Forecasting future volcanic activity in long time series orbital infrared data using machine learning

Abstract

Well established baseline data are critical for volcanic change detection and forecasting associated hazards. High resolution thermal infrared (TIR) data captured synoptically over the entire volcanic system and spanning a long time period provide this forecasting capability for many volcanoes around the world. Foundational to these patterns is the subtle (1-2 K) thermal changes, which are easily overlooked using the current lower spatial resolution (1-2 km) TIR data. Therefore, despite decades of spaceborne data acquisition, orbital volcano science still lacks the fundamental ability to forecast a new eruption. Fortunately, several high spatial resolution TIR missions are planned for the coming decade and their data will be crucial to constrain volcanic activity patterns throughout the pre- and post-eruption phases. One of these is the Surface Biology and Geology (SBG) TIR instrument being jointly developed between NASA in the US and ASI in Italy. It is planned to have high spatial (~ 60 m) and much improved temporal (1-3 days) resolutions with the regular production of volcano-specific data products. In preparation for the data from these new missions, we conducted the first study using the entire data record of higher spatial, lower temporal resolution TIR data from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor. ASTER data are the most similar to SBG TIR and most critically, ASTER’s twenty-two-year archive presents a unique opportunity to quantify low-level temperature anomalies to establish baseline behavior. We developed a new statistical algorithm to automatically detect the full range of thermal activity and applied it to >5000 ASTER scenes of five volcanoes with well-documented eruptions. Unique to this algorithm is its ability to use both day and night data, account for clouds, and quantify accurate background temperatures by dynamically scaling depending on the anomaly size. Despite the less frequent temporal coverage of ASTER, the results are an improvement over prior studies that used lower spatial resolution data and show that high spatial resolution TIR data are more effective. Most significant was the finding that the smaller, subtle thermal detections served as precursory signals in ~81% of eruptions. These results also create a framework for classifying future eruptive styles and produced a labeled dataset for use with more advanced machine learning (ML) modeling. Using ML trained on the ASTER results, data from a sixth volcano that was not part of the original study were modeled and the thermally-elevated pixels accurately identified. This thermal anomaly detection approach will be incorporated into the SBG data processing stream to produce crucial daily orbital forecasting of the volcanic activity across the world.