Distinguishing Cyst Nematode Species Using Image Textures and Artificial Neural Networks

Abstract

Cyst nematodes are plant parasitic nematodes infecting crops, causing extensive crop damage and annual losses, and affecting food production. The precise species identification is significant to initiate their control. The repeatable, less expensive, and less laborious distinguishing cyst nematode species using image processing and artificial intelligence can be advantageous. The objective of this study was to distinguish cyst nematodes belonging to the species Globodera pallida, Globodera rostochiensis, and Heterodera schachtii based on image parameters using artificial neural networks (ANN). The application of parameters selected from a set of 2172 textures of images in color channels L, a, b, X, Y, Z, R, G, B, V, U, and S to build classification models using a narrow neural network, medium neural network, wide neural network, trilayered neural network, WiSARD, multilayer perceptron, and RBF network is a great novelty of the present study. Algorithms allowed for distinguishing cyst nematode species with an average accuracy reaching 89.67% for a model developed using WiSARD. The highest correctness was obtained for H. schachtii and this species was distinguished from each other with the highest accuracy of 95–98% depending on the classifier. Whereas the highest number of misclassified cases occurred between G. pallida, G. rostochiensis belonging to the same genus Globodera. The developed procedure involving image parameters and artificial neural networks can be useful for non-destructive and objective distinguishing cyst nematode species.

1. Introduction

Plant parasitic nematodes are significant pests due to spreading disease and viruses and their feeding habits. Cyst nematodes (Heterodera spp. and Globodera spp.) are ranked on the list of plant parasitic nematodes. Potato cyst nematodes are one of the most damaging species. They can be found on all continents, in the southern tropical, tropical, and temperate zones, at sea level as well as higher altitudes. G. rostochiensis and G. pallida are well-known quarantine species [1]. Plant parasitic nematodes infect crops and thus affect food production. They cause extensive crop damage and annual losses. Cyst nematodes as a major group of plant parasitic nematodes, especially the most economically significant species within the Heterodera and Globodera genera, result in the greatest economic losses. These cysts are very persistent and can be viable for up to 20 years in the soil. This makes them the economically important group of plant parasitic nematodes [2]. Cyst nematodes, such as Globodera spp. and Heterodera spp., are obligate endoparasites of roots [3]. For example, G. rostochiensis and G. pallida are parasites of Solanaceae plants. They infect mainly potato, tomato, eggplant, and other Solanum spp. [4].

Cyst nematodes Globodera and Heterodera induce a syncytium, which is a feeding site in the roots of host plants and is composed of ultrastructurally modified root cells [5]. The syncytium is active and provides nutrients throughout the nematode development. The feeding relationships with the host plants are very specialized and complex [6]. The cyst nematodes use secreted protein effectors, which are encoded by parasitism genes to establish the interaction between the host and parasite [7]. Plant parasitic nematodes overcome the plant cell wall to invade hosts. They are willful and difficult to manage pests [8]. G. pallida and G. rostochiensis co-evolved with hosts in Bolivia, South America, and the Andean regions of Peru. Due to the ability to survive without a host for extended periods, they are very difficult to eradicate once established [9].

The precise nematode species identification and nematode number determination in soil samples are key points to initiate control of them [10]. Epidemiological monitoring and epidemic surveillance require the identification of the nematode species. The identification of nematodes to the species or genus can be performed by nematologists using morphobiometric techniques. Additionally, species can be identified using molecular testing or genome sequencing. However, in the case of emergent nematodes or new quarantine species, it is not always available. Morphobiometric techniques can be used alone or be associated with molecular tests for nematode species identification. The discrimination of nematode species using morphological characteristics through qualitative or quantitative analysis requires expertise and can be time-consuming for nematologists. The morphological identification of nematodes is often performed based on analysis of the cysts or second-stage juveniles. In the case of cysts, Granek’s ratio, the number of ridges between the anus and vulva, or the distance between the anus and vulva can be determined [1]. Due to the difficulties and time-consuming of cyst nematode identification using morphometrics and morphology, DNA-based techniques were applied. In the case of Globodera and Heterodera genera, molecular markers and genes were applied [4]. DNA-based methods are quicker than the traditional analysis using morphology and morphometrics for plant-parasitic nematode identification [11]. These methods are efficient for rapid, reliable, and precise identification of cyst nematodes. The application of PCR with species-specific primers was a significant step in the diagnostic technology development for sensitive species identification. In the case of H. schachtii, G. pallida, and G. rostochiensis, species-specific primers from sequence differences in the ITS (internal transcribed spacer) regions of rDNA were developed. Furthermore, real-time PCR with SYBR green I dye allows a faster detection simultaneously with quantification of target DNA [10]. The species identification can be carried out using a combination of molecular and morphological techniques [12].

