IndabaX Madagascar 2024
Empowering AI Solutions for Local Challenges with a Storm Nowcasting Hackathon
Young African scientists have been able to create innovative new weather prediction tools in an AI hackathon held in Madagascar in December 2024.
Overview
IndabaX Madagascar 2024, held in Antananarivo from 13 to 15 December 2024, was a key event that brought together Malagasy students, researchers, and Artificial Intelligence (AI) enthusiasts to explore the applications of AI in addressing local challenges. The event featured a variety of activities, including workshops, keynotes, and an ideathon. One of the most exciting aspects was the storm “nowcasting” hackathon (https://www.kaggle.com/competitions/hackathon-DL-IndabaX-MDG-2024/overview), the first of its kind in Madagascar. “Nowcasting” uses data-science methods to predict the development of severe weather over a period of a few hours, and is vital for creating alerts of Africa’s weather extremes. Hackathon participants developed innovative AI solutions for predicting thunderstorms in real-time, focusing on Nosy Be, an important economic region of Madagascar prone to frequent thunderstorms.
Problem statement: Using real-time data from the nearest storm (colored blob) at time t0, including location, size, intensity, and distance, to forecast storm activity in Nosy Be (red plus sign +) for the next 1 to 3 hours.
Inspiration Behind the Hackathon: Mendrika Rakotomanga's Research
This hackathon was inspired by the recent work of Mendrika Rakotomanga (https://eps.leeds.ac.uk/maths/pgr/13569/mendrika-rakotomanga), a Postgraduate Researcher in Applied Mathematics, and his supervisory team, led by Professor Douglas Parker (https://environment.leeds.ac.uk/see/staff/1469/professor-douglas-parker). They developed an effective storm-based machine learning approach for nowcasting thunderstorms at specific locations over the next few hours. This approach uses real-time data from nearby storms, incorporating key factors such as time of observation, location, intensity, size, and distance, all extracted from satellite observations. The data have been made available by the UK Centre for Ecology and Hydrology (UKCEH), through a collaboration with the University of Leeds.
Keynotes to Introduce Storm Nowcasting
The hackathon began with two keynotes that sparked interest both in person and online. Each session had about 25 in-person attendees and was streamed on Facebook. Mendrika Rakotomanga gave a talk titled "Advancing Thunderstorm Prediction: The Role of AI and Satellite Data in Africa." He discussed his work on AI-based thunderstorm nowcasting and highlighted international projects like FASTA, WISER-EWSA, and NAIAR that focus on advancing nowcasting in Africa. The session received 1,600 plays and 65 engagements (https://www.facebook.com/share/v/1BAfragPmW/?mibextid=WC7FNe). Dr. Stephan Bojinski from EUMETSAT presented "New-generation Satellite Data for Nowcasting from Meteosat Third Generation," explaining how new satellite data can enhance nowcasting. His talk reached 920 plays and 40 engagements (https://www.facebook.com/share/v/17bSWRoiq7/?mibextid=WC7FNe). These talks provided valuable insights to participants, particularly those involved in the hackathon.
Some photos from the keynote sessions
Hackathon Overview and Results
The hackathon was hosted on Kaggle as an open competition, attracting 129 entrants, 65 participants, and 1,941 submissions. Among the 65 participants, 22 were from Madagascar, most of whom were STEM majors (https://www.kaggle.com/competitions/hackathon-DL-IndabaX-MDG-2024/overview). Remarkably, 18 of these local participants developed AI solutions that outperformed an operational thunderstorm nowcasting system. Prizes were awarded to the top five winning solutions from participants based in Madagascar.
The competition showcased impressive models, including those leveraging Artificial Neural Networks and Boosting Methods such as XGBoost and LGBM (https://www.kaggle.com/competitions/hackathon-DL-IndabaX-MDG-2024/code). Feature engineering played a vital role in optimising model performance by incorporating temporal, spatial, and storm-specific characteristics. Participants created interactions, lagged variables, and derived metrics—such as cyclone season weights, radial distances, and intensity-density—to enhance predictive accuracy.
This success highlighted the potential of local African participants to create innovative solutions using AI for storm nowcasting for their own countries. Hosting the hackathon on Kaggle also attracted international participants from countries such as India, Pakistan, Turkey, and Uganda (https://www.kaggle.com/competitions/hackathon-DL-IndabaX-MDG-2024/leaderboard). Some participants suggested broadening the competition, as prizes were limited to participants residing in Madagascar. They noted that opening the event to a wider audience, perhaps across all Africa, could reach greater engagement and innovation.
Hackathon and ideathon winners
Event Supporters and Partners
We thank EUMETSAT, Dr Cornelia Klein (UKCEH, and Mendrika Rakotomanga (University of Leeds) for the datasets, and extend our gratitude to Deep Learning Indaba, Professor Douglas Parker, the University of Leeds, and the WISER-EWSA project for their funding and support. Finally, we acknowledge the contributions of our partners: AlgoMada, VirtuoCode, the Laboratoire d'Intelligence Artificielle de Madagascar, AMLD Africa-MG, WeR, Zindi, Ikala STEM, Optimize Insights, and Centre de Recherche en Education Environnementale .