Introduction

Of the many pathogens disrupting healthy growth of brassica species in the valley is the pathogen commonly referred to as Blackleg. This fungal pathogen has been reported to nearly wipe out canola production in Europe, Australia, Canada and in more recent years has devastated the United States (West et al., 2001). In 2013 the pathogen was reported for the first time in the pacific northwest since the 1970’s (Agostini et al., 2013) and has since been reported in the Willamette Valley (Claassen, 2016). There are two known species of Blackleg, Leptosphaeria maculuns and Leptosphaeria biglobosa. These are not to be mistaken with the potato bacterial pathogen Pectobacterium atrosepticum, which is also referred to as Blackleg. While much of the crop failure in canola has been associated with Leptosphaeria maculuns, both species are found in the valley and seem to be of similar consequence to turnip.

Classification of cercospera leaf spot for instance has been accomplished applying a support vector machine model but utilized a hyperspectral camera with high spectral and spatial resolution (Rumpf et al., 2010). Because plant diseases can oftentimes be difficult to see even with the naked eye, researchers have struggled to successfully detect specific plant diseases as spatial resolution decreased. While this analysis focuses solely on detection at 1.5 meters, it is possible for detection of blackleg despite lowered spatial resolution as result of increased flying elevations.

Here we consider how spatial patterns from diseased leaves is related to ground truth disease ratings of turnip leaves based on spectral signatures. With the application of a support vector machine model, classification of diseased versus non-diseased tissue is expected to generate a predictive model. This model will be used to determine if single turnip leaves which are diseased and non-diseased are accurately categorized based on the ground truth.

Data

This project will be using a data set derived from roughly 500 turnip leaves, harvested here in the valley during the months of February to April of 2019. Roughly 200 of these leaves will be used in training the data model. The remaining leaves will be used for the data analysis in determining accuracy of the model versus ground truth. The spatial resolution is less than 1 mm and images are in raster format of single turnip leaves with five bands (blue, green, red, far-red, NIR). I do not anticipate using every band for the analysis and will likely apply some VI as a variable. The camera being used is a mica sense Red-edge which is 12-bit radiometric resolution.

Hypothesis

I hypothesize that spatial patterns of pixels based on image classification are related to manually classified spatial patterns of observed disease on turnip leaves because disease has a characteristic spectral signature of infection on the leaves.

Approaches

In order to accomplish this analysis, I will be using ArcGIS Pro where I have quite a bit of experience, but not particularly on this subject or type of analysis. The workflow for analysis will begin with image processing where I have little experience but don’t require expertise in this area. I hope to conduct the image processing in Pix4D where I will begin with image calibration based on the reflectance panel in each image. Followed by cropping down to simply the leaf under assessment. From here there may be some smoothing and enhancing the contrast of the image but is still undetermined.

Images will then be brought into ArcGIS Pro for conducting a spatial analysis. I intend to use spatial pattern analysis of manually classified disease versus unsupervised segmentation of the leaves for exercise 1. Next I plan on then using this information in spatial regression to improve image-based classification for exercise 2. For exercise 3 I intend to use the support vector model wizard, which will be used for training a model. This involves highlighting regions of diseased tissue and regions of non-diseased tissue to obtain a trained model when a sufficient number of pixels to create support vectors is reached. The x and y-axis for the model are yet to be determined but will likely be NIR and red-edge digital number values. Some alternatives are using different VI’s such as NDVI as explanatory variable. Turnip images which were never used for training the model will be used for analysis of the support vector machine’s ability to classify diseased or non-diseased regions and then leaves entirely. I anticipate every leaf to have at least a few pixels which will classify as diseased and will therefore set a threshold for a maximum number of diseased pixels in the image, while yet classifying it as non-diseased. I also might require a certain number of pixels to be bordering one another qualify as diseased region. The methodology may require some troubleshooting, but the expectations are clear and the methods to reach that outcome are mostly drawn out.

Expected Outcome

I expect the model to have very high accuracy after the model is fine tuned for accuracy based on contrast in spectral signatures I expect to see between diseased versus non-diseased leaves. Below I have outlined the three outcomes I would like to ultimately achieve. Due to time restrictions, the scope of my research is limited to outcomes 1 & 2 below.

1. Train a support vector machine model for classification of pixels in turnip leaves as either diseased or non-diseased.
2. Accurately apply the SVM model on turnip leaves from many geographical locations in the valley with different levels of diseases severity and different times in the year.
3. Scale up from 1.5 meters and test the ability of the model to maintain accurate classification of blackleg on turnip.

Significance

I intend to publish and further the collective knowledge in aerial remote sensing. This more specifically applies to those in the area of agronomy or plant pathology. This is very applied science and is a resource for those in the industry of agriculture.

Traditionally detection for this pathogen has depended on a reliable field scout who may need to cover fifty acres or more looking for signs or symptoms of this disease. Nowadays, precision agriculture has introduced the use of drones to perform unbiased field scouting for the grower. This saves time and can be very reliable if done properly. An important aspect of disease control relies on early detection. If early detection can be accomplished, growers have time to respond accordingly. This may allow for early sprays with lighter applications rates or less controlled substances, cultural control of nearby fields, etc. in order to stop the spread of disease.

Works cited

Agostini, A., Johnson, D. A., Hulbert, S., Demoz, B., Fernando, W. G. D., & Paulitz, T. (2013). First report of blackleg caused by Leptosphaeria maculans on canola in Idaho. Plant disease, 97(6), 842-842.

Claassen, B. J. (2016). Investigations of Black Leg and Light Leaf Spot on Brassicaceae Hosts in Oregon.

Rumpf, T., Mahlein, A. K., Steiner, U., Oerke, E. C., Dehne, H. W., & Plümer, L. (2010). Early detection and classification of plant diseases with Support Vector Machines based on hyperspectral reflectance. Computers and Electronics in Agriculture, 74(1), 91-99.

West, J. S., Kharbanda, P. D., Barbetti, M. J., & Fitt, B. D. (2001). Epidemiology and management of Leptosphaeria maculans (phoma stem canker) on oilseed rape in Australia, Canada and Europe. Plant pathology, 50(1), 10-27.