Using Image Analysis and Regression Modeling to Develop a Diagnostic Tool for Peanut Foliar Symptoms

Abstract

Peanut foliar diseases and disorders can be difficult to rapidly diagnose with little experience because some abiotic and biotic symptoms present similar symptoms. Developing algorithms for automated identification of peanut foliar diseases and disorders could potentially provide a quick, affordable, and easy method for diagnosing peanut symptoms. To examine this, images of peanut leaves were captured from various angles, distances, and lighting conditions using various cameras. Color space data from all images was subsequently extracted and subjected to logistic regression. Separate algorithms were developed for each symptom to include healthy, hopperburn, late leaf spot, Provost injury, tomato spotted wilt, paraquat injury, or surfactant injury. The majority of these symptoms are not included within currently available disease identification mobile apps. All of the algorithms developed for peanut foliar diagnostics were ≥ 86% accurate. These diagnostic algorithms have the potential to be a valuable tool for growers if made available via a web-accessible platform, which is the next step of this work.

1. Introduction

Peanut is an important food crop in the U.S. where annual production is valued greater than USD 1.1 billion [1]. Numerous pathogens have the ability to cause yield-reducing disease, with disease often considered the most limiting factor with respect to peanut production [2]. In South Carolina, late leaf spot and tomato spotted wilt are among the most economically important diseases causing predominantly foliar symptoms [3]. Several abiotic incitants cause foliar symptoms in peanut including nutrient deficiency and pesticide injury, some of which can resemble those caused by diseases. Diagnostics of foliar symptoms can be time consuming when visual assessments are conducted manually or when lab-based molecular methods are used [4]. Early and accurate diagnosis of peanut diseases or disorders are vital for reducing economic losses through unnecessary purchases, wasted time and labor, and yield losses resulting from suboptimal management. Furthermore, misdiagnosis can contribute to the development of fungicide resistance among populations of pathogenic fungi through additional exposure to non-target fungicide applications [5]. Although reference resources are available to assist growers in the manual identification of peanut foliar symptoms (e.g., [2] and https://images.bugwood.org/, accessed on 8 April 2022), the availability of an economically accessible and easy-to-use tool to aid in diagnostics would be beneficial when making management decisions.

Imaging analysis and processing techniques for assessing plant phenotypes have previously been developed and used by plant scientists and researchers for classifying phenotypes ranging from basic morphology to quantifying plant disease severity [6,7,8]. The general process is image collection, image analysis (or processing), and then model development. For image acquisition, there are several types of images that have been used for detecting biotic or abiotic symptoms in plants, such as visible spectrum (VIS) [8,9,10,11], fluorescence [12,13,14], infrared [15,16,17], normalized difference vegetation index (NDVI) [18,19,20], mulit-spectral [21,22], and hyperspectral [12,23]. However, all imaging methods beyond simple VIS require specialized cameras or equipment. The image processing step normally requires a software or program to analyze and extract color space data from the uploaded image(s). A few examples of software used to process symptomatic plant images include Assess (American Phytopathological Society, St. Paul, MN, USA) [24,25], Adobe Photoshop (Adobe Inc, San Jose, CA, USA) [22,26], and ImageJ [8,10,27]. An advantage to ImageJ over Assess and Adobe is that it is free and opensource; however, in this study we used Batch Load Image Processor v.1.1 (BLIP) software developed by Clemson University [28]. Although BLIP is not currently accessible to the public, it was preferrable over other processing programs due to our ability to customize the type of color space data extracted from images. Segmentation is a step during image analysis where image pixels are classified by color or texture to facilitate differentiating between diseased and healthy tissue [11,29]; this process can be manual or automated depending on the individual software used. Edge detection is another method of segmentation where adjacent pixels with abrupt changes in color space values are detected as an edge or boundary [29]. Machine learning techniques are very commonly used for the development of image-based models, such as artificial neural networks (ANN), support vector machines (SVM), random forest (RF), linear discriminant analysis (LDA), regression analysis, and convolutional neural network (CNN) [11,20,30,31]. Models for classifying disease symptoms have been developed for tomato [32,33], rice [34], wheat [35], avocado [20], lettuce [18], and citrus leaves [36], as well as potato tubers [9].

