About This Project

Flooding affects ~1.81 billion people worldwide, rising with urban growth. Flood patterns and impacts differ across locations due to varying geography and severity. Deucalion and Pyrrha datasets enable AI to detect flooded areas, wet surfaces, damages, vehicles and people in photos. Expanding them with geographically diverse images is expected to improve AI accuracy and effectiveness, enhancing disaster response by sourcing from thousands of social media posts and protecting lives and property.

Ask the Scientists

What is the context of this research?

Accurate detection of flood-related objects in social media photos has become a major focus in disaster management research in recent years. Can AI models accurately identify and extract these objects while ensuring user privacy? Flood events produce complex and highly variable scenes, whereas past research focused mostly on simpler and more uniform images. We hypothesize that AI models trained on geographically diverse flood images from multiple locations can accurately detect flooded areas, damages, vehicles, and people, extracting only flood-relevant information, while not processing any personally identifiable data, including facial features or other sensitive elements. Analyzing hundreds or thousands of such images—especially with location data—can provide valuable, actionable insights for scientists, authorities, and emergency responders, supporting more effective disaster management.

What is the significance of this project?

Floods are increasingly frequent, and timely, accurate information is crucial for effective disaster management. This project expands the Deucalion and Pyrrha flood image datasets, enhancing AI models’ ability to detect flooded areas, wet surfaces, damages, vehicles, and people in complex and variable real-world scenarios. By providing geographically diverse, quality flood images, the datasets enable scientists, authorities, and emergency responders to train and validate object-detection models for rapid response. Any public disaster management agency worldwide can use these datasets to develop their own models and integrate the resulting data into operational workflows, supporting faster, data-driven decision-making and improved protection of lives and property. Additionally, the datasets can serve as a benchmarking tool for evaluating the performance of new research models, facilitating reproducibility and progress in the field -accessible to any scientist or researcher globally.

What are the goals of the project?

The main goals of the project are related to the expansion of Deucalion and Pyrrha flood image datasets, in terms of volume of photos for both datasets, and in terms of volume of extracted masks respectively. Moreover, even the classification schema of Pyrrha v1.0 is advanced, including 20 different categories, the v2.0 will also include categories that can be more compatible to practical flood models like e.g. flood level.

Target for Deucalion v2.0 is 60,000–70,000 photos (a roughly sixfold
increase from 10,240 in v1.0), half of them related to various flood
events. Target for Pyrrha v2.0 is 8,000–9,000 photos (a roughly fourfold
increase from 2,004 in v1.0) and 45,000–55,000 objects (a roughly
fivefold increase from 11,393 in v1.0).