Abstract
Purpose Sudden Death Syndrome (SDS) in soybeans causes significant yield losses, with no effective biological or chemical control. Accurate detection and severity assessment can help mitigate losses through preventive measures and breeding programs. Traditional severity assessment is labor-intensive and requires expertise, highlighting the need for automated and precise detection.
Methods This study aimed to develop an automated workflow to detect SDS and classify severity in soybean leaves. Hyperspectral reflectance images (398–1011 nm) were captured from both sides of 284 leaves, including SDS-symptomatic and asymptomatic samples. From these images, five distinct datasets were generated using key spectral bands, along with a pseudo-RGB dataset for model evaluation.
Results A You Only Look Once version 11 (YOLO11) model was trained and fine-tuned to distinguish asymptomatic leaves from symptomatic ones, achieving a mean average precision (mAP) of 0.97. To further classify regions of SDS-symptomatic leaves into categories (unaffected, mildly affected, severely affected, and dehydrated), both supervised and unsupervised methods were employed using reflectance and Normalized Difference Vegetation Index (NDVI) values. Among these methods, unsupervised k-means clustering using reflectance values demonstrated promising performance, achieving an overall accuracy of 82.69%. However, 24% of dehydrated pixels were misclassified as severe, 17% of mild pixels as unaffected, and 12% of unaffected pixels as mildly affected. Future work will seek to evaluate the progression of SDS throughout the growing season, which was a limitation of this study due to the late season data collection.
Conclusions The findings from this study could help farmers accurately assess SDS severity for better crop management and support breeding programs in developing SDS-resistant varieties.