Soybean looper (SBL), Chrysodeixis includens (Walker) (Lepidoptera: Noctuidae: Plusiinae), is a polyphagous economic pest that is widely distributed across its native range of the Americas (Mitchell et al., 1975; Eichlin and Cunningham, 1978). The larvae of SBL feed on over 170 plants from 39 families, including crops such as the common bean, sunflower, tomato, cotton, and peanut (Specht et al., 2015). It is also a major soybean pest (Bortolotto et al., 2015; Huseth et al., 2021; Debnath et al., 2024). In 2023, the total crop losses and management costs of foliar insecticide applications for managing SBL in soybean-producing U.S. states were valued at nearly US$84.0 million, where southern states represented 97% of the economic impact of this pest in soybean production (Musser et al., 2024). Reports of insecticide resistance in SBL populations represent an additional challenge for the management of this pest in the U.S., leading to a growing concern for the continued economic impact of SBL in cultivated systems, especially in soybean (Reisig, 2015; Catchot et al., 2016). Field sampling techniques, such as beat sheet knockdown and sweep netting, are some of the standard methods for infestation estimation of larval defoliators (Pedigo and Rice, 2014), and according to Huseth et al. (2021), management decisions for SBL have been primarily based on defoliation estimation. However, the challenges related to SBL management are due to the larval feeding behavior of initial instars, which is mainly concentrated on the lower canopy of soybeans (Carter and Gillett-Kaufman, 2018). The larval infestation detection is difficult until the later 4–6 instars when 97% of economic damage occurs (Reid and Greene, 1973; Boldt et al., 1975). Curative chemical control is recognized as a challenge for this pest because of lower susceptibility to insecticides in late instars, compromising the efficacy of insecticide applications. In this way, detecting initial infestations of SBL is critical for timing management decisions in IPM programs.

Due to the limitations of larval and defoliation estimations as tools for SBL monitoring, sex pheromone trapping is used to detect adult moths and, consequently, egg and initial larval instars to promote efficient adoption of management measures. In general, traps baited with commercial formulations of sex pheromones have been successfully applied for detection and population monitoring of several pests with the advantages of being inexpensive, reliable, and less labor-intensive (Witzgall et al., 2010). Using flight phenology information combined with moth trapping of target species, sex pheromone trapping represents a proactive IPM approach by triggering timely field sampling and consequent adoption of management decisions before economic damage levels are reached by the pest (Pedigo and Buntin, 1994). However, the commercial formulations of sex pheromone for SBL and other pests of the Plusiinae subfamily are non-specific, resulting in cross-attraction of morphologically similar species (Jones and Duckworth, 2017; Shaw et al., 2021). An inventory performed by Shaw et al. (2021) listed four plusiines cross-attracted in SBL monitoring programs, including Ctenoplusia oxygramma (Geyer); Trichoplusia ni (Hübner), cabbage looper (CBL); Rachiplusia ou (Guenée), gray looper moth (GLM); and Argyrogramma verruca (Fabricius). Of the four plusiines, C. oxygramma was the primary crossed-attracted species in SBL sex pheromone trapping. However, moths of this species can be easily distinguished from the others based on distinct forewing patterns using the naked eye (Lafontaine and Poole, 1991). Most CBL, GLM, and A. verruca represented less than one percent of the cross-attracted moths (Shaw et al., 2021). However, these plusiines overlap in the flight phenology with SBL, especially with CBL and GLM, which are also ranked as secondary pests in soybean and peanut (Sutherland and Greene, 1984; Allen et al., 2021; Huseth et al., 2021). Cross-attracted CBL and GLM in trapping programs for SBL can compromise timely detection and appropriate insecticide selection to control SBL, considering the growing reports of insecticide resistance in SBL (Reisig, 2015; Catchot et al., 2016).

