Multi-Branch Convolutional Neural Networks for Volcanic Activity Classification Using Thermal Imagery (Cotopaxi and El Reventador volcanoes)

Over the last 25 years, five volcanoes have erupted on mainland Ecuador, generating eruptive columns, pyroclastic density currents, lava flows, etc. Currently, El Reventador and Sangay are erupting and are being monitored by the Instituto Geofísico of the Escuela Politécnica Nacional (IGEPN) using different techniques, including thermal surveillance. The volcanic products emitted by these volcanoes are identified through thermal and visual image analysis. The timely identification of these products can greatly influence decision-making by authorities and the response of vulnerable populations.

This study presents a novel approach to the automated classification of volcanic states using thermal imagery from multiple Ecuadorian volcanoes, acquired by the IGEPN. We developed a Multi-Branch Convolutional Neural Network architecture that processes three-dimensional tensor representations of thermal data to distinguish between clear conditions, cloudy conditions, emission events, and lava flow events. The system processes raw FLIR camera images (.fff format) through a pipeline that includes metadata extraction, thermal analysis, and classification. Our architecture utilizes three parallel branches processing base thermal information, edge detection features, and volcano-specific thermal thresholds simultaneously.

The model was trained and validated on a dataset of more than 10,000 thermal images from two active Ecuadorian volcanoes: Cotopaxi (7,024 images) and Reventador (3,536 images). The dataset encompasses four volcanic states: cloudy conditions, emission events, clear conditions, and lava flow events. Our multi-volcano approach incorporates volcano-specific thermal threshold parameters, recognizing the distinct thermal characteristics of different volcanic systems. The model achieved robust performance with 94.74% validation accuracy and 94.58% training accuracy across all volcanic states and locations. Per-class validation performance demonstrates excellent discrimination capability: <95% for clear conditions, <96% for cloudy conditions, <94% for emission events, and <90% for lava flow events. The confusion matrix reveals minimal inter-class confusion, indicating the model's ability to distinguish between complex volcanic phenomena. This approach addresses key challenges in manual analysis of thermal imagery while providing a scalable framework that can be adapted to different volcanic systems and integrated into existing monitoring networks.