AI-Driven Crisis Communication for Vulnerable Populations: Automated Classification and Simplification of Weather Disaster Advisories
Abstract
Natural disasters threaten lives and disrupt communities, especially in rural areas where people rely on SMS or radio and have limited internet connectivity. This research tackles a major communication problem in natural disaster management: how do we reach vulnerable populations with urgent weather alerts? To address this, we built a machine learning system to automatically classify weather advisories into urgent versus routine categories and condense urgent warnings into short text messages suitable for SMS delivery. We collected 200 weather advisories from the National Weather Service API and compared three machine learning methods: a simple baseline using TF-IDF and Logistic Regression, an LSTM neural network, and a transformer model (MiniLM) combined with Logistic Regression. Only 19 of 200 advisories (9.5%) were urgent, which reflects reality: most weather alerts are routine rather than life-threatening. The baseline model achieved 100% recall on urgent advisories, correctly identifying all urgent warnings while maintaining 97.5% overall accuracy. The MiniLM model achieved 75% recall. The LSTM model, however, failed to detect any urgent advisories (0% recall). These results revealed that when data is limited, simpler models may outperform complex architectures. The finding matters because emergency data is inherently scarce, so organizations building similar systems will likely face the same constraint. To address accessibility in low-bandwidth environments, we implemented abstractive summarization using DistilBART to condense lengthy advisories into 160-character SMS messages while preserving critical warning information. Our work shows that automated disaster alert systems are feasible even with limited data. We prioritized detecting all urgent advisories (100% recall) over perfect overall accuracy because missing a life-threatening warning is far more dangerous than triggering an occasional false alarm. This work demonstrates that low-cost AI tools can meaningfully improve disaster communication for vulnerable populations. Future work will expand the dataset and conduct human evaluation of SMS summaries with target populations.
Start Time
15-4-2026 10:00 AM
End Time
15-4-2026 11:00 AM
Room Number
303
Presentation Type
Oral Presentation
Presentation Subtype
Grad/Comp Orals
Presentation Category
Science, Technology, and Engineering
Student Type
Graduate
Faculty Mentor
Ahmad Al Doulat
AI-Driven Crisis Communication for Vulnerable Populations: Automated Classification and Simplification of Weather Disaster Advisories
303
Natural disasters threaten lives and disrupt communities, especially in rural areas where people rely on SMS or radio and have limited internet connectivity. This research tackles a major communication problem in natural disaster management: how do we reach vulnerable populations with urgent weather alerts? To address this, we built a machine learning system to automatically classify weather advisories into urgent versus routine categories and condense urgent warnings into short text messages suitable for SMS delivery. We collected 200 weather advisories from the National Weather Service API and compared three machine learning methods: a simple baseline using TF-IDF and Logistic Regression, an LSTM neural network, and a transformer model (MiniLM) combined with Logistic Regression. Only 19 of 200 advisories (9.5%) were urgent, which reflects reality: most weather alerts are routine rather than life-threatening. The baseline model achieved 100% recall on urgent advisories, correctly identifying all urgent warnings while maintaining 97.5% overall accuracy. The MiniLM model achieved 75% recall. The LSTM model, however, failed to detect any urgent advisories (0% recall). These results revealed that when data is limited, simpler models may outperform complex architectures. The finding matters because emergency data is inherently scarce, so organizations building similar systems will likely face the same constraint. To address accessibility in low-bandwidth environments, we implemented abstractive summarization using DistilBART to condense lengthy advisories into 160-character SMS messages while preserving critical warning information. Our work shows that automated disaster alert systems are feasible even with limited data. We prioritized detecting all urgent advisories (100% recall) over perfect overall accuracy because missing a life-threatening warning is far more dangerous than triggering an occasional false alarm. This work demonstrates that low-cost AI tools can meaningfully improve disaster communication for vulnerable populations. Future work will expand the dataset and conduct human evaluation of SMS summaries with target populations.