G. pallida and H. schachtii are distributed in Asian and European countries and the Americas [10]. H. schachtii is the beet cyst nematode [13]. White potato nematode G. pallida (Stone, 1973) occurred in Poland as a very rare species, as a neighboring population of G. rostochiensis. Both species are on the EPPO Quarantine list [14]. The last note of G. pallida on the potato plantation in Central Poland in 1986, then deceased and several times after controlling shows the ban was effectively lifted (or: successfully extinguished) by Quarantine Officials from Bydgoszcz WSKIOR [15]. In 2010, in two batches of potato, tubers imported from Cyprus new population of G. pallida was found. Cyst nematodes can impair the growth and deform tubers of potatoes [16]. Till now PCR analysis distinguished these two cyst species parasitizing the potato most accurately [17]. A single nematode transcriptomic profiling with long-read sequencing can be used to reannotate the cyst nematode genome [18]. The analysis of the population genetic structure is characterized by the important role in the determination of the variability in the nematode–plant interaction [19].

In addition to the use of destructive methods to evaluate nematode cysts, objective and non-destructive evaluation can be performed using imaging. Computer vision is characterized by better repeatability, a formal quantification of errors of measurements, automation of species identification by computers, and elimination of the effort of human experts [1]. Due to image processing and pattern recognition, machine vision enables the quantitative analysis of qualitative criteria [20]. The application of image analysis can result in the determination of many textural and geometric parameters of images [21]. The image features can be analyzed using artificial intelligence. An approach combining image analysis and traditional machine learning or deep learning was successfully used in previous studies for classification [22,23,24,25].

This study was aimed at classifying cyst nematodes of Globodera pallida, Globodera rostochiensis, and Heterodera schachtii based on texture parameters using artificial neural networks (ANNs). The application of features selected from 2172 texture parameters of images in color channels L, a, b, X, Y, Z, R, G, B, V, U, and S to build classification models is a novelty of the present study. Due to involving the combination of image processing and artificial intelligence, the discrimination of cyst nematode species was performed in a non-destructive, rapid, effective, and objective manner.

2. Materials and Methods

2.1. Material

The research material consisted of cyst nematodes belonging to the species Globodera pallida, Globodera rostochiensis, and Heterodera schachtii. We sampled cyst nematodes from the cultures of the Nematology Laboratory in Gdynia and the Department of Botany, Institute of Biology, Warsaw University of Life Sciences—SGGW in Warsaw, Poland. Then, we subjected obtained cyst nematodes to imaging and distinguishing based on image textures using neural network models. The applied procedure is presented in Figure 1.

2.2. Methods

2.2.1. Image Acquisition and Processing

We imaged the cysts of G. pallida, G. rostochiensis, and H. schachtii using an Epson Perfection flatbed scanner (Epson, Suwa, Nagano, Japan). We performed color calibration and collected the images on a black background at a resolution of 1200 dpi and saved the acquired scans in TIFF format. Fifty cysts belonging to one nematode species were in one image. We acquired two images for each species. Then, we divided each image into fifty images including one cyst nematode each. The sample images are shown in Figure 2. In total, we obtained images of one hundred cysts of each nematode species. For a dataset including images of three hundred cysts belonging to three nematode species, we performed the image segmentation, ROI determination, and image texture computation separately for each image including one cyst nematode. Before image processing, we changed the file format of cyst images to BMP and processed the images using the MaZda software (Łódź University of Technology, Institute of Electronics, Łódź, Poland) [26,27,28] to segment the images, determine regions of interest (ROIs), and extract image features. We segmented the images into lighter cysts and a black background. Then, we separated cyst nematodes from the background and considered each cyst as one ROI. For each cyst, we computed 2172 textures of images in color channels L, a, b, X, Y, Z, R, G, B, V, U, and S based on the gradient map, histogram, run-length matrix, co-occurrence matrix, autoregressive model, and Haar wavelet transform.

