PASSING THE TEST. After training, the team tested their CNN using about 40 other images of each type of particle. They found that the system was 92 percent successful in accurately categorizing an image. For images it couldn't categorize, it provided probability ratios (for example, a 90 percent probability that a particle is vesicular and a 10 percent probability that it's blocky).
As is, the system could already prove useful in eruption response efforts, but the researchers hope to upgrade their CNN to analyze additional aspects of volcanic ash, including its color and texture, providing even more valuable insights into the type of eruption behind the ash.