Publication Date

Spring 2022

Degree Type

Master's Project

Degree Name

Master of Science (MS)


Computer Science

First Advisor

Phil Heller

Second Advisor

Nada Attar

Third Advisor

Amanda Kahn


CNN, Sponge Spicules, Porifera


The phylum Porifera includes the aquatic organisms known as sponges. Sponges are classified into four classes: Calcarea, Hexactinellida, Demospongiae, and Homoscleromorpha. Within Demospongiae and Hexactinellida, sponges’ skeletons are needle-like spicules made of silica. With a wide variety of shapes and sizes, these siliceous spicules’ morphology plays a pivotal role in assessing and understanding sponges' taxonomic diversity and evolution. In marine ecosystems, when sponges die their bodies disintegrate over time, but their spicules remain in the sediments as fossilized records that bear ample taxonomic information to reconstruct the evolution of sponge communities and sponge phylogeny.

Traditional methods of identifying spicules from core samples of marine sediments are labor-intensive and cannot scale to the scope needed for large analysis. Through the incorporation of high-throughput microscopy and deep learning, image classification has made significant strides toward automating the task of species recognition and taxonomic classification. Even with sparse training data and highly specific image domains, deep convolutional neural networks (DCNNs) were able to extract taxonomic features among morphologically diverse microfossils. Using transfer learning, training a classifier on pretrained DCNNs has achieved recent successes in classifying similar microfossils, such as diatom frustules and radiolarian skeletons.

In this project, I address the reliability of pretrained models to perform spicule identification and class-level classification. Using FlowCam technology to photograph individual microparticles, our dataset consists of spicule and non-spicule types without additional image segmentation and augmentation. Our proposed method is a pre-trained model with a custom classifier that performs two different binary classifications: a spicule vs non-spicule classification, and a taxonomic classification of Demospongiae vs. Hexactinellida. We evaluate the effect of implementing different DCNN architectures, data set sizes, and classifiers on image classification performance. Surprisingly, MobileNet, a relatively new and small architecture, showed the best performance while still being the most computationally efficient.

Other studies that didn’t involve MobileNet had similar high accuracies for multi-class classifications with fewer training images. The reliability of DCNNs for binary spicule classification implicates the promising approach of a more nuanced multi-class/taxonomic classification. Future work should build multi-class classification that ranges more biogenic materials for the identification or more sponge taxonomic levels for species classification.