Most plusiine moths are generic in appearance, with small gray or brown-colored bodies and a characteristic metallic stigma on the forewing with intra- and interspecific variation (Eichlin and Cunningham, 1978). Using wing patterns for species identification is challenging due to the subjectivity of interpreting descriptions. Shaw et al. (2021) indicated that the native plusiines from the U.S. cross-attracted to SBL commercial sex pheromone lures are described with ambiguous terminology, abstract characters that appear similar, and highly variable characters, based on descriptions in Lafontaine and Poole (1991). For example, the adults of CBL and GLM are ambiguously described as possessing a “grizzled appearance” by Lafontaine and Poole (1991). The quality of a given specimen and the lighting conditions also influence the presence and integrity of the wing pattern characters required to identify a moth. While alternative identification methods, primarily male genitalia dissections (Eichlin and Cunningham, 1978; Shaw et al., 2021) and DNA analyses (Zink et al., 2023; Gotardi et al., 2024), have been developed to address these challenges, they are costly and time-consuming for rapid management decision-making and forecasting initial infestations of plusiines. Advances in species identification in this group were recently proposed based on geometric morphometric analysis (Smith-Pardo et al., 2025). The wing shapes of the native SBL, invasive Chrysodeixis chalcites (Esper), and potentially other plusiines were distinct using geometric morphometrics for species identification of plusiine pests. However, even though this method represents an alternative for the detection of the invasive C. chalcites in survey programs, it is still time-consuming and inappropriate as a pest monitoring tool in IPM programs.

Recent advancements in computer vision and deep learning methods offer novel capabilities to automate rapid insect identification using images of external morphology (Wäldchen and Mäder, 2018). Convolutional neural networks (CNNs) are particularly powerful tools for image classification and object detection tasks and have demonstrated success in wing-based insect identification at the species level while producing inferences in only a few milliseconds (Chang et al., 2017; Nolte et al., 2024; Spiesman et al., 2024). Because CNNs are designed to process structured grid-like data, they can effectively analyze complex visual features from images, providing taxonomist-level accuracy in distinguishing morphologically similar species (Valan et al., 2019). They have the potential to establish automated identification systems that allow taxonomists to spend their time on conceptually difficult problems beyond simple routine identification (MacLeod et al., 2010; Valan et al., 2019). In this study, we explore the application of CNNs to recognize and distinguish magnified images of the wing patterns of morphologically similar plusiine pests, focusing particularly on identifying the economic pest SBL. Five image classification CNN models, including ResNet50, ShuffleNetv2, MobileNetv3, EfficientNetv2-B0, and YOLOv8s-cls, were trained and evaluated on three different fine-grained image classification tasks (hereafter referred to as classification experiments): the two-species classification of SBL and CBL, the species-sex classification of male and female SBL and CBL, and the three-species classification of SBL, CBL, and GLM. Moreover, the t-SNE algorithm (Maaten and Hinton, 2008) was utilized in this study to visualize the class separability of a model, and the explainable AI technique, GradCAM (Selvaraju et al., 2017), was employed to understand how a model achieves the predictions and identify potential biases. The goal of this study is to provide a foundation for future automated plusiine identification and increase the applicability of pheromone trapping for rapid SBL detection in IPM programs.

In this study, deep learning models demonstrated promising performances on the classification of three economic plusiine pests, SBL, CBL, and GLM, using fine-scale differences in wing pattern morphology. Pest detection and identification have been at the forefront of entomological applications of deep learning methods (Høye et al., 2021; Li et al., 2021; Teixeira et al., 2023). Deep learning-based CNNs have the capacity to recognize and distinguish features in images, making them a promising tool for classification tasks in pest identification and monitoring in IPM programs (Kasinathan et al., 2021; Korsch et al., 2021; Ullah et al., 2022; Qin et al., 2024). In Lepidoptera specifically, the wings are well-suited as features for deep learning tasks due to their large size relative to the insect, ease of image acquisition as a two-dimensional feature, and unique wing pattern morphology (Feng et al., 2016; Chang et al., 2017). While deep learning applications are robust tools for classifying the large diversity of Lepidoptera (Chang et al., 2017; Böhlke et al., 2021; Xi et al., 2022), deep learning models are also capable of separating visually similar objects, including cryptic and morphologically similar species (Böhlke et al., 2021; Nolte et al., 2024; Qin et al., 2024; Spiesman et al., 2024). This type of classification task, known as fine-grained classification, has not been well-explored on morphologically similar lepidopteran adults for pest monitoring and survey programs, though it has received attention for the study of pests on a broader scale (Li et al., 2022; Zhang et al., 2022; Xu et al., 2024). This study introduces an efficient identification method for SBL and CBL, two predominant pests in agroecosystems across several regions in the Americas that share similar wing patterns and close phenology of flight (Shaw et al., 2021), overcoming plusiine identification challenges. Using deep learning models and explainable AI techniques, our study provides a foundation for future applications of automated pest identification in IPM programs.