2.2.2. Neural Network Models

We built neural network models using MATLAB (MathWorks, Inc., Natick, MA, USA) and WEKA (Machine Learning Group, University of Waikato, Hamilton, New Zealand) involving selected image textures. We performed the attribute selection using the Best First and CFS (Correlation-based Feature Selection) subset evaluator and selected thirty texture parameters so that the ratio of the number of image textures (30) to cases (300) was 1:10. The set of selected image parameters included 7 textures from color channel B, 7 textures from color channel Z, 4 textures from color channel U, 3 textures from color channel L, 2 textures from color channel G, 2 textures from color channel X, 2 textures from color channel Y, 1 texture from color channel b, 1 texture from color channel R, and 1 texture of images in color channel S (

Table 1. The selected image texture parameters used to build artificial neural network models to classify accuracies cyst nematode species.

Color Channel	Selected Image Textures
L	LHMean LHPerc90 LS5SZ3SumAverg
b	bS5SZ3SumAverg
X	XHMean XHPerc90
Y	YHMean YHDomn10
Z	ZHMean ZHVariance ZHPerc50 ZHPerc99 ZS5SH3SumVarnc ZS5SZ3SumAverg ZS4RVGLevNonU
R	RS5SV3DifVarnc
G	GHMean GHVariance
B	BHMean BHVariance BHPerc50 BHPerc90 BHDomn10 BS4RHGLevNonU BS4RVFraction
U	UHMaxm10 US5SN3AngScMom US5SN5SumAverg UATeta2
S	SS5SN5Contrast

). We performed the classification using a 10-fold cross-validation test mode applying a random division of the image texture dataset into ten parts and treating each part in turn as the test set, and the remaining nine parts of a dataset as the training sets. We performed the learning ten times on different training sets and estimated the overall error as the average of ten errors.

The neural network models with the specified parameters developed using MATLAB and WEKA are presented in

Table 2. The parameters of classifiers used to build models using MATLAB and WEKA.

Application	Classifier Type	Parameters
MATLAB	Narrow Neural Network	first layer size: 10; number of fully connected layers: 1; activation: ReLU; iteration limit: 1000; standardize data: yes; regularization strength (Lambda): 0
	Medium Neural Network	first layer size: 25; number of fully connected layers: 1; activation: ReLU; iteration limit: 1000; standardize data: yes; regularization strength (Lambda): 0
	Wide Neural Network	first layer size: 100; number of fully connected layers: 1; activation: ReLU; iteration limit: 1000; standardize data: yes; regularization strength (Lambda): 0
	Bilayered Neural Network	first layer size: 10; second layer size: 10; number of fully connected layers: 2; activation: ReLU; iteration limit: 1000; standardize data: yes; regularization strength (Lambda): 0
	Trilayered Neural Network	first layer size: 10; second layer size: 10; third layer size: 10; number of fully connected layers: 3; activation: ReLU; iteration limit: 1000; standardize data: yes; regularization strength (Lambda): 0
WEKA	WiSARD	batchSize: 100; bitNo: 8; bleachConfidence: 0.01; bleachFlag: False; bleachStep: 1.0; debug: False; doNotCheckCapabilities: False; mapType: RANDOM; seed: −1; ticNo: 256
	Multilayer Perceptron	batchSize: 100; debug: False; doNotCheckCapabilities: False; decay: hiddenLayers: a; learningRate: 0.3; momenetum: 0.2; nominalToBinaryFilter:True; normalizeAttributes: True; normalizeNumericClass: True; reset: True; resume: False; seed: 0; trainingTime: 500; validationTreshold: 20
	RBF (Radial basis function) Network	batchSize: 100; clusteringSeed: 1; debug: False; doNotCheckCapabilities: False; maxIts: −1; minStdDev: 0.1; numClusters: 2; ridge: 1.0 × 10−8

. We chose the classifiers providing the highest accuracies to be presented in this paper. The classifier types of narrow neural network, medium neural network, and wide neural network used in MATLAB depended on the first layer size. The classifier types of bilayered neural network and trilayered neural network depended on the number of layers. The exemplary neural network architecture is presented in Figure 3. For models built using MATLAB, we determined the confusion matrices, average accuracies, true positive rate (TPR), false positive rate (FPR), ROC (receiver operating characteristic) curves, and AUC (area under the ROC curve). In the case of models developed using WEKA [29,30,31], we computed the confusion matrices, overall accuracies, Kappa statistic, and TPR (true positive rate), FPR (false positive rate), ROC area (receiver operating characteristic area), PRC area (precision-recall area), precision, F-Measure, MCC (Matthews correlation coefficient) [32,33,34] for individual classes and weighted averages. We used the following Equations (1)–(9):

