Identification of pod borers and pod bugs of cowpea using deep learning techniques

Abstract

The study on “Identification of pod borers and pod bugs of cowpea using deep learning techniques” was conducted during 2022-2024. The objectives of the study were to create a database containing images of pod borers and pod bugs along with symptoms of infestation; selection of the best deep-learning model for accurate pest identification; and documentation of the crop-pest-weather relationship. High-quality images of pod borers and pod bugs at various life stages, including eggs, nymphs (pod bugs), pupae (pod borers) and adults were systematically collected from cowpea fields in the Thiruvananthapuram and Kollam districts of Kerala. All images were preprocessed using the Roboflow platform to ensure consistency and improve model performance. The preprocessing steps involved cropping, resizing, auto-orientation, and contrast adjustment to standardize the images. For the identification of pod borers and pod bugs, separate classes were created based on species. The pod borer model consisted of 10 classes, while the pod bugs model had 4 classes, corresponding to different species and life stages. A total of 20 models (M1–M20) were developed using the You Only Look Once (YOLO) architecture on Google Colab to identify pod bugs, experimenting with different hyperparameters to determine the optimal model. This optimization process involved variation in the number of epochs (20 to 300), image sizes (640×640 pixels and 512×512 pixels) and training-to-testing ratios (80:20 and 75:25). The YOLO-based model (M14) which was trained with 200 epochs, an image size of 512×512 pixels and a training-to-testing ratio of 80:20, achieved an impressive precision of 99.5 per cent in identifying different species of pod bugs. A class-wise performance evaluation was conducted across all models, revealing that the M14 model achieved a precision score of over 90 per cent for each class. The optimized hyperparameters were subsequently applied to additional models developed for pod borer identification and life stage differentiation. The results were impressive, with nearly all classes achieving precision rates exceeding 85 per cent in identifying pod borers. This highlights the effectiveness of the models in accurately identifying both pod borers and pod bugs of cowpea, along with their respective life stages. 94

The seasonal incidence of pest populations was further analyzed to understand their correlations with weather parameters. The pod borer, Maruca vitrata showed a peak population density of 2.65 larvae per five plants during late April (16th SMW) under warmer temperatures (25–30°C). The population of M. vitrata exhibited a significant positive correlation with minimum temperature (r = 0.630) and a negative correlation with evening relative humidity (RH II, r = -0.396). The seasonal data revealed a peak population density of Riptortus pedestris at 7.54 bugs per five plants during late April (16th SMW). This peak correlated positively with maximum temperature (r = 0.637) and minimum temperature (r = 0.559), and negatively with evening RH (r = -0.480). Aphid (Aphis craccivora) populations peaked during the 15th, 22nd and 28th–31st SMWs, reaching up to 75 aphids per leaf. Aphid populations showed a positive correlation with evening RH (r = 0.520) and a non-significant negative correlation with maximum temperature (r = -0.354) and rainfall (r = -0.400). The study successfully developed and optimized YOLO-based models for the precise identification of pod borers and pod bugs, achieving high precision rates of over 85 per cent for pod borers and 90 per cent for pod bugs across different species and life stages. These results establish a robust framework for scalable pest monitoring, with potential applications in mobile or web-based tools for real-time pest identification. Future research could further enrich the database with additional pest species and integrate predictive analytics for proactive pest management. The seasonal incidence analysis provided valuable insights into the relationships between pest populations and weather parameters. Both pod borers (M. vitrata) and pod bugs (R. pedestris) reached peak population densities during warmer periods (16th SMW), demonstrating a significant correlation with temperature and relative humidity. Aphid (A. craccivora) populations also exhibited multiple peaks, showing a positive correlation with evening RH. These findings underscore the importance of considering abiotic factors in developing effective pest monitoring and control strategies for sustainable cowpea cultivation.