254-10 Cracking the Code of Fragmented Fossils: A Machine Learning Approach to Irregular Echinoid Identification

Session: Recent Advances in Fossil Imaging

Presenting Author:

Authors:

Abstract:

Since the skeletal remains of many marine invertebrates are preserved as fragments, a significant portion of the invertebrate fossil record is frequently overlooked or discarded. To address this long-standing gap in paleontology, it is essential to develop methods capable of accurately identifying fossil fragments. As an initial step, we developed a machine learning model to identify fragments of irregular echinoids, with the goal of eventually extending this approach to regular echinoid fragments. Our method first employed the Segment Anything Model (SAM) to perform zero-shot instance segmentation, automatically isolating the primary object within an image. From this initial mask, we programmatically generated a dataset of polygonal fragments of Clypeaster rosaceus, Clypeaster subdepressus, Encope aberrans, Encope michelini, Leodia sexiesperforata, Mellita tenuis, and Meoma ventricosa where the fragment size is systematically controlled as a percentage of the object's diameter (e.g., 100%, 50%, 25%). This allowed us to precisely titrate the amount of visual information available. Then a pre-trained ResNet-18 classifier was trained on these fragment datasets to evaluate how classification accuracy changes with the level of partial visibility. Classification accuracy drops sharply with a decreasing fragment size, but doesn't fall below ~50%, even at the smallest size. This suggests that very small fragments still retain some weakly informative features, even though accuracy is low. This challenges the idea that fragmentary remains are not useful. Certain regions of the irregular echinoid test carry more diagnostic features that help the model distinguish among species, while other regions do not provide enough information when isolated as small fragments. However, fragments altered by taphonomic processes may exhibit more complexity than artificially generated fragments. These findings emphasize the potential to recover meaningful taxonomic information from material that is typically disregarded, initializing a method for more unbiased and representative paleontological analyses.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-8447

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.