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N})}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}}$$

(1)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\mathrm{T}\mathrm{P}\mathrm{R}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(2)

$$\mathrm{F}\mathrm{P}\mathrm{R}=\frac{\mathrm{F}\mathrm{P}}{\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}}$$

(3)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$

(4)

$$\mathrm{R}\mathrm{O}\mathrm{C}\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}=\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{u}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{T}\mathrm{P}\mathrm{R}\mathrm{v}\mathrm{s}.\mathrm{F}\mathrm{P}\mathrm{R}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(5)

$$\mathrm{P}\mathrm{R}\mathrm{C}\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}=\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{u}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{s}.\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(6)

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}=\frac{\frac{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}+\frac{\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}$$

(7)

$$\mathrm{F}-\mathrm{Measure}=\frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{(\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l})}$$

(8)

$$\mathrm{M}\mathrm{C}\mathrm{C}=\frac{\left(\mathrm{T}\mathrm{P}\times \mathrm{T}\mathrm{N}-\mathrm{F}\mathrm{P}\times \mathrm{F}\mathrm{N}\right)}{\sqrt{\left(\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)\right)}}$$

(9)

where TP: true positive; TN: true negative; FP: false positive; and FN: false negative.

3. Results and Discussion

We classified the cysts of nematodes belonging to Globodera pallida, Globodera rostochiensis, and Heterodera schachtii using artificial neural network models built based on selected image textures. We developed models using various algorithms. The MATLAB software allowed for building quite successful models using a narrow neural network, medium neural network, wide neural network, bilayered neural network, and trilayered neural network (

Table 3. The classification accuracies of cyst nematode species using artificial neural network models built based on selected image textures.

Algorithm	True Class	Predicted Class (%)			Average Accuracy (%)
		G. pallida	G. rostochiensis	H. schachtii
Narrow Neural Network	G. pallida	81	19	0	86.3
	G. rostochiensis	17	80	3
	H. schachtii	1	1	98
Medium Neural Network	G. pallida	85	15	0	86.3
	G. rostochiensis	20	79	1
	H. schachtii	1	4	95
Wide Neural Network	G. pallida	81	17	2	86.3
	G. rostochiensis	15	83	2
	H. schachtii	0	5	95
Bilayered Neural Network	G. pallida	72	25	3	83.7
	G. rostochiensis	18	81	1
	H. schachtii	0	2	98
Trilayered Neural Network	G. pallida	79	19	2	83.7
	G. rostochiensis	23	75	2
	H. schachtii	2	1	97

). The average accuracies ranged from 83.7% (bilayered neural network, trilayered neural network) to 86.3% (narrow neural network, medium neural network, wide neural network). The cysts H. schachtii were classified with the highest correctness of up to 98% for narrow neural network and bilayered neural network. This meant that the cysts of this species were the most different from the other species in terms of image texture parameters. The greatest number of misclassified cases was between G. pallida and G. rostochiensis. This meant that these cysts were the most similar to each other. In the case of bilayered neural network, G. pallida cysts were distinguished with an accuracy of 72%, and as many as 25% of cases were misclassified as G. rostochiensis, and only 3% of cases as H. schachtii. Whereas 81% of cysts belonging to the true class G. rostochiensis were correctly classified as G. rostochiensis, 18% of cases were incorrectly included in the predicted class G. pallida and 1% in the class H. schachtii. For the trilayered neural network, G. pallida and G. rostochiensis were correctly classified with accuracies of 79% and 75%, respectively. As many as 19% of G. pallida cysts were misclassified as G. rostochiensis, and 2% as H. schachtii. A total of 23% of G. rostochiensis cysts were incorrectly classified as G. pallida and 2% as H. schachtii.

The ROC curves for the classification of cyst nematode species based on selected texture parameters of images confirmed the highest correctness of distinguishing H. schachtii cysts for all models (Figure 4). The values of AUC (area under the ROC curve) for H. schachtii were the highest reaching 0.9911 in the case of narrow neural network.

The next models were built using WiSARD, multilayer perceptron, and RBF Network (

Table 4. The accuracies of distinguishing cyst nematode species based on selected image texture parameters using algorithms from artificial neural networks.