Image analysis using regression modeling for peanut foliar symptoms is one such means that can be used to automate the process of diagnosis to further facilitate information-based crop management decisions. Using image analysis and regression modeling for identifying (or classifying) peanut diseases based upon field-collected, smartphone images with native backgrounds would be a novel development and one that could be accessible with a greater degree of flexibility for end-users. The objective of this project was to develop algorithms for automated identification of peanut foliar symptoms as an initial step towards this effort.

2. Materials and Methods

2.1. Image Acquisition

Still images of peanut foliage were taken from various smartphones (e.g., Samsung Galaxy S5 and S10e, Google Pixel 3) and camera (Canon EOS Rebel T6i) types at various angles, light conditions, and distances from the canopy in a field setting (Figure 1). The size range of the images were 2397 to 15,872,256 pixels with approximately 74% of images used having ≤ 150,000 pixels. Images were sorted by symptomology for seven symptom types to include healthy, hopperburn, late leaf spot (caused by Nothopassalora personata ((Berk, and M. A. Curtis) S. A. Kahn and M. Kamal), Provost injury, tomato spotted wilt, paraquat injury, and surfactant injury (Figure 2). From each original (composite) image, individual leaflets were identified and cropped as a new image with the native background retained. Leaflets and canopy images from a range of varying presentation angles and shadow coverage were included.

2.2. Image Processing

Each image was manually rated on a binary scale for each of the seven foliar symptoms where 0 = nonevent (absent) and 1 = event (applicable symptom present) in a master database along with picture name, type (canopy or cropped), and location. Images were processed in BLIP, which extracts 544 variables of color space data for each image pixel, such as RGB, hue, Hope Color Index (HCI), and edge detection. The HCI was developed by manually recording more than 70 unique RGB values, which were grouped to generally characterize greens, yellows, browns, dark browns or black, and whites. These colors were recorded by manually assessing the images of healthy and symptomatic leaves and collecting the RGB from healthy areas, lesions, chlorotic areas, injury, etc., so that these colors were specific to asymptomatic and symptomatic tissue on peanut leaves. Colors were analyzed using K Means Clusters for more specific groupings and were added to the image processing software as additional data columns (

Table 1. Color clusters collected from healthy and symptomatic peanut images that were added to Batch Load Image Processor v.1.1 software as the Hope Color Index for analysis of the images used to develop the diagnostic algorithms.

HCI Color a	Red	Green	Blue	Corresponding Color
0	185	152	49
1	81	116	60
2	50	37	41
3	159	201	81
4	193	204	72
5	129	417	36
6	150	168	137
7	209	195	114
8	96	73	75
9	237	227	49
10	231	238	218
11	202	213	155
12	152	111	104

^a Final 13 distinct color groupings based on K Means Cluster analysis of over 70 unique colors (RGB) specific to peanut leaves.

). Manual binary ratings and the BLIP output variables for each image were compiled into a master database to be used for model development.

2.3. Data Analysis

The final master database size was 4208 peanut images. Models were developed for asymptomatic (or healthy) leaves and symptomatic leaves to include paraquat injury, hopperburn, late leaf spot, Provost injury, surfactant injury, and tomato spotted wilt for a total of seven individual models using the numerical color space values generated by BLIP. The data were subjected to logistic regression analysis (PROC LOGISTIC) in SAS 9.4. The best subset selection was used to reduce the number of parameters for each model. Logistic regression was selected as the classifier of choice due to our interest in statistically modeling the binary nature of the data and reported success by other researchers to develop similar diagnostic models using logistic regression [11,13].

Models were trained using 80% of the dataset (or 3367 images), and the remaining 20% (or 841) of the images were used for the validation. The probability of an event was calculated using the following equation:

Probability of symptom (event = 1) = exp(L)/(1 + exp(L))

The probabilities were on a scale of 0 to 1 for each image, with images assigned a probability ≥ 0.50 being classified as an event for each respective model. The models were assessed based on model accuracy, defined as the percentage of images that were correctly classified, sensitivity, defined as ‘true positive’ or the proportion of event correctly classified as an event (1), and specificity, defined as ‘true negative’ or the proportion of non-event images being correctly classified as a non-event (0), as well as receiver operating characteristic (ROC) curves, Akaike’s information criterion (AIC), and loess calibration plots.

