Identification of Natural Hydrogen Seeps: Leveraging AI for Automated Classification of Sub-Circular Depressions

Abstract

Hydrogen has long been used as an energy vector, but the recent discovery of natural hydrogen (H₂) opens the door for its use as a direct energy source. Identifying H₂ seepages is therefore crucial to advance exploration. Although the scientific community does not yet fully understand the parameters controlling H₂ leaks from underground, sub-circular depressions (SCDs) appear to be key indicators associated with these emissions. However, distinguishing SCDs from similar landforms remains a challenge. This study leverages open-source multispectral and high-resolution imagery to train a deep learning model (YOLOv8) for classifying rounded landforms and detecting H₂-related structures (i.e., SCDs). The model achieved 90% accuracy with Google Maps© imagery, outperforming Sentinel-2 multispectral data. Applied to a pre-existing data set from Brazil, the model allowed a large-scale screening, discarding 52% of the structures as non-H₂ emitting ones and pinpointing high-potential areas for field validation. Future enhancements, including, for example, higher-resolution input data and morphometric analysis, would aim to reduce false positives and boost predictive accuracy. This approach significantly improves H₂ exploration efficiency, with global applicability including some region-specific adjustments during post-processing analyses.