Women in Big Data Global

Donate
×

Blogs

From Data to Disaster Response: Inside the WiBD Spring Hackathon 2026

Women in Big Data

By WiBD Team,

June 29, 2026

hack image

Blog is co-authored by: Dhara Awasthi, Amina Wasiq, Ekaterina Kuznetsova, Deepika Ramanathan

WiBD Spring Hackathon 2026: Applying AI to Real-World Wildfire Prediction

On May 22, Women in Big Data (WiBD) successfully concluded its Spring Hackathon 2026 under the leadership of WiBD Global Hackathon Director Rupa Gangatirkar and Technical Advisor Stuti Patel.

Rupa, Lead AI/ML Solution Architect and Senior Data Scientist at Siemens, championed the program and served as the primary organizer, while Stuti, Lead Data Scientist at Samsung, provided technical guidance and coordination throughout the eight-week program.

Rupa opened the closing session by highlighting the remarkable milestones WiBD has achieved over the past 10 years and sharing a preview of exciting initiatives ahead as the organization celebrates its 10th anniversary. Stuti then guided attendees through the hackathon journey, reflecting on the teams’ learnings, technical experiments, survival analysis approaches, and concluding with an engaging Q&A session.

The Spring Hackathon Challenge 2026

On May 22, we celebrated seven incredible teams representing four continents and three WiBD Chapters. Over the previous eight weeks, these teams tackled the WiDS Global Datathon 2026 challenge in collaboration with Watch Duty, building survival models to predict wildfire impacts on critical infrastructure.

By analyzing geospatial data from the first five hours after ignition, participants developed actionable insights to help emergency responders better protect at-risk communities.

Teams worked on a global machine learning competition focused on predicting the time-to-threat for evacuation zones during wildfire events. The challenge was framed as a survival analysis problem using real-world wildfire data, where each of the 316 wildfire incidents (221 training and 95 test samples) included only the first five hours of early perimeter observations.

In total, participants worked with 221 training fires, 34 features, and four probability predictions for every wildfire—leaving very little room for overconfidence.

The objective was to estimate the probability of a wildfire reaching within five kilometers of evacuation zones across multiple time horizons (12, 24, 48, and 72 hours), while appropriately handling right-censored outcomes, where many fires did not reach evacuation thresholds during the observation period. This required careful modeling of incomplete observations while reflecting the uncertainty that emergency responders face in real-world decision making.

Survival-Aware Modeling

One of the most important technical learnings throughout the hackathon was the application of survival-aware modeling.

Unlike traditional supervised learning tasks, this competition required teams to explicitly account for censored data—situations where the final outcome was not fully observed. This shifted the focus from traditional accuracy metrics toward probability calibration and accurate interpretation of time-dependent risk.

Understanding how to model uncertainty and partial observations became central to building meaningful predictive frameworks for real-world wildfire risk assessment.

Collaboration and Learning Across Continents

Beyond the technical challenges, the hackathon became a powerful example of collaboration, resilience, and continuous learning. Teams from around the world brought together diverse backgrounds, experience levels, and perspectives while working toward a shared goal—using data science to solve a meaningful real-world problem.

Team Afghanistan Muzhgan Shams, Blqais Pari Haidari, Muhaddesa Noori, and Hanzala Sirat participated in their very first hackathon.

Beyond the technical modeling, WiBD Afghanistan team shared that the most valuable part of the experience was collaborating as a team. While working with complex data and constrained observations, they continuously exchanged ideas, explored different modeling approaches, and learned from one another throughout the process.

This collaborative environment strengthened both their technical knowledge and teamwork skills, demonstrating the value of collective learning in data science.

Team Inferno Dhara Awasthi, Zuhra Habibi, and Hetavi Mehta shared that the WiDS Datathon 2026 gave them an incredible opportunity to apply their data science skills to a real-world challenge—predicting the probability of a wildfire reaching an evacuation zone within 12, 24, 48, and 72 hours using survival analysis with complex censored data and significant class imbalance.

