Automated Mapping of Lava: Classification of the Lava Flow Field from the 2021 Fagradalsfjall Eruption

Abstract

In March 2021 an effusive eruption began in the Fagradalsfjall volcanic system in Iceland. It was the first in a new series of volcanic events on the Reykjanes Peninsula that threaten communities and infrastructure in the area. Near real-time mapping of the lava flow field is essential for hazard assessment during an eruption. This study aims to develop efficient automated mapping methods of lava flows that could be applied to future Reykjanes eruptions. Aerial photogrammetric data used to monitor the 2021 Fagradalsfjall eruption is utilized to develop and test classification workflows for mapping of both the lava flow field outline and lava channels. Pixel-level classification was first tested and determined to be sufficient for mapping the lava flow field outline based on favorable results. Orthomosaics and layers derived from Digital Elevation Models (DEMs) were used for training of a Random Forest classifier. Manually digitized outlines of the lava flow field were used as ground truth; the result is a binary classification (lava or not-lava). The unsupervised K-means algorithm could be added as an intermediate step to create a multi-class model. After training, the models were tested on the entire survey area and post-processing steps were performed to minimize error. This workflow was applied to both the 30 September post-eruptive survey and the 26 June syneruptive survey. The post-eruptive binary model achieves a 96.7% accuracy against the manually digitized ground truth, while the multi-class model achieves a 99.8% accuracy. The syn-eruptive binary model achieves a 98.6% accuracy, and the multi-class model achieves a 97.3% accuracy. In all cases, the classified outline follows the manual outline well. The multi-class model can correctly classify incandescent lava and lava with sulfur deposits. However, gas clouds present in the syn-eruptive survey negatively affect classification. The addition of a thermal layer was also tested by applying the workflow to the February 2024 Sundhnúkagígar eruption and achieves a 98.3% classification accuracy. The results demonstrate that the workflow developed is a fast and efficient method of mapping the lava flow field outline of an active eruption. A pixel-based approach was not sufficient to map lava channels; Object-Based Image Analysis (OBIA) was needed instead. Mean-shift segmentation on the 30 September data was performed with the open-source Orfeo toolbox. Zonal statistics were calculated for each segment. A Random Forest model was then trained using channels identified by Hutchinson (2023) as ground truth, and then tested on the entire lava flow field. The model succeeds in mapping some channels but also produces significant error. Resampling of the training data and addition of spatial statistics appears to marginally improve accuracy. A more complex classification technique may be needed to accurately map nonincandescent channels for hazard assessment; however, incandescent channels are mapped well by the pixel-based workflow.