teamwork.
Join us on our journey to develop a hurricane trajectory prediction model based on machine learning.
The aim of AI Saturdays is to democratize how artificial intelligence is being taught, bringing students from different backgrounds together and fostering a bottom-up, hands-on approach as an alternative to many of the standard AI courses. Even more so, as a key component of each course is to apply the newly gained knowledge in a final project with a social impact.
This third edition of AI Saturdays Madrid started off as any of the previous editions with a key focus on face-to-face collaboration and learning by doing together with all of our fellow students during the sessions each Saturday in Madrid.
Then the ingredients for the perfect storm hit in:
Contact restrictions, no presential classes, a complete lockdown, and uncertainty wherever you would look. At the beginning it looked indeed quite challenging to finish the course in a similar way than any of the previous editions, with a key component like the students´ direct interactions being quite limited in a virtual environment requiring remote teamwork as the foundation to get the maximum number of participants to the finishing line. Yet the AI Saturdays Madrid team managed to hold up the spirit and got the project teams off for a start.
Having voted for different project ideas related with atmospheric phenomena, our remotely created team decided to apply our newly gained knowledge in machine learning to develop a model to predict one of the most critical moments and characteristics of a hurricane´s trajectory: the landfall.
The Caribbean region and the Gulf of Mexico are one of the areas on our planet that are most affected by hurricanes. Though due to the global climate change and rising sea temperatures, these phenomena start to become more frequent in other regions too.
As a starting point we evaluated different data sources and were able to leverage the extensive datasets of previous hurricanes made available by the National Hurricane Center. More specifically we concentrated on the Atlantic hurricane database (HURDAT2) which covers data between 1851 and 2019. After an initial data cleansing and the definition of which elements would be crucial, we extracted the relevant data in order to set up our dataset. Then we developed a model based on machine learning to start analyzing and then predicting the trajectory of the hurricanes with a specific focus on whether the storm would hit land or not.
In order to determine the condition of the landfall we set up a detailed geojson defining the coast of the Mexican Gulf and all of the Caribbean region. Applying a random forest to the dataset and the geojson we trained the model to determine which hurricanes hit land and which didn´t. Leveraging a confusion matrix and a ROC Curve to determine false positive and false negative, with the following results:
Following this step, we then applied a grid search in order to obtain the best parameters and got to the following improvement:
Clearly what we saw was that in our model the two most important features are the longitude and latitude, with little variance coming from other data values such as the wind and so on. As a next step we would therefore analyze how the model behaves without these two measures. In order to determine the accuracy of our model, we applied f1 as a metric. On Github you can also have a closer look at the code we developed.
Apart from applying the newly acquired knowledge in this project, a key challenge in this edition of AI Saturdays was surely to overcome the additional complexity due to the new context. Without being able to meet and collaborate with each other face-to-face, nor exchange ideas with other groups, one of the key pillars was surely to keep up the “remote team spirit”.
Thanks to AI Saturdays Madrid we´re not only taking away new knowledge and an initial insight into the amazing world of artificial intelligence, yet also new friends.
If you have the chance to participate in any of the following editions, our team would definitively recommend to take the chance!
For any feedback and questions, please contact us: