Deep learning approaches to trace fossil analysis

Trace fossils are direct records of organism activity and are often abundant where body fossils are absent. Because the responsible tracemaker taxon often cannot be identified, trace fossils are named and classified in an independent parataxonomy called ichnotaxonomy. In vertebrate track ichnotaxonomy, only morphological features related to anatomy should be used, although behaviour, substrate properties and preservation also contribute to the shape of the track. Because of these multiple sources of variation, and the continuous rather than discrete nature of morphological features, ichnotaxonomy involves many subjective decisions that make rigorous quantitative analyses difficult.

Deep learning is a promising approach to the study of trace fossils and can be used to independently test ichnotaxonomy and its conclusions. Here I present a novel similarity learning approach where different models are trained on dinosaur tracks of the same morphotype and locality, modern bird tracks of known species, and/or synthetic data, with the aim of autonomously learning anatomically relevant features. New tracks can be plotted based on these features, revealing clusters that are conceptually similar to, but independent of, traditional ichnotaxa. The models suggest widespread oversplitting of dinosaur ichnotaxa and major inconsistencies in ichnotaxa definitions. They confirm that the dinosaurian ichnofauna of East Asia differs significantly from that of Europe and North America in the Lower Cretaceous. However, this difference does not appear to result from a geographic isolation of East Asia, as commonly assumed, but from the absence of an ecomorphotype in Europe and North America.