With the goal to make the algorithms available to the public, and with the probability of there being low quality images analyzed, it was important to test the algorithms for accuracy on images with reduced pixel availability and various levels of brightness. The validation dataset was sorted by image size and classified into 4 pixel classes for each model, where: Class 1 = number of pixels < 50,000, Class 2 = 50,000 < number of pixels < 100,000, Class 3 = 100,000 < number of pixels < 150,000, and Class 4 = number of pixels > 150,000. In a separate analysis, the validation dataset was split into 4 classes of brightness where: Class 1 = brightness value < 150 (dark images), Class 2 = brightness value > 150 and < 190, Class 3 = brightness value > 190 and < 240, and Class 4 = brightness class > 240 (very bright images).

3. Results

3.1. Model Performance

The algorithms developed for peanut foliar diagnostics (Table S1) were accurate (89.4 to 98.7%), specific (94.9 to 99.3%), and sensitive (70.0 to 93.1%;

Table 2. Performance of paraquat injury, healthy, hopperburn, late leaf spot, Provost injury, surfactant injury, and tomato spotted wilt diagnostic models.

Model	Accuracy (%) a	Sensitivity (%) b	Specificity (%) c
Paraquat injury	98.7	86.6	99.3
Healthy	91.6	76.3	95.8
Hopperburn	96.3	70.0	98.6
Late leaf spot	89.4	71.7	94.9
Provost injury	97.8	93.1	98.7
Surfactant injury	94.3	91.3	96.1
Tomato spotted wilt	95.2	73.5	98.1

^a Accuracy is defined as the percentage of images that were correctly classified. ^b Sensitivity is defined as the proportion of events correctly classified as an event, or true positives. ^c Specificity is defined as the proportion of non-event images being correctly classified as a non-event, or true negatives.

). The area under the ROC curves for training and validation performance were > 93% for all models. Of the incorrect predictions for the LLS model (89.4% accuracy), only eight images were of Provost injury incorrectly predicted to be LLS. For the Provost injury model (97.8% accuracy), only two images were incorrectly classified as Provost injury that were actually LLS. None of the hopperburn images misclassified with the hopperburn model (96.3% accuracy) were incorrectly predicted to be paraquat injury; however, one image of hopperburn was incorrectly classified as paraquat injury when the dataset was analyzed using the paraquat injury model (98.7% accuracy).

3.2. Impact of Image Quality on Model Accuracy

For each symptomology, the majority of images in the validation dataset had ≤ 150,000 pixels (Classes 1–3), mostly single leaflet images that were cropped from larger canopy images (

Table 3. The total number of images, the number of event images and the accuracy of each pixel size class by model.

Model	Size Class a	Total Images	Event Images b	Accuracy (%) c
Paraquat injury	Class 1	377	4	94.2
	Class 2	164	11	95.7
	Class 3	99	11	85.9
	Class 4	201	20	96.0
Healthy	Class 1	377	169	86.7
	Class 2	164	14	82.3
	Class 3	99	6	72.7
	Class 4	201	8	82.1
Hopperburn	Class 1	377	26	95.2
	Class 2	164	6	84.1
	Class 3	99	9	72.7
	Class 4	201	23	83.1
Late leaf spot	Class 1	377	115	82.0
	Class 2	164	42	75.6
	Class 3	99	12	58.6
	Class 4	201	31	76.6
Provost injury	Class 1	377	36	95.2
	Class 2	164	28	82.9
	Class 3	99	19	65.7
	Class 4	201	49	92.5
Surfactant injury	Class 1	377	44	89.7
	Class 2	164	77	78.7
	Class 3	99	60	61.6
	Class 4	201	124	78.1
Tomato spotted wilt	Class 1	377	19	90.5
	Class 2	164	24	81.7
	Class 3	99	12	71.7
	Class 4	201	27	93.0

^a Class 1 includes images with > 50,000 pixels, Class 2 images have > 50,0000 and < 100,000 pixels, Class 3 images have > 100,000 and < 150,000 pixels, and Class 4 images have > 150,000 pixels. ^b Event images are defined as images that contain the respective symptomology for the model. ^c Accuracy is defined as the percentage of images that were correctly classified.