The two-species classification experiment found that the five models distinguished SBL and CBL with extremely high accuracy (>99%) on the lab dataset, and three models continued to perform well and generalize to the field (>94% accuracy) and noise datasets (>86% accuracy). Overall, the YOLOv8 model demonstrated the highest performance and generalization ability across the three datasets. The lightweight MobileNetv3 model had low accuracy results on the generalization datasets (<53%) and feature maps that highlighted image artifacts, suggesting that this model was overfitted to the lab dataset. The fewer parameters in MobileNetv3 may have contributed to its lower performance relative to the other four models. The GradCAM visualizations of the feature maps in this experiment were valuable in understanding which features the model used to make its predictions. Interestingly, the YOLOv8s-cls feature maps for both species did not often use the discal area or wing stigma for classification. We hypothesized that these regions were used in earlier prediction stages because GradCAM only displays the last convolutional layer for final classification decisions. The feature maps from correct CBL and SBL predictions demonstrated the use of the basal area of the wing, where the basal line tends to be more prominent and clearly defined in SBL wings. A short black basal dash is also present in CBL but not in SBL wings. In the submarginal area, CBL has a discontinuous band of white scales that is not present in SBL. The features used for SBL predictions may be related to an overall subtle difference in wing pattern characteristics between SBL and CBL. Lafontaine and Poole (1991) had ambiguously described CBL wings as having a “grizzled appearance” caused by scattered black scales on a sandy-gray colored pattern. This trait tends to be reduced in SBL due to the general golden-brown color of SBL wings, particularly in the regions that are highlighted by the SBL feature maps. Despite the feature map interpretations described in this study, it is possible that the model used other features that are not easily discernable by the naked eye. The GradCAM visualizations and interpretations contribute hypothesized explanations to the wing features recognized by the model and offer useful context of CBL and SBL wing pattern morphology as an identification aid to current taxonomic descriptions.

The species-sex classification experiment showed that only the YOLOv8s-cls model could simultaneously predict the species and sex of SBL and CBL. Overall, classification performance on SBL was slightly higher than CBL, likely due to the larger number of training images for SBL. On the lab dataset, the five models maintained high accuracy (>90%), but the performance decreased considerably on the field and noise datasets. The loss of accuracy could be attributable to the quality of specimens in the two generalization datasets, indicating that species-sex classification relies on very fine differences in wing pattern morphology. However, the GradCAM feature maps of the EfficientNetv2-B0 model showed a high influence of image artifacts associated with all predictions, suggesting that EfficientNetv2-B0 and the other low-performing models were overfit to the lab dataset. In the case of the YOLOv8s-cls model, it had the highest accuracy results on the lab and field datasets, and the feature maps indicated that the correct predictions were derived from wing pattern features and not random image artifacts. The accuracy results and verification of the feature maps reveal that the YOLOv8s-cls model found species-specific sexual dimorphic wing patterns in SBL and CBL. From the feature maps, it is visually unclear which features were used to make this distinction. There is no evidence or record of sexually dimorphic wing patterns in CBL or SBL, though this has been described in Rachiplusia and a few other plusiines (Kitching, 1987).

