What Volcanologists Do - Using Artificial Intelligence to detect lava lake thermal patterns at Kilauea volcano

Goal: Since gases are so difficult to measure repeatedly (the technology is expensive, so it can't be left out for long periods of time), can we detect patterns on the lava lake that are indicative of high gas levels? If we can then we can simply look at the pattern of the lava lake and give warnings about high gas readings even when we can’t directly measure the gas levels.

Why is this important: So much gas comes off the lava lake that it has affected the health of downwind residents, forcing some to move. It would be helpful to provide warnings to downwind residents when gases are extremely high. However, gases can not yet be measured continually, but we do have cameras that run round the clock and provide views of the lake surface that could indicate high gas levels. Below you can see the gas plume (red arrow) from Kilauea blowing primarily to the southwest.

Below shows some of the patterns on the lava lake detected by thermal cameras that are permanently set up on the edge of the crater.

What did we do? We obtained several thousand thermal images of different patterns and used an artificial intelligence (AI) application called “self organizing maps” that looks for patterns in 2d data. When you have thousands of images, the human mind simply can’t process that much data. Once AI lumps similar images into categories, we look at the times for all of the images in a category and compare them to everything we can, such as gas, seismicity, tilt and see if the pattern is indicative of high levels of any of these eruption processes.

What did we find?  

We found that AI first lumped the images according to average temperature of the surface, and not the pattern. In order to isolate the pattern, we “normalized” all of our images so that they were the same average temperature while preserving the pattern. When we applied AI to the normalized images that isolated the patterns we found that the highest gas levels were indeed tied to one of the categories. This is a very useful result because we can indeed detect high gas levels by looking at lake patterns, and someday may be able to warn downwind residents of high gas levels.