PROJECTS
Team Members: Diana Mosquera, Francisco Gallegos,
Juan Daniel Vásconez, Karla Mosquera, Pedro Merino, Silvia Vallejo
INstituto geofísico EPN + MODEMAT
Development of a Volcanic Monitoring System with Convolutional Neural Networks (CNN)
Our research, in collaboration with Modemat and the Geophysical Institute of Escuela Politécnica Nacional, presents a methodological framework for developing an advanced volcanic monitoring system that combines thermal imaging with artificial intelligence, specifically designed for volcanic surveillance needs in Ecuador. The system processes raw images captured by FLIR cameras (.fff format) through metadata extraction, thermal analysis and automated classification. The key element of the system lies in its three-dimensional tensor processing that captures both the spatial dimensions (x,y pixel coordinates) and the temporal evolution (z-dimension) of the thermal patterns. This approach simultaneously analyzes three fundamental aspects: thermal information, edge detection and thresholds at different temperature levels.
For analyzing these complex data, we implemented a Multi-Branch Convolutional Neural Network architecture. This architecture processes the three types of thermal information in parallel, later merging the extracted features to generate accurate classification of volcanic state. The model was trained with an extensive thermal dataset (approximately 7 GB), implementing regularization techniques to ensure its performance under variable conditions.
The results are exceptional: the system achieves 98.86% accuracy in detecting volcanic emission events, with robust performance in distinguishing between clear (87.70%) and cloudy (81.25%) conditions. This accuracy makes the system a reliable tool for real-time monitoring. Practical applications include early warning systems for communities near volcanoes, continuous monitoring of volcanic activity, improved risk management through timely information, and advances in volcanological research through more precise and constant data.
The project has validated its applicability in real monitoring scenarios, working effectively even under different atmospheric conditions and observation angles. This methodology can be implemented on other volcanoes, significantly contributing to the safety of vulnerable populations. The results of this research will culminate in the publication of a scientific paper in a specialized journal, sharing this methodology with the international volcanology community and establishing a new standard in AI-assisted volcanic monitoring.
Return ⇐