Dhara reflected on her first hackathon experience:

“It was my first ever hackathon experience, and with no prior background in survival analysis, I started with a very naive approach—a simple Random Forest and Gradient Boosting ensemble that treated each prediction horizon as an independent binary classification problem, without truly appreciating the statistical complexity underneath.”

As the team iterated through different approaches, they realized the biggest challenge wasn’t choosing a model—it was correctly calibrating prediction probabilities across each time horizon.

Their journey explored LightGBM, isotonic calibration, power transforms, meta-learners, and eventually Gradient Boosting Survival Analysis (GBSA). Their strongest results came from blending LightGBM’s classification strengths with GBSA’s ability to model time-to-event behavior and censored observations.

Although they didn’t finish at the top of the leaderboard, Dhara described the experience as deeply rewarding.

“The best model is not always the most complex one, but the one that is honest about uncertainty, validated rigorously through cross-validation, and built with a genuine understanding of the problem structure rather than a blind search for a higher leaderboard number.”

Team Europe  Lisa Fadelli and Ekaterina Kuznetsova shared that their Datathon journey was as much about people as it was about data.

Coming from different backgrounds and skill levels, the team first invested time in building a shared understanding of the challenge before beginning any modeling. That early collaboration became one of their greatest strengths.

Many members explored survival analysis, censored data, class imbalance, and Random Survival Forests for the first time, turning classroom concepts into real-world experience.

Even when only one team member was ultimately able to complete the final submission, they emphasized that the finished solution reflected everyone’s ideas, effort, and encouragement.

Their biggest takeaway was the power of collaboration, curiosity, and persistence.

Team ZenHer Deepika Ramanathan and Swarnima Khadka emphasized that this wasn’t a simple yes/no classification problem.

It was right-censored survival data—the same statistical framework commonly used in medical research to answer questions such as “When will a patient relapse?”—applied to wildfire prediction.

Of the 221 fires:

  • 69 reached an evacuation zone (the timing was known).
  • 152 did not reach an evacuation zone within 72 hours—but potentially could have afterward.

Ignoring censoring would have produced misleading models. Respecting it required survival analysis techniques, including:

  • Kaplan-Meier curves to establish baseline probabilities
  • Cox Proportional Hazards models for per-fire predictions
  • Calibration plots and Brier Scores to evaluate probability estimates
  • Concordance Index (C-index) to measure ranking performance

One of their most valuable discoveries through exploratory data analysis was that distance to the evacuation zone and fire alignment direction were among the strongest predictors of wildfire risk.

Closing Reflections

Being part of the global Women in Big Data community reinforced the importance of inclusion, mentorship, and shared growth in data science.

For the WiBD Afghanistan team, the experience extended well beyond predictive modeling. It represented participation in a global initiative where women in data science collaborate to solve meaningful, real-world challenges using machine learning.

The hackathon demonstrated how data-driven approaches can contribute to disaster preparedness and emergency response while fostering a supportive and empowering global community.

While every team approached the challenge differently, they all shared one common experience—working together across countries and time zones to interpret complex spatio-temporal wildfire data, discuss modeling strategies, and translate early-stage fire observations into meaningful predictive insights.

The environment encouraged continuous learning, idea sharing, mentorship, and peer support throughout the competition.

One participant summed up the experience perfectly:

“The problem was small in data (316 fires) but enormous in depth—exactly the kind of challenge that makes you a better data scientist.”

For anyone who has never explored survival analysis, our teams highly recommend it. The same techniques are widely applied across industries, including customer churn prediction, equipment failure, clinical trials, and now, wildfire response.

Thank You

A heartfelt thank you to our Hackathon Leaders:

 

A special thank you to our mentors:

Sireesha Pulipati, Deepika Ramanathan, Richa Awasthi, Praveen Akunuru, Ekaterina Kuznetsova, and Isha Agarwal for guiding these teams in technical excellence, collaboration, and leadership throughout the program.

Finally, congratulations to every participant who dedicated eight weeks to learning, experimenting, and solving one of today’s important real-world challenges. Your curiosity, perseverance, and willingness to collaborate across borders truly embody the spirit of the Women in Big Data community.

Watch the Closing Session recording here.

Related Posts