Other studies have explored the ability of deep learning and machine learning models to distinguish insect species and sex (Tuda and Luna-Maldonado, 2020; Kittichai et al., 2021; Genaev et al., 2022). However, these studies were performed with whole insect specimens and species that exhibited dimorphism through traits like ovipositor presence, body size, and body color. Silva et al. (2025) described the potential of using deep learning models to diagnose monomorphic species as sexually dimorphic. This application of deep learning has not been applied in entomology until recently, as in this study and Nawoya et al. (2025), where they used deep learning models to separate the sex of monomorphic black soldier fly larvae. The studies that have explored this ability of deep learning models reported similar accuracies as this study that are higher than 50% (Wang et al., 2019; Nawoya et al., 2025; Silva et al., 2025). This finding of sexual dimorphism in SBL and CBL wing patterns highlights the potential of deep learning techniques to improve insect identification.

In the three-species classification experiment, adding GLM provided insights into model performance on an imbalanced training dataset. The models had high accuracy scores (>99%) on the combined lab and field dataset and lower performance on the generalization datasets (52-93%), similar to the results when classifying SBL and CBL in the two-species classification experiment. However, the feature maps of YOLOv8s-cls and EfficientNetv2-B0 predictions displayed a high influence of image artifacts on the species predictions. The EfficientNetv2-B0 model demonstrated the ability to separate at least two species, GLM and SBL, but not all predictions were based solely on wing features. Overall, the model performances suggest that none of the models learned to distinguish wing pattern features across all three species. The imbalanced data, with a limited amount of training data for GLM, likely contributed to model overfitting. These results indicate that larger training datasets (>1,000 images) are necessary for the fine-grained classification of plusiine wings when included in deep learning studies. Increasing the number of training images for GLM relative to the other classes could improve model performance and generalization, as observed in the two-species classification experiment. However, it is noteworthy that this species has a relatively minor occurrence in cultivated systems of soybean and other host plants.

In each classification experiment, the five models demonstrated a high performance on the lab dataset used for training and testing, and the performance decreased when model generalization was evaluated on the field and noise datasets. The loss in accuracy on the field and noise testing sets was expected because these datasets were assembled to differ in specimen quality from the data for model training. The models were not optimized for high performance because the objective was to determine whether deep learning models could distinguish the similar wing pattern morphology of plusiines. The models were trained on specimens in ideal conditions, where the wing and scales were not damaged, obscured, or missing. The field dataset maintained the same image standardization as the training data, but the wing quality was more representative of specimens recovered from field trapping. Field-collected specimens likely incurred discoloration or damage to the wing or scales from the trapping process and the storage, freezing, and drying process (Park et al., 2020; Kittichai et al., 2021). Similarly, the age of the collected specimen can impact the quality of the wing scales and the overall integrity of the wing. The noise dataset differs from the other two datasets because it has images of pinned specimens. The same image magnification was used, but the specimens were collected from the field and remained intact (i.e., forewings were not removed). The three-dimensional nature of the specimen and the presence of other insect parts influenced the quality of the image and background colors, creating additional challenges for model predictions. By the progression of specimen and image quality using the three datasets and verification with GradCAM visualizations, the YOLOv8s-cls model was robust in generalizing to new images in the two-species and species-sex classification experiments, and it used wing characteristics to make predictions and not random artifacts caused by the imaging process.

