A robust segmentation method combined with classification algorithms for field-based diagnosis of maize plant phytosanitary state

Thierry Tchokogoué; Auguste Vigny Noumsi; Marcellin Atemkeng; Michèle Flore Yimga Fonkou; Louis Aimé Fono

doi:10.1515/jisys-2023-0049

Article Open Access

A robust segmentation method combined with classification algorithms for field-based diagnosis of maize plant phytosanitary state

Thierry Tchokogoué , Auguste Vigny Noumsi , Marcellin Atemkeng , Michèle Flore Yimga Fonkou and Louis Aimé Fono

Published/Copyright: May 31, 2024

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Journal of Intelligent Systems Volume 33 Issue 1

Abstract

Early diagnosis of maize-plant phytosanitary state in the field is crucial to prevent crop damage and optimize yield. However, this field diagnosis presents a challenge due to the variable background of the field environment, which can hinder the performance of classification algorithms. In this article, we introduced a novel segmentation technique using a combined normalized difference vegetation index that effectively isolates the features of interest, such as the leaves, from the surrounding image, which includes the diverse field background. To assess the effectiveness of our segmentation approach, we conducted early diagnosis of maize plants in the field using supervised classification algorithms. We generated a dataset that incorporated four essential texture features: energy, entropy, contrast, and inverse. These features were extracted from each of the segmented images using grayscale co-occurrence matrices. We employed four different classification methods, namely Adaboost, Random Forest, K-Nearest Neighbors, and support vector machine. When combined with the proposed segmentation technique, the support vector machine outperformed the other models, achieving an accuracy rate of 97%.

Keywords: phytosanitary; diagnosis; maize; segmentation; Adaboost; random forest; k-nearest neighbors; support vector machine

1 Introduction

The rapid growth of the world population, leading to a significant increase in food requirements, coupled with a reduction in cultivable areas due to urbanization and residential development, underscores the need to develop techniques ensuring both the quality and quantity of agricultural production on available arable land. These techniques should also safeguard crops against various pests and diseases, as highlighted in the study by Hanssen and Lapidot [1]. The conventional approach involves the use of fertilizers to improve soil quality and pesticides to manage pests and diseases, which are significant contributors to decreasing agricultural yields [2]. However, it is essential to acknowledge that the excessive or routine application of these pesticides and fertilizers has its limitations. It leads to environmental pollution, poses health risks to those handling it, and strains the national economy.

A recent approach aims at early, automatic, and robust disease detection using artificial intelligence in cultivated fields to promote sustainable agriculture with enhanced yields. Artificial intelligence has emerged as a transformative force with vast potential across various fields of application. Its impact extends far beyond traditional boundaries, with notable contributions observed in the medical [3,4], transportation [5], ecological [6,7], and agricultural sectors [8,9], among others. Using artificial intelligence to automate plant disease detection can significantly reduce the reliance on fertilizers and pesticides, making it environmentally friendly, and optimize the use of available human resources in plantations. Modern monitoring methods necessitate the development of automatic recognition systems based on spectral image sensor data combined with algorithms to identify infected areas of a crop [10]. This challenge gives rise to the ambition to create more robust and accurate automatic recognition algorithms capable of running on promising fully embedded processing units. The results are expected to be employed in designing a digital model for the automatic diagnosis of the phytosanitary (indicate whether the leaves of the plant are healthy or infected) condition of food crop plots. This model will be carried on board a drone or an agricultural vehicle, flying over known areas to detect infected plants. Machine learning and deep learning methods are increasingly essential in the advancement of precision agriculture, enabling farmers to forgo the need for expert diagnosis and expediting the detection of infected areas in the field, significantly accelerating interventions.

In this context, we propose an automated diagnostic system for corn plant diseases in the fields and in Cameroon. Corn, one of the primary cereals consumed in Cameroon and the fourth staple crop after cassava, plantains, and taro [11], is regularly targeted by pests, with the most common being Spodoptera frugiperda ^[1]. More precisely, this system is designed for the automatic classification of real-world images of corn plants captured in multiple fields, involving a five-step process: image acquisition, preprocessing, segmentation, feature extraction, and classification. It is worth noting that, due to the absence of a database on diseased corn plants, the initial step of this work was to collect the data in the field. Image segmentation, a critical phase in the image processing pipeline, focuses on distinguishing the leaves from the rest of the image (which can be highly diverse) using a novel approach, namely, the combined normalized difference vegetation index (CNDVI) technique applied to color images.