Algorithm	True Class	Predicted Class (%)			Average Accuracy (%)
		G. pallida	G. rostochiensis	H. schachtii
WiSARD	G. pallida	87	13	0	89.67
	G. rostochiensis	14	85	1
	H. schachtii	0	3	97
Multilayer Perceptron	G. pallida	75	25	0	84.67
	G. rostochiensis	19	81	0
	H. schachtii	1	1	98
RBF Network	G. pallida	86	14	0	86.00
	G. rostochiensis	24	74	2
	H. schachtii	0	2	98

). WiSARD neural network model from the group of functions classified cyst nematode species of G. pallida, G. rostochiensis, and H. schachtii with the highest average accuracy of 89.67%, and the accuracies for individual species were 87%, 85%, and 97%, respectively. It was found that G. pallida and H. schachtii were completely correctly from each other and no cases were misclassified between these species. Whereas the greatest misclassification of cases between G. pallida and G. rostochiensis was observed. Similar results were found for a model developed using the RBF Network. No cases were mixed between G. pallida and H. schachtii and the highest number of misclassified cases were between G. pallida and G. rostochiensis. RBF Network distinguished cyst nematode species with an average accuracy of 86.00%, and accuracies of 86%, 74%, and 98% for G. pallida, G. rostochiensis, and H. schachtii, respectively. A model developed using multilayer perceptron was characterized by the lowest average accuracy equal to 84.67%.

A total number of correctly and incorrectly classified cases, as well as the Kappa statistic, and weighted averages of TPR (true positive rate), FPR (false positive rate), ROC Area (receiver operating characteristic area), PRC Area (precision-recall Area), Precision, F-Measure, and MCC for all three cyst nematode species are presented in

Table 5. The results of the classification of all cyst nematode species of G. pallida, G. rostochiensis, and H. schachtii based on models built using image texture parameters.

Algorithm	Correctly Classified Cases	Incorrectly Classified Cases	Kappa Statistic	Weighted Averages
				TPR	FPR	ROC Area	PRC Area	Precision	F-Measure	MCC
WiSARD	269	31	0.845	0.897	0.052	0.918	0.833	0.898	0.897	0.845
Multilayer Perceptron	254	46	0.770	0.847	0.077	0.943	0.889	0.849	0.847	0.771
RBF Network	258	42	0.790	0.860	0.070	0.942	0.870	0.861	0.859	0.791

TPR—true positive rate, FPR—false positive rate, ROC Area—receiver operating characteristic area, PRC Area—precision-recall area, MCC—Matthews correlation coefficient.

. It was found that WiSARD correctly classified the highest number of cyst nematodes, 267 out of 300. Also, the highest values were observed for other metrics. The Kappa statistic of 0.845, TPR of 0.897, FPR of 0.052, ROC Area of 0.918, PRC Area of 0.833, Precision of 0.898, F-Measure of 0.897, and MCC of 0.845 were determined.

The classification performance metrics for individual cyst nematode species of G. pallida, G. rostochiensis, and H. schachtii are presented in

Table 6. The performance metrics of the classification of individual cyst nematode species based on selected image textures using algorithms from artificial neural networks.

Algorithm	Class	TPR	FPR	ROC Area	PRC Area	Precision	F-Measure	MCC
WiSARD	G. pallida	0.870	0.070	0.901	0.810	0.861	0.866	0.798
	G. rostochiensis	0.850	0.080	0.863	0.702	0.842	0.846	0.768
	H. schachtii	0.970	0.005	0.991	0.989	0.990	0.980	0.970
Multilayer Perceptron	G. pallida	0.750	0.100	0.928	0.799	0.789	0.769	0.659
	G. rostochiensis	0.810	0.130	0.913	0.880	0.757	0.783	0.669
	H. schachtii	0.980	0.000	0.987	0.989	1.000	0.990	0.985
RBF Network	G. pallida	0.860	0.120	0.932	0.800	0.782	0.819	0.724
	G. rostochiensis	0.740	0.080	0.911	0.834	0.822	0.779	0.679
	H. schachtii	0.980	0.010	0.982	0.976	0.980	0.980	0.970

TPR—true positive rate, FPR—false positive rate, ROC Area—receiver operating characteristic area, PRC Area—precision-recall area, MCC—Matthews correlation coefficient.

