2026 Volume 35 Issue 3 Pages 143-150
To estimate stalk numbers during the early growth stage in sugarcane breeding fields, we evaluated deep-learning–based regression models using unmanned aerial vehicle (UAV) imagery. The sugarcane was grown in a breeding field in Nishinoomote City, Kagoshima Prefecture (Tanegashima Island), Japan. Ground surveys and UAV flights were conducted to estimate stalk numbers in each plot and acquire RGB images, from which canopy cover images were generated. A conventional linear regression model based on canopy cover ratio was used as a baseline to evaluate the performance of the deep learning models. The deep learning regression analysis comprised 7 ImageNet-pretrained models across 14 architectures, with fine-tuning performed on cropped RGB images from individual plots. The estimation performances of the linear regression model and each deep learning model were assessed using root mean square error (RMSE). The linear regression model yielded an RMSE of 17.19, which was outperformed by all deep learning models. ConvNeXt-Base achieved the highest accuracy (RMSE=11.16). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations revealed that it focused on soil background regions in images with low stalk numbers but on stalks and leaves in images with high stalk numbers. These findings suggest that deep learning regression models used morphological information, such as leaf shape and texture, which could not be fully captured by canopy cover alone, resulting in more accurate estimation of stalk number. Overall, this study shows the efficacy of image feature extraction by deep learning for estimating stalk number in sugarcane at the early growth stage in sugarcane breeding fields.