). For each model, Class 3 images were the least accurately categorized, ranging from 58.6% to 85.9%. Model accuracies ranged from 82.0% to 95.2% for Class 1 images, from 75.6% to 95.7% for Class 2 images, and from 76.6% to 96.0% for Class 4 images.

Overall, it was observed that models were less likely to accurately categorize images in brightness Classes 1 and 4 compared to brightness Classes 2 and 3 (

Table 4. The total number of images, the number of event images and the accuracy of each brightness class by model.

Model	Brightness Class a	Total Images	Event Images b	Accuracy (%) c
Paraquat injury	Class 1	125	1	84.0
	Class 2	259	4	92.7
	Class 3	352	31	96.6
	Class 4	105	10	83.8
Healthy	Class 1	125	16	72.8
	Class 2	259	66	83.8
	Class 3	352	87	87.5
	Class 4	105	28	66.7
Hopperburn	Class 1	125	9	74.4
	Class 2	259	25	83.8
	Class 3	352	27	90.3
	Class 4	105	3	76.2
Late leaf spot	Class 1	125	34	61.6
	Class 2	259	51	81.5
	Class 3	352	80	83.0
	Class 4	105	35	54.3
Provost injury	Class 1	125	42	72.8
	Class 2	259	35	93.1
	Class 3	352	47	92.6
	Class 4	105	8	87.6
Surfactant injury	Class 1	125	62	67.2
	Class 2	259	93	85.3
	Class 3	352	123	91.8
	Class 4	105	27	72.4
Tomato spotted wilt	Class 1	125	5	86.4
	Class 2	259	26	92.3
	Class 3	352	40	88.1
	Class 4	105	11	76.2

^a Class 1 includes images with brightness values < 150; Class 2 images have brightness values > 150 and < 190; Class 3 images have brightness values > 190 and < 240; and Class 4 images have brightness values > 240. ^b Event images are defined as images that contain the respective symptomology for the model. ^c Accuracy is defined as the percentage of images that were correctly classified.

). Accuracy of the models’ classification of images ranged from 61.6% to 86.4%, and from 54.3% to 87.6% for brightness Classes 1 and 4, respectively. For brightness Classes 2 and 3, accuracy ranged from 81.5% to 93.1% and from 83.0% to 96.6%, respectively, among the peanut symptomology models.

4. Discussion

The algorithms developed for peanut foliar symptom diagnostics using logistic regression analysis were accurate and acceptable when compared to similar diagnostic models developed using either traditional modeling or machine learning. Pérez-Bueno et al. [20] developed similar logistic regression algorithms for detecting white root rot of avocado trees using NDVI data extracted from images. Their reported logistic regression analysis model was 82% accurate, 79% sensitive, and 85% specific, and performed better when compared to other classifiers, ANN, linear discriminant analysis, and SVM models developed using the same images [20]. Oppenheim et al. [12] used a CNN to identify and differentiate between healthy tubers and symptomatic tubers for four diseases with an accuracy of approximately 96% with an 80/20 train-test split as used to develop our algorithms.

Olivoto et al. [10] used logistic regression to develop models for measuring foliar disease severity for six individual diseases. In their study, images with an average range from 84,480 to 972,774 pixels were used, and it was observed that image resolutions did not significantly affect the accuracy of their analysis; however, a slight reduction in accuracy was observed with images reduced to ≤ 45,900 pixels [10]. Steddom et al. [24] compared the effects of image resolution and formatting on the accuracy of their predictions for wheat foliar disease severity. They found that resized images with low resolution (858–337,000 pixels/image) had little to no effect on the accuracy when uploaded to Assess software for analysis. Retention of accurate classification with a large range of image resolutions represents a potential means of increasing the throughput for the final server-side framework, though the exact magnitude of time savings would depend on the specific framework involved. Overall, a reduction in accuracy (%) was not observed with low resolution images for these peanut foliar symptomology models. In fact, for the majority of the models, images in pixel class 1 (>50,000 pixels) were the most accurately (%) classified. This was most likely due to the fact that the majority of the image dataset were cropped leaflet images with low resolution. Image brightness did have some impact on model accuracy (%) where images with brightness values > 150 and < 240 were most likely to be classified correctly. Such information pertaining to image quality will be useful in developing the website (or app) for model availability where the user could be alerted that the image uploaded is outside the desired range and that there is a corresponding decrease in confidence in the result based on this data.

