Machine Learning Quantifies Fine‐Scale Hairiness in Shore Flies (Diptera: Ephydridae)

TL;DR

This paper introduces a machine learning method to efficiently measure tiny hairs on shore flies, helping compare species traits related to water resistance.

Contribution

A novel machine learning workflow using Ilastik and Fiji for quantifying microtrichia in small insects.

Findings

The workflow produces results consistent with manual counts of microtrichia in shore flies.

Paracoenia species from hot springs have more but shorter hairs compared to Parydra species.

Percent coverage of microtrichia per unit area did not differentiate species except at the anterior thoracic spiracle.

Abstract

Morphological analysis of fine structures on small insects is often labor intensive, scale‐limited, and biased by sampling or organismal life history. We used a pixel classification machine‐learning workflow with the open source programs Ilastik and Fiji to identify and quantify microtrichia in semiaquatic shore flies (Ephydridae). This methodology semi‐automates quantification of hairs by counting objects or groups of class‐assigned pixels and determining their percent coverage at a given magnification using scanning electron micrographs. Our results are consistent with manual counts, with Paracoenia species that tolerate hot springs having more hairs than less aquatic Parydra. However, Paracoenia hairs tend to be shorter, and the percent coverage of microtrichia per unit surface area did not differentiate species except for the anterior thoracic spiracle. Our workflow is adaptable for use in other taxonomic groups or beyond the quantification of hairs, with the upper limits of applicability determined by overlap in the feature of interest. As molecular datasets continue to grow and proliferate in the multi‐omics age, efficient morphological workflows become even more critical to allowing proportionally robust, complementary biological inferences grounded in phenotypic data.

We present an efficient, broadly applicable, and open source approach to quantifying micron scale features using a machine learning pixel classification workflow. Using this method, we assess microtrichial variation in semiaquatic Ephydridae species and compare the results to manual assessment. Results are concordant and suggest a relationship between microtrichial length and number in the context of hydrophobicity.

Figures