Optimizing Corn Tar Spot Measurement: A Deep Learning Approach Using Red-Green-Blue Imaging and the Stromata Contour...

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision

Background:

• Accurate quantification of tar spot disease in corn relies on visual detection of stromata.
• Existing methods for stromata detection face limitations in accuracy and parameter optimization.
• Early-season disease monitoring is crucial for effective tar spot management.

Purpose of the Study:

• To develop and validate an improved algorithm (SCDA v2) for automated detection of tar spot stromata on corn leaves.
• To address the limitations of SCDA v1, particularly the need for empirical parameter searching.
• To achieve higher and more consistent accuracy in stromata detection compared to previous methods.

Main Methods:

• SCDA v2 combines SCDA v1 for region proposal generation with a pretrained convolutional neural network (CNN) classifier.
• The algorithm was tested on Red-Green-Blue (RGB) images of corn leaves from field and glasshouse conditions.
• Performance was evaluated using accuracy metrics including F1 score, linear regression, and Lin's concordance correlation against human annotations.

Main Results:

• SCDA v2 demonstrated significantly higher agreement with reference data than SCDA v1.
• The mean Dice value (overall accuracy) for SCDA v2 was 73.7%, compared to 30.8% for SCDA v1.
• The CNN classifier in SCDA v2 effectively reduced overestimation, enhancing the F1 score.

Conclusions:

• SCDA v2 offers a robust and accurate solution for automated tar spot stromata detection.
• The algorithm shows significant potential for large-scale applications in crop surveillance and phenotyping.
• This advancement can aid in more timely and effective management strategies for tar spot disease in corn.