. For all models, cyst nematodes H. schachtii were characterized by the highest values of TPR, ROC Area, PRC Area, Precision, F-Measure, and MCC and the lowest FPR. It indicated the most correct distinguishing H. schachtii cyst nematodes from other species. For H. schachtii, TPR reached 0.980 (multilayer perceptron, RBF Network), ROC Area—0.991 (WiSARD), PRC Area—0.989 (WiSARD, multilayer perceptron), Precision—1.000 (multilayer perceptron), F-Measure—0.990 (multilayer perceptron), and MCC—0.985 (multilayer perceptron). The value of FPR was the lowest (0.000) for a model developed using multilayer perceptron.

In the next step of the analysis, algorithms providing the highest accuracies of distinguishing all three cyst nematode species were used to build further classification models. Firstly, cyst nematodes belonging to Heterodera and Globodera genera were distinguished (

Table 7. The results of the classification of cyst nematodes of Heterodera and Globodera spp. based on models built using image texture parameters.

Algorithm	True Class	Predicted Class (%)		Average Accuracy (%)
		Heterodera spp.	Globodera spp.
WiSARD	Heterodera spp.	98	2	98.50
	Globodera spp.	1	99
Multilayer Perceptron	Heterodera spp.	98	2	99.00
	Globodera spp.	0	100
RBF Network	Heterodera spp.	98	2	98.50
	Globodera spp.	1	99

and

Table 8. The performance metrics of the classification of cyst nematode of Heterodera spp. and Globodera spp. using models developed based on image textures.

Algorithm	True Class	TPR	FPR	ROC Area	PRC Area	Precision	F-Measure	MCC	Kappa Statistic
WiSARD	Heterodera spp.	0.980	0.010	0.992	0.989	0.980	0.980	0.970	0.9700
	Globodera spp.	0.990	0.020	0.994	0.996	0.990	0.990	0.970
Multilayer Perceptron	Heterodera spp.	0.980	0.000	0.995	0.995	1.000	0.990	0.985	0.9849
	Globodera spp.	1.000	0.020	0.995	0.995	0.990	0.995	0.985
RBF Network	Heterodera spp.	0.980	0.010	0.992	0.989	0.980	0.980	0.970	0.9700
	Globodera spp.	0.990	0.020	0.994	0.996	0.990	0.990	0.970

). Then, models were developed to classify species of G. pallida and G. rostochiensis belonging to the genus Globodera (

Table 9. The distinguishing cyst nematode species of G. pallida and G.rostochiensis belonging to the genus Globodera.

Algorithm	True Class	Predicted Class (%)		Average Accuracy (%)
		G. pallida	G. rostochiensis
WiSARD	G. pallida	87	13	85.5
	G. rostochiensis	16	84
Multilayer Perceptron	G. pallida	83	17	81.5
	G. rostochiensis	20	80
RBF Network	G. pallida	82	18	76.5
	G. rostochiensis	29	71

and

Table 10. The performance metrics of distinguishing cyst nematode of G. pallida and G. rostochiensis based on models built using image texture parameters.

Algorithm	True Class	TPR	FPR	ROC Area	PRC Area	Precision	F-Measure	MCC	Kappa Statistic
WiSARD	G. pallida	0.870	0.160	0.807	0.800	0.845	0.857	0.710	0.7100
	G. rostochiensis	0.840	0.130	0.734	0.725	0.866	0.853	0.710
Multilayer Perceptron	G. pallida	0.830	0.200	0.869	0.839	0.806	0.818	0.630	0.6300
	G. rostochiensis	0.800	0.170	0.869	0.870	0.825	0.812	0.630
RBF Network	G. pallida	0.820	0.290	0.864	0.826	0.739	0.777	0.533	0.5300
	G. rostochiensis	0.710	0.180	0.864	0.856	0.798	0.751	0.533

). Cyst nematodes of Heterodera spp. and Globodera spp. were correctly classified in 99.00% for the model built using multilayer perceptron and 98.50% for WiSARD and RBF Network (Table 7). In the case of a model built using multilayer perceptron, the highest Kappa statistic of 0.9849 was found. The values reaching 1.000 were observed in the case of TPR for Globodera spp. and Precision for Heterodera spp. The lowest FPR of 0.000 was determined for Heterodera spp. (Table 8).

