AUTHOR=Torres Karina M. , Tan Chenjiao , Mollet Kayla A. , Smith-Pardo Allan H. , Xu Rui , Li Changying , Paula-Moraes Silvana V. 

TITLE=Integration of wing pattern morphology and deep learning to support Plusiinae (Lepidoptera: Noctuidae) pest identification

JOURNAL=Frontiers in Agronomy

VOLUME=Volume 7 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/agronomy/articles/10.3389/fagro.2025.1602164

DOI=10.3389/fagro.2025.1602164

ISSN=2673-3218

ABSTRACT=Soybean looper (SBL), Chrysodeixis includens (Walker) (Lepidoptera: Noctuidae: Plusiinae), a major pest native to the Americas, poses considerable management challenges. Sex pheromone trapping in IPM programs represents a tool to detect initial infestations and promote timely management decisions. However, commercial formulations of sex pheromone for SBL are non-specific, leading to the cross-attraction of morphologically similar plusiines, such as cabbage looper (CBL), Trichoplusia ni (Hübner), and gray looper moth (GLM), Rachiplusia ou (Guenée). Current identification methods of plusiine adults are laborious, expensive, and thus inefficient for rapid detection of pests like SBL. This study explores the use of deep learning models and visualization techniques to explain the learned features from forewing patterns as an identification tool for SBL and differentiation from morphologically similar plusiines. A total of 3,788 unique wing images were captured from specimens collected from field and laboratory populations with validated species identification. Five deep learning models were trained on lab-reared specimens with high-quality wing patterns and evaluated for model generalization using field-collected specimens for three classification tasks: classification of SBL and CBL; male and female SBL and CBL; and SBL, CBL, and GLM. Our results demonstrate that deep learning models and the visualization methods are effective tools for identifying plusiine pests, like SBL and CBL, whose wing patterns are difficult to distinguish by the naked human eye. This study introduces a novel application of existing deep learning models and techniques for quickly identifying plusiine pests, with potential uses for pest monitoring programs targeting economic plusiine pests beyond SBL.