Improved Deep Learning Model for Detecting Pest-Induced Feeding Damage on Soybean Leaves

• Published in: Korean journal of crop science pp. 79 - 91
• Main Authors: 강지우, 신태환, 고종한
• Format: Journal Article
• Language: English
• Published: 한국작물학회 01.06.2025
• Subjects: 농학
• ISSN: 0252-9777; 2287-8432

The accurate assessment of leaf damage is essential for monitoring crop health, optimizing yield, and ensuring crop quality. Convolutional neural networks (CNNs) have demonstrated considerable potential in precision agricultural applications, particularly in tasks such as plant classification and pest detection. In this study, we developed an enhanced Faster Region-based CNN (Faster R-CNN) model to detect pest-induced feeding damage on soybean (Glycine max) leaves under field conditions. Images of ‘Daepung’ and ‘Pungsannamul’ soybean cultivars grown in fields at the Chonnam National University (Gwangju) and National Institute of Crop Science (Wanju-gun, Jeollabuk-do) were collected in 2021 and 2022. The dataset comprised 4,827 leaf images for classification and 795 canopy images for object detection, captured under diverse environmental conditions and featuring complex backgrounds, including soil, weeds, and overlapping foliage. The optimized Faster R-CNN model achieved a mean average precision (mAP) of 72.6%, demonstrating robust performance in detecting pest damage across varying conditions. While the model performed well in detecting partially damaged leaves, its detection performance for severely damaged leaves was lower owing to a class imbalance in the training data. These findings highlight the potential of CNN-based models to accurately detect pest damage in real-time field settings, with opportunities for further improvement through balanced data collection and refined detection strategies. KCI Citation Count: 0