Decision forests for machine learning classification of large, noisy seafloor feature sets

Extremely randomized trees (ET) classifiers, an extension of random forests (RF) are applied to classification of features such as seamounts derived from bathymetry data. This data is characterized by sparse training data from by large noisy features sets such as often found in other geospatial data. A variety of feature metrics may be useful for this task and we use a large number of metrics relevant to the task of finding seamounts. The major significant results to be described include: an outstanding seamount classification accuracy of 97%; an automated process to produce the most useful classification features that are relevant to geophysical scientists (as represented by the feature metrics); demonstration that topography provides the most important data representation for classification. As well as achieving good accuracy in classification, the human-understandable set of metrics generated by the classifier that are most relevant for the results are discussed.