The remainder of this article is organized as follows: Section 2 discusses related work; Section 3 presents the methodology with emphasis on image acquisition and preprocessing, the new proposed segmentation method, features extraction, and the four classification methods; Section 4 discusses the results; and Section 5 gives concluding remarks.

2 Related work and purpose of the article

In recent years, many researchers have been interested in remote sensing to perform phytosanitary diagnostics of plants using colorimetric and texture parameters combined with artificial intelligence. Let us summarize some of the obtained results. Salathé et al. [12] trained a neural network to identify plant diseases. They worked on 14 plant species and 26 diseases, and the system was able to identify diseases in plant images extracted from a dataset containing images photographed under lighting conditions, color, background, etc. Their methods have a limitation in that the images are not captured directly in the field. Instead, the leaves are first collected and taken to a different environment, where pictures are taken against a white background. Wäldchen et al. [13] presented a new convolutional neural network (CNN)-based rice disease identification method using a dataset of 500 natural images of healthy and diseased rice leaves achieving 90 % accuracy. Aït et al. [14] tested an early detection model of eight tomato diseases using a CNN classifier based on deep learning. With 2,000 scanned images (200 per disease for eight diseases and 400 healthy leaf images) obtained from the net Plant Village dataset, they achieved an accuracy of 94%. The authors in [14–20] used the images taken in open datasets (PlantVillage, crop diseases, etc.) containing images of corn, rice, and many other species to train their learning model, thus obtaining very good detection precision. However, these precisions deteriorate considerably when testing their models on other datasets or on images photographed under real field conditions. Other authors [13,21–23] used natural photographs to train their models and obtained weak points compared to the first category of authors. Seyed et al. [20] used several deep learning model architectures that were trained to achieve a success rate of 99.53% for the diagnosis of various diseases of different plants extracted from the open database PlantVillage of 87,848 scanned images (also captured with a monochrome background and uniform lighting conditions). They proposed a methodology using segmentation by K-means clustering while using gray level co-occurrence matrix (GLCM) for the calculation of second-order texture characteristics and a support vector machine (SVM) classifier. However, this methodology is applied to images taken in the laboratory under controlled lighting and background conditions. Table 1 briefly summarizes all these related works.

Table 1

Related works

Authors	Year	Classifier deep – machine learning	Application	Types of images	Accuracy	Limit
Salathé et al. [12]	2016	CNN (deep learning)	Detection of 26 diseases on 14 species of plants	54,306 Scans images	99.35%	Images taken on a plain background and under controlled lighting
Wäldchen et al. [13]	2017	CNN (deep learning)	Identification of rice diseases	500 Natural images	90%	Image taken zoom only on the parts of the sheet (low image definition)
Ferentinos [15]	2018	CNN (deep learning)	Detection of 58 diseases on 25 species of plants	87,848 Scans	99.53%	Images taken on a plain background and under controlled lighting
Aït et al. [14]	2020	CNN (deep learning)	Early detection of eight tomato diseases	2,000 Scans	94%	Images taken on a plain background and under controlled lighting
Sumita et al. [22]	2020	CNN avec ajustement des hyper paramètres (deep learning)	Maize disease detection	4,382 Natural and scanned images of plant village	88.66% with natural images	Image taken zoom only on the parts of the sheet (low image definition)
Yu et al. [17]	2021	K-means associated with a CNN (deep learning)	Diagnosis of three maize diseases	900 scanned images selected from Crop disease	93%	Images taken on a plain background and under controlled lighting
Deng et al. [21]	2021	CNN (deep learning)	Detection of 06 types of rice diseases	33,026 Naturals	91%	Image taken zoom only on the parts of the sheet (low image definition)
Pan et al. [23]	2022	CNN (deep learning)	Detection of northern corn leaf blight	985 Natural images increase to 30,655 images	99.94%	Natural but extensive image base
Prajapati et al. [25]	2017	SVM (machine learning)	Detection and classification of rice plant diseases	120 Semi-natural images	88.57%	Images taken on a white background
Mousavi et al. [18]	2018	Expert system using bayes theorem method (machine learning)	Maize disease detection	Scans images	90.04%	Images taken on a plain background and under controlled lighting
Goel and Nagpal [16]	2022	SVM (machine learning)	Classification of plant diseases	Scans images	99.98%	Images taken on a plain background and under controlled lighting
Pawan et al. [19]	2022	Expert system developed using the forward chaining method, which is a forward trace (machine learning)	Early detection of maize diseases	Database of scanned images	84%	Images taken on a plain background and under controlled lighting and low precision
Seyed et al. [20]	2022	SVM (machine learning)	Grape leaf disease detection and classification	3,885 Scanned images of plant village	98.71%	Images taken on a plain background and under controlled lighting

In this article, we propose a system trained with natural photographs for deployment in real fields, with a segmentation approach based on a CNDVI. This approach, which applies to color images, makes it possible to better separate the regions of interest in the photographed images despite the variations in light and the diversified background.

