Deep learning architectures for semantic segmentation and automatic estimation of severity of foliar symptoms caused by diseases or pests

Colour-thresholding digital imaging methods are generally accurate for measuring the percentage of foliar area affected by disease or pests (severity), but they perform poorly when scene illumination and background not uniform. In this study, six convolutional neural network (CNN) architectures were trained semantic segmentation in images individual leaves exhibiting necrotic lesions and/or yellowing, caused insect pest coffee leaf miner (CLM), two fungal diseases: soybean rust (SBR) wheat tan spot (WTS). All manually annotated three classes: (B), healthy (H) injured (I). Precision, recall, Intersection over Union (IoU) metrics test image set highest B, followed H I classes, regardless architecture. When pixel-level predictions used to calculate percent severity, Feature Pyramid Network (FPN), Unet DeepLabv3+ (Xception) performed best among architectures: concordance coefficients greater than 0.95, 0.96 0.98 CLM, SBR WTS datasets, respectively, confronting with severity. The other tended misclassify pixels as injured, leading overestimation Results highlight value a CNN-based automatic method determine severity on diseases obtained under challenging conditions brightness background. accuracy levels estimated FPN, DeepLabv3 + similar those standard commercial software, which requires adjustment parameters removal complex images, tasks that slow down process.