A machine learning approach to species identification of European sepsid dung flies (Diptera: Sepsidae)

2026/07/30 by Ramon Dallo, Jonathan Rosenthal, Jeannine Roy +1
Agricultural and Biological Sciences · Mathematics · #Diptera species taxonomy and behavior #Forensic Entomology and Diptera Studies #Morphological variations and asymmetry

paper · doi:10.35929/rsz.161

crossref issued 2026/07/30 · crossref published 2026/07/30 · crossref published-print 2026/07/30 · openalex publication_date 2026/07/30 · crossref created 2026/07/31 · crossref deposited 2026/07/31 · openalex created_date 2026/08/01 · crossref indexed 2026/08/01 · openalex updated_date 2026/08/02

Abstract

Widespread species are important to deepen our understanding of ecology and evolution, and field experiments are particularly useful to observe natural processes. A crucial aspect of field studies is the correct identification of species, which remains a challenge for small insects. Species identification in insects is traditionally often based on male genitals or secondary sexual traits, so in many cases it is difficult if not impossible to determine the species of wild-caught female specimens. Here we combine geometric morphometric data of wing shape with a machine learning algorithm employing a random forest classifier for differentiating closely related European dung fly species of the same genus Sepsis (Diptera: Sepsidae). The training dataset included 190 samples for each of nine similar species coexisting on pastures in Switzerland. We ultimately tested performance with a novel dataset of 270 male and female specimens (30 per species). Interestingly, the algorithm was better at identifying female specimens, with an accuracy of 82%. Population and sample age variation in training data sets somewhat influenced the accuracy of species predictions. Overall, accurate species identification within a single genus proved difficult for these closely related sister species displaying very similar traits, limiting our machine learning approach for general species identification of sepsid flies. However, females of some species showed promising precision scores, so implementing more complex algorithms using full pictures of the wing or the entire fly might enhance performance to optimize differentiation of very similar species.

Citations

• ONTbarcoder 2.0: rapid species discovery and identification with real‐time barcoding facilitated by Oxford Nanopore R10 .4
• Scientific Machine Learning Benchmarks
• Automated Taxonomic Identification of Insects with Expert-Level Accuracy Using Effective Feature Transfer from Convolutional Networks
• Discrimination of Eubazus (Hymenoptera, Braconidae) sibling species using geometric morphometrics analysis of wing venation
• Biological identifications through DNA barcodes
• Backpropagation Applied to Handwritten Zip Code Recognition
• The Elements of Statistical Learning: Data Mining, Inference, and Prediction
• Does better taxon sampling help? A new phylogenetic hypothesis for Sepsidae (Diptera: Cyclorrhapha) based on 50 new taxa and the same old mitochondrial and nuclear markers
• Behavioural mechanisms of reproductive isolation between two hybridizing dung fly species
• Geographic patterns of postzygotic isolation between two closely related widespread dung fly species (Sepsis cynipseaandSepsis neocynipsea; Diptera: Sepsidae)