AI Model Challenge for Earthquake Response – Final Results Announced
The AI Model Challenge for Disaster Response, organized by the International Charter: Space & Major Disasters in collaboration with ESA’s Φ-lab, has successfully concluded. Over the past three months, teams from around the world competed to design artificial intelligence (AI) models capable of automatically detecting building-level earthquake damage from high resolution satellite imagery furnished by the Charter archive.
Charter activations for earthquakes are fairly common, averaging about four per year over the last five years. Rapid assessments of building damage are essential to guide emergency response efforts. Today, this process is still largely manual and time-consuming. The challenge was created to push forward the state of the art by encouraging the development of AI models that can work across diverse regions, datasets, and sensors, mirroring the real operational conditions of Charter activations.
Participation and Engagement
The challenge attracted 143 registered participants from 40 countries, who were given access to the dataset. Of these, 18 teams (maximum 5 members each) completed Phase 1 submissions, and 12 teams carried their models through to Phase 2.
Participants represented a wide mix of expertise:
- 60% AI experts
- 20% Earth Observation experts
- 20% students and early-career researchers
Technical Highlights
The dataset was compiled from the Charter archive, with contributions from KARI, USGS/Maxar, CNES/Airbus, Blacksky, and CNSA, ensuring a sufficient and diverse training resource. The full dataset was 475 GB and contained over 200 images, making it one of the largest and most operationally realistic collections of EO data ever assembled for AI model training concerning disaster response.
Across both phases, seven sites were scored (five in Phase 1, two in Phase 2). Participants faced several technical challenges:
- Class imbalance: For the Mandalay, Myanmar site, fewer than 0.2% of nearly half a million buildings were damaged. Some models under-detected damage, while others over-predicted.
- Multiple sensors: Pre-event imagery (usually from WorldView/GeoEye) was paired with post-event imagery (usually from Pleiades), requiring pan-sharpening and co-registration
- Misalignment issues: Differences in acquisition angles and sensors required additional correction steps
These conditions made the competition robust and realistic. In practice, the ideal model identifies nearly all damaged buildings with limited false detections of undamaged ones. For large-scale events, over-detection is less problematic than under-detection: flagged areas can be verified manually but missed damage risks leaving affected communities unnoticed.
Final Results
Final rankings were based on F1 scores averaged per site, with Phase 1 and Phase 2 weighted 40/60. Four teams emerged as clear leaders, separated by only small margins:
- TelePIX – 0.7067
- Datalayer – 0.6698
- DisasterM3 – 0.6598
- Thales – 0.6469
Because of the closeness of results and the quality of the submissions, the challenge organizers and Charter partners decided to recognize the top four for this challenge. Below are examples of successful predictions by each of the four winning teams. The examples shown are over the Mandalay, Myanmar site used in phase 2 of the challenge. The post-event images shown are from Pleiades © CNES 2025, distribution Airbus DS.
Meet the Winning Teams
- TelePIX (Republic of Korea) Led by JaeWan “Eric” Park, TelePIX Co., Ltd. is a Seoul-based NewSpace startup founded in 2019. The company develops ultra high-resolution EO/IR nanosatellites, big data analytics, and ground systems. Its onboard AI processor, TetraPLEX, demonstrated real-time in-orbit processing on a SpaceX Falcon 9 in 2024. TelePIX is recognized as a World Economic Forum Technology Pioneer and has partnered internationally on advanced EO missions. Notably, TelePIX also won the last AI4EO competition, MapYourCity.
- Datalayer (Belgium) Represented by Eléonore Charles, a data scientist with expertise in NLP, computer vision, and MLOps. Eléonore currently works as Product Manager at Datalayer, which provides a platform for scaling Jupyter workflows to the cloud. Datalayer enables data scientists to run large-scale training and analysis seamlessly across computing environments.
- DisasterM3 (Japan) Led by Dr. Junjue Wang, researcher at the University of Tokyo’s Machine Learning and Statistical Data Analysis Lab. Dr. Wang’s work focuses on multi-modal remote sensing, large-scale land-cover mapping, and landslide detection. His academic contributions span NeurIPS, AAAI, and ISPRS. DisasterM3 applied this strong research background to develop an adaptable geospatial AI solution for the challenge.
- Thales (France) Led by Nicolas Dublé, an Image Processing and Deep Learning Engineer at Thales Services Numériques. Nicolas has extensive experience in remote sensing and AI. Remarkably, he developed his solution single-handedly, with only minimal external support.
The top four teams were invited to present their results at the Charter Board Meeting, held in Strasbourg, France from October 6-10 and hosted by CNES. Here the teams were able to present their unique solution to the challenge and converse with Charter members.
Looking Ahead
The results of this challenge show the promise of AI in accelerating post-earthquake assessments. The winning teams demonstrated innovative approaches and strong generalization capabilities, bringing us closer to operational integration within the Charter.
The Charter members are also exploring the possibility of future AI4EO challenges, potentially focused on a different hazard scenario of the Charter, once again leveraging the rich Charter archive. The International Charter thanks all participants for their dedication and ingenuity. Together, we are advancing the use of space data and AI for faster, more effective disaster response worldwide. Image insert here