Euphytica Генетика

Automated detection and quantification of southern corn leaf blight severity in popcorn maize using computer vision

dos Santos Pereira, Hércules, Vivas, Marcelo, Pequeno de Souza, Yure, Cesar Gradice Saluci, Julio, Rangel de Souza, Ana Lúcia, Pereira Pinto, Luciana, Roberto Francisco, Felipe, Grover, Kashish, Gupta, Kajal, Fritsche-Neto, Roberto
Springer Nature · 2026-11-19 · DOI 10.1007/s10681-026-03831-5

Аннотация

Southern corn leaf blight (SCLB), caused by Bipolaris maydis , is one of the major fungal diseases responsible for significant losses in popcorn maize ( Zea mays L. var. everta ). Accurate assessment of disease severity is essential for selecting resistant genotypes in breeding programs. In this context, convolutional neural networks have emerged as a promising alternative, reducing analysis time and subjectivity compared with conventional visual assessment. Therefore, this study aimed to develop a model based on the YOLOv11l architecture to estimate SCLB severity in popcorn maize. A total of 67 genotypes were evaluated under controlled greenhouse conditions, resulting in a dataset of 1093 images. The model achieved mAP50 values of 1.00 and 0.99 for detection and segmentation of the Leaf class, respectively, and 0.78 and 0.74 for the Disease class, with an IoU of 0.64 and a Dice coefficient of 0.78. The estimates showed a strong correlation with visual assessment (r = 0.79–0.82) and Lin’s concordance correlation coefficient ranging from 0.62 to 0.82. Image-based phenotyping showed higher H2 than visual assessment, and the assessment method influenced the ranking of genotypes for resistance. Relative selection efficiency ranged from 88 to 110%. The model proved effective for quantifying SCLB severity and showed potential to support the selection of resistant genotypes. Graphical abstract A three-stage workflow assesses disease severity in popcorn maize. First, 67 genotypes, with four replicates and two plants per pot, are inoculated with Bipolaris maydis and imaged in a phenotyping box. Leaf, ruler, and disease masks are annotated; the dataset is split into 70 percent training, 15 percent validation, and 15 percent testing. A YOLOv11-seg model is trained for 200 epochs at 1,024 pixels. Predictions include detection and segmentation, with severity calculated as lesion area in pixels divided by total leaf area in pixels. The example prediction reports 15 percent severity. Evaluation includes a confusion matrix, mean average precision, recall, precision, and F1 score.

Читать у издателя →