Cyst nematode species of G. pallida and G. rostochiensis were distinguished with an average accuracy of up to 85.5% for a model built using WiSARD. Cyst nematodes of G. pallida were correctly classified in 87% and G. rostochiensis in 84% (Table 9). In the case of WiSARD algorithm, the model was also characterized by the highest Kappa statistic of 0.7100, TPR of 0.870 and 0.840, Precision of 0.845 and 0.866, F-Measure of 0.857 and 0.853, MCC of 0.710 and 0.710, and the lowest FPR of 0.160 and 0.130 for G. pallida and G. rostochiensis, respectively (Table 10).

It was observed that in the case of each model, the highest accuracies were for H. schachtii, and the greatest misclassification was between G. pallida and G. rostochiensis. The reason for such classification results could be that species G. pallida and G. rostochiensis belong to the same genus Globodera and are more similar to each other. Whereas cysts H. schachtii belonging to a different genus Heterodera were more distinct in appearance and structure. G. pallida and G. rostochiensis are closely related and are morphometrically and morphologically similar. Therefore, these species are difficult to distinguish based on morphology. The cysts of both species are brown. However, cysts of G. pallida change during maturation from creamy white directly to brown, and G. rostochiensis from white to yellow and then brown [12].

Our study revealed that the combination of image processing and machine learning was useful for distinguishing cyst nematode species. The applied approach is innovative in distinguishing cyst nematode species. However, slightly different applications of machine learning and imaging are present in the literature. In the previous studies performed by Vlaar et al. [13], machine learning and metabolomics were used for the identification of a cyst nematode hatching factor. Very high Pearson’s correlation coefficients reaching 0.89 between metabolite features present in the root exudate, such as a compound of molecular weight 526.17 and solanoeclepin A and hitching, were observed [13]. Microscopic images combined with deep learning were applied to discern and count nematode eggs. The average detection accuracy of 97.00%, including algorithm count, human count, and error margin, was determined [35]. Whereas deep learning and computer vision were used for the successful identification of species of Globodera quarantine nematodes, such as G. pallida and G. rostochiensis. The accuracy of distinguishing both species reaching 0.88 was found [1]. Kranse et al. [36] reported that a cost-effective and easy-to-build imaging device to acquire images of the root system of a host plant infected with H. schachtii can replace costly microscopy equipment and the combination with machine learning can increase screening speed. The authors [36] found a correlation between automatic and manual counts of the area reaching 0.83 and 0.44 for the nematode females and nematode cysts, respectively. Furthermore, machine learning techniques and remote multispectral reflectance sensors were used to identify nematode damage on a soybean. The accuracy of the classification of asymptomatic and nematode-symptomatic soybean plants reached 0.71 [37]. Remote sensing was also used for the detection of plant stress induced by H. schachtii in sugar beet fields with a 100% discrimination between affected with beet cyst nematode (BCN) or Rhizoctonia crown and root rot (RCRR) and healthy plants [38] and by soybean cyst nematodes with the correlation of 0.58 between initial soybean cyst nematode population density and satellite image intensities [39]. Whereas Lu et al. [40] used fluorescence imaging to count soybean cyst nematodes and in comparison with the microscope counting, the imaging system was about 60% faster. Baretto et al. [41] applied hyperspectral imaging and machine learning to determine symptoms caused by Rizoctonia solani in sugar beet and the scoring of disease incidence was up to five times higher than the human visual rating. Our research is innovative compared to research reported in the available literature. The applied approach combining color image processing and artificial neural networks is a new direction of research in distinguishing cyst nematode species. The developed procedure can be used in further studies for the examination of other species of cyst nematodes. Furthermore, research can be expanded with the use of deep learning.

4. Conclusions

The obtained results revealed that the approach combining image analysis and artificial neural networks can be successful in distinguishing cyst nematode species. Cyst nematodes Globodera pallida, Globodera rostochiensis, and Heterodera schachtii were correctly classified with an average accuracy of up to 89.67%, and the highest correctness of 97% was found for H. schachtii. The most effective classification model was built using the WiSARD algorithm. The performed research is innovative. The application of texture parameters extracted from images in color channels L, a, b, X, Y, Z, R, G, B, V, U, and S to build classification models is a great novelty. Therefore, the present study set a new research direction in distinguishing cyst nematode species involving color image processing and artificial neural networks. Further studies can be expanded with other cyst nematode species and another field of artificial intelligence, such as deep learning.