Accurate identification of thirteen fly species from three families using wing venation patterns with machine learning approaches / Ling Min Hao

Ling Min Hao, Min Hao (2024) Accurate identification of thirteen fly species from three families using wing venation patterns with machine learning approaches / Ling Min Hao. Masters thesis, Universiti Malaya.

PDF (Thesis M.A) Restricted to Repository staff only until 31 December 2025. Download (1663Kb)

Abstract

The ease and the affordability of image data acquisition have made whole-image analysis an attractive analytical approach in biological research. Coupled with machine learning, whole-image analysis has the potential to complement or even supplant traditional morphometric approaches for species identification in medical, veterinary, and forensic entomology. Here, I used a substantially expanded dataset (n = 759; 13 species and a species variant; 3 families) to consolidate findings from a pilot study (n = 74; 15 species; 2 families) for automated species identification of fly species based on their wing venation patterns, using classical Krawtchouk moment invariants coupled with a random forest model. To leverage on state-on-the-art methods on image analysis, I conducted a comparative analysis using ResNet, a deep learning model. Five-fold cross validation results show impressive mean identification accuracies of 98.56 ± 0.38% and 99.60 ± 0.27% at the family level, and 91.04 ± 1.33% and 97.87 ± 1.01% at the species level, for the classical and deep learning approaches, respectively. Additionally, the mean F1- scores of 0.89 ± 0.02 and 0.97 ± 0.01 respectively indicate a good balance of precision and recall for both models. Importantly, the regions on the fly wings that are used by ResNet for species identification were successfully visualised using Grad-CAM heatmaps, thus facilitating the interpretation of putative biological bases of identifications using ResNet. In summary, this study demonstrates the extent to which species differences in the studied dipteran species can be expressed in wing morphology, both quantitatively and qualitatively, through image data. Specifically, the findings from interpretable deep learning are potentially useful for generating hypotheses about putative wing anatomies that hold taxonomic value.

Actions (For repository staff only : Login required)