Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. North Dakota is a major wheat production state in the U.S. Growers face a variety of disease challenges (e.g., tan spot and leaf rust) during the wheat production, which not only reduce grain yield but also lower their quality. Wheat growers usually walk into the field and visually evaluate wheat disease conditions. This conventional method is low in efficiency as growers need to walk through a large field, and less accurate because they generally use the results of several spots to estimate the condition of a large plot. In addition, the approach is objective because different inspectors may differ the evaluation results. Machine vision is a technology that has a potential to replace human inspectors to realize automatic evaluation by providing objective inspection results. This study focuses on recognizing and distinguishing wheat diseases – tan spot and leaf rust. Wheat crops were pot-planted in the greenhouses, and diseases were inoculated at proper stages. After inoculation and the appearance of disease, color images were collected with an RGB camera for successive 25 days. The dataset was manually created including 500 samples for tan spot diseased, 200 samples for tan spot control, 500 sample for leaf rust diseased, and 200 samples for leaf rust control. Classification algorithms (i.e., SVM, NN, GoogLeNet, and VGG16) were used to recognize the diseased leaves as well as distinguishing them. Prediction accuracies of different algorithms were reported and compared. For leaf rust diseased vs leaf rust control, the SVM and NN had accuracies of 86% and 85%, respectively. For tan spot diseased vs tan spot control, the detection accuracies for SVM and NN were 83% and 84%, respectively. For rust diseased and rust control, the GoogLeNet model accuracy was 93% and VGG-16 model accuracy was 92%. For the tan spot diseased vs tan spot control, the GoogLeNet accuracy was 95% and 98% for VGG-16. Based on the model accuracies, VGG16 (98%) performed more satisfactorily compared to other models, and it is therefore recommended to apply VGG16 to detect wheat diseases.