The models investigated in this study can potentially increase the applicability of pheromone trapping for detecting infestations of SBL and CBL. Pheromone traps for detecting plusiine pests, such as SBL, are challenging because of the non-specificity of sex pheromone formulations (Huseth et al., 2021; Debnath et al., 2024). Still, pheromone traps have the benefit of detecting initial pest infestations. This is particularly effective for SBL, as it is challenging to manage once established due to its larval feeding behavior and resistance to multiple insecticides (Catchot et al., 2016; Reisig and Goldsworthy, 2024). The deep learning models in this study offer an efficient alternative to current identification methods because they can be used independently with ease and produce reliable predictions in a matter of seconds, requiring only an image of the forewing. For applying the models to pest detection using sex pheromone lures, the context of the species included in the model is an important consideration. This study focused on SBL, CBL, and GLM primarily due to their high morphologic similarity and phenology of occurrence in cultivated systems. In addition, reports of cross-attraction of high abundance of C. oxygramma moths represent a concern for management decisions (Shaw et al., 2021). Although C. oxygramma is easily distinguished by the forewing patterns, the current models cannot predict beyond the defined classes. Future research for practical applications of the models should consider including C. oxygramma, allowing the model to differentiate between all commonly trapped plusiine species in the U.S (Eichlin and Cunningham, 1978; Lafontaine and Poole, 1991; Pogue, 2005; Shaw et al., 2021). In addition, due to the quarantine risk of the invasive C. chalcites and the potential of pheromone trapping to detect low-density populations (Witzgall et al., 2010), studies should consider including this species in model classification tasks to improve pest survey programs.

Although the models effectively distinguish SBL, CBL, and, to some extent, GLM, the image acquisition process is highly standardized. The process requires a specialized imaging system and associated workflow for the standardization of images, which limits the operation to one person capturing one image sample at a time. The tradeoff between image standardization and model accuracy is a well-known challenge in deep learning studies. When applying models to different domains, such as images taken with a smartphone, this challenge can be particularly evident (Schneider et al., 2023). Here, the specialized imaging process was important for evaluating the ability of the models to detect a difference in wing pattern morphology. However, the performance of the models beyond this standardization is unknown. Future studies should validate whether the models in this study can generalize to images taken at different resolutions or by other devices. Nolte et al. (2024) specifically explored the use of multiple image-capture devices on mosquito wing classification and found that their models struggled to accurately classify images taken by devices that differed from the device used in the training data. Image magnification and resolution should also be considered when using multiple devices because a model trained on high-magnified images can predict poorly on lesser-magnified images due to the loss of features important to the model (Kittichai et al., 2021; Nolte et al., 2024). In the case of the fine-grained wing morphology classification task, it is unclear what level of magnification is required for accurate classification. The present study is the first to explore SBL and CBL identification based on wing patterns with reliable accuracy at a high magnification. The challenges associated with the image acquisition process by incorporating multiple image-capturing devices and exploring different magnification levels of the wing patterns represent one of the next steps to be addressed to reduce standardization and increase the variability of image quality under field conditions.

This study explored the use of five deep learning models on the identification of economic pests SBL, CBL, and GLM based on their morphologic forewing patterns. It is the first study validating wing patterns as a diagnostic trait for pest identification in the Plusiinae subfamily. In the two-species classification, three of the five models demonstrated robust performances on the identification and generalization of CBL and SBL wings. The results of the species-sex classification revealed sexually dimorphic wing patterns of both CBL and SBL. Based on the GradCAM visualizations, only the YOLOv8s-cls model simultaneously classified the species and sex of the plusiines using the wing pattern features. The sexual dimorphism identified by the model is a novel discovery not only for the species themselves but also for the potential of deep learning techniques for other entomological applications. The three-species classification that included GLM, SBL, and CBL showed similar performance to the two-species classification. However, the GradCAM feature maps indicated model overfitting that was likely caused by the class imbalance, where there was a smaller amount of training data for GLM. In the two-species and species-sex classification experiments, the YOLOv8s-cls model consistently achieved the highest performance by identifying distinguishable wing features for classification and reaching the highest accuracy scores compared to the other models. The deep learning models in this study provide an efficient alternative to current identification methods of moths from the Plusiinae subfamily, including genitalia dissections, DNA analyses, and morphometrics. Moreover, this study lays a foundation for future work in the automated identification of plusiines, particularly SBL in IPM programs. The findings of the study open the opportunity of using deep learning models to classify wing patterns in the group of Plusiinae, such as for identifying the invasive C. chalcites for survey programs. This study takes the first step to the widespread application of automated trapping for pest monitoring and detection in IPM programs for economic pests in the Plusiinae subfamily.