Though a range in accuracy was observed, the most accurate (%) was the algorithm of peanut leaves symptomatic for paraquat injury with the least accurate (%) being the algorithm for peanut leaves symptomatic of late leaf spot. This may have been contributed to by late leaf spot images varying greatly in the severity and color of symptoms (e.g., lesions can vary in shade and the irregular presence of a yellow halo around lesions), whereas paraquat injury symptoms were generally more uniform in severity and color. Regardless, models were able to accurately classify respective symptoms against a photo database of all foliar symptoms. Additionally, the individual models were accurate when used to ‘diagnose’ or differentiate symptoms that appear similar, such as the late leaf spot and Provost injury models and the hopperburn and paraquat injury models. With respect to image classification, each model was more specific than sensitive, with the sensitivity and specificity of these models having been comparable to those developed by Pérez-Bueno et al. [11]. Although the models for healthy, hopperburn, late leaf spot, and tomato spotted wilt, exhibited a greater chance for ‘false negative’ predictions (i.e., lower sensitivities circa 70 to 76% compared with 87 to 93%), the sensitivity of those four models’ predictions were functionally acceptable. This follows the availability of readily observable and characteristic features that can be paired with image collection. Hopperburn symptoms, in addition to being recognizable, often occur together first near field edges and, when feeding is active, are accompanied in the field by leafhoppers whose movement may be seen. Image identification of late leaf spot lesions may be supplemented with in-field examination for the presence of bumpy conidiophores on abaxial surfaces. Tomato spotted wilt infection, as a whole, can produce a wide variety of symptoms; when infections occur, it is rarely limited to an individual plant, which consequently facilitates multiple symptomatic plants from which to collect images in a given field for subsequent analysis. Lastly, the nature of a non-omniscient tool aimed at declaring or predicting “healthy” is limited by design to only being able to identify instances where there is a lack of previously identified problems that have been (explicitly or implicitly) incorporated into its development.

While the app Plantix offers some models for peanut symptom identification, it is only available for Android users and does not include models for paraquat injury, tomato spotted wilt (caused by Tomato spotted wilt virus), surfactant injury, nor Provost injury on peanut [37]. Therefore, there are currently no automated tools available to identify all of the peanut foliar symptoms presented in this paper, meaning that these diagnostic algorithms have the potential to be a valuable tool for growers if made available via a web-accessible platform. In anticipation of making these models accessible and in a user-friendly manner, images with reduced quality (≤150,000) relative to the average resolution of most devices (≥1 million pixels) and native backgrounds were used in order to eliminate the need for special equipment or for users to perform additional modification to the image prior to upload. Similar to how traditional machine learning techniques use segmentation to remove backgrounds or classify diseased areas, BLIP uses edge detection procedures to differentiate between plant and soil pixels as well as the HCI colors, which were greatly important in the development of the presented models.

The availability of a simple, quick, and accurate tool to correctly identify symptoms could aid producers in decision-making regarding the potential need for pesticide applications. Although diagnostics are a crucial part of the decision-making process, it is important for growers to consider all available information before finalizing management decisions. Factors such as cultivar, growth stage, incidence of symptoms throughout the field, and management decisions prior to diagnosis add invaluable context to help inform how individual situations should be addressed.

5. Conclusions

In this study, image analysis and regression modeling were successfully used in the development of peanut foliar symptom diagnostic algorithms that are not currently available, to our knowledge, via other disease identification tools on the market. The development of peanut foliar models using point and click, field-based images is a novel approach to automated diagnostics that would potentially be a great resource for growers wanting to quickly identify foliar symptoms found in their field. The accessibility of these models to process images would be an advancement in the current market of automated plant diagnostic applications and one that is currently being pursued.