Abstract:
Maize yields worldwide are limited by foliar diseases that could be fungal, oomycete,
bacterial, or viral in origin. Correct disease identification is critical for farmers to apply the correct
control measures, such as fungicide sprays. Deep learning has the potential for automated disease
classification from images of leaf symptoms. We aimed to develop a classifier to identify gray leaf
spot (GLS) disease of maize in field images where mixed diseases were present (18,656 images
after augmentation). In this study, we compare deep learning models trained on mixed disease
field images with and without background subtraction. Performance was compared with models
trained on PlantVillage images with single diseases and uniform backgrounds. First, we developed a
modified VGG16 network referred to as “GLS_net” to perform binary classification of GLS, which
achieved a 73.4% accuracy. Second, we used MaskRCNN to dynamically segment leaves from
backgrounds in combination with GLS_net to identify GLS, resulting in a 72.6% accuracy. Models
trained on PlantVillage images were 94.1% accurate at GLS classification with the PlantVillage testing
set but performed poorly with the field image dataset (55.1% accuracy). In contrast, the GLS_net
model was 78% accurate on the PlantVillage testing set. We conclude that deep learning models
trained with realistic mixed disease field data obtain superior degrees of generalizability and external
validity when compared to models trained using idealized datasets.