Journal of Korean Society of Agricultural Engineers ISSN:1738-3692(Print) 2093-7709(Online) 한국농공학회논문집

This study investigates the feasibility of soybean cultivar classification under open-field conditions in South Korea using unmanned aerial vehicle (UAV)-based RGB imagery and evaluates the performance of deep learning models. High-resolution RGB images were acquired at nine different growth seasons for four major soybean cultivars—Daewon, Seonpung, Pungsan, and Aram—and the temporal variations in RGB pixel value distributions were quantitatively analyzed. Among the RGB channels, the green channel showed the highest sensitivity to growth stage changes, making it a key feature for classification. Two types of deep learning models, Attention U-Net and DeepLabv3+ with ResNet-50 and ResNet-101 backbones, were applied to the cultivar classification task. The experimental results demonstrated that DeepLabv3+ with a ResNet-101 backbone achieved the highest average F1-scores and the lowest variance across all cultivars, indicating distinguished performance and stability. Classification performance was particularly high for the Daewon and Aram cultivars, while Pungsan showed relatively lower accuracy and higher variability, likely due to its less distinctive visual traits across growth stages. These findings suggest that UAV-based RGB imagery, combined with robust deep learning models, holds strong potential for reliable and scalable open-field soybean cultivar identification.