3 Methodology

We proposed a process to diagnose maize diseases in the field. The methodology is twofold: (1) the first phase consists of the acquisition of images in the field, followed by data pre-processing, segmentation of the images, and extraction of the relevant features; (2) The second phase applies four machine learning methods to classify the different states of maize diseases. This methodology is summarized in two phases and presented in Figure 1.

Figure 1

Image processing process: Constitution of dataset (top panel), determination of the best model (middle panel), and image diagnosis process after deployment of the model in the field (bottom panel).

3.1 Dataset: data collection in the field and preprocessing

The first step of the automatic diagnostic process is to acquire images of corn plants. Wäldchen and Mäder [24] suggested that the image could be of the whole plant, leaf, flower, stem, or even fruit. They also suggested that there are three categories of images based on the mode of image acquisition, namely “Scan,” “Pseudo-scan,” and “Photos.” In the scanning and pseudo-scanning categories, leaf images are captured through the scanning and photography process, respectively, meaning that images are taken in front of a uniform color background and under fixed lighting conditions. For the third category, images are plants captured in their natural environment (with variable backgrounds and lighting conditions). For this phase, we collected 130 photos of maize leaves photographed in their natural environment in two fields. In the first field located in the Douala 5 district (referred to as field A with GPS coordinates 4.100501–9.802453), we collected 44 images of leaves attacked by Spodoptera frugiperda and 31 images of healthy leaves. In the second field located in the Douala 3 district (referred to as field B with GPS coordinates 4.059465–9.817352), we collected 25 images of leaves attacked by Spodoptera frugiperda and 30 images of healthy leaves. In total, there are 69 images of leaves infected by Spodoptera frugiperda and 61 images of healthy leaves. These images, with an average size of 6 MB per image, were taken using a Samsung SM-A750FN camera with an aperture of F1.7, a focal length of 3.93 mm, and the flash disabled. Figure 2 shows one of the coauthors of this work in one of the fields taking the images.

Figure 2

Field work for data collection.

To facilitate the identification or location of each image in the image database, an encoding format was defined and applied to all images in the database. Table 2 provides a description of the different attributes of the coding format. For example, the code S-S2-CA-11.04.21-1 means the photo of a healthy corn leaf (S) at its second week of growth (S2) taken in field A (CA) on April 11, 2021 (11.04.21), and numbered 1.

Table 2

Description of each element contained in the image encoding format

Format components	Descriptions	Component codes
State	Indicates the condition of the leaves	S (healthy) M1 (Spodoptera frugiperda)
Maturity	Indicates the age of the plant	n , n = 1 , 2 , 3 , … number of weeks (S4)
Field	Indicates the image collection field	CA (Field A) CB, …
PickupDate	Indicates the date and month of the image capture	10.05.21 (May, 10, 2021)
Number	Indicates the order number of the image	1, 2, 3, …

Wäldchen and Mäder [24] and Wäldchen et al. [26] justified that various factors (such as visibility, brightness and contrast of gray-level pixels, and amount of noise) impact the quality of the image captured in natural environments under various lighting conditions. To obtain suitable images for processing from those in our database, we performed pre-processing to highlight certain interesting features of an image. To preprocess our images from our database captured in natural environments, we applied the Gaussian filter (a classic example of preprocessing that increases contrast and removes noise from the image for improvement). The principle of this preprocessing method involves convolving the raw input image I with a Gaussian kernel G to obtain a preprocessed or filtered image: Figure 3 shows two original images taken randomly from our image database, as well as the result obtained after applying the Gaussian filter to both images.

$Figure 3 Original and filtered pictures with σ = 2 \sigma =2 and σ = 5 \sigma =5 .$

Figure 3

Original and filtered pictures with σ = 2 and σ = 5 .

The images in our database, even after preprocessing, contain a diverse and random background. In order to better identify their contents using computer vision techniques, it is important to take into account the interaction between light and matter (green leaf for healthy plants or spotted green for infected leaves) to propose a new method of segmentation for pre-processed leaves. This segmentation principle is based on the theories of radiative transfer, which state that when a surface is illuminated by light, four reactions (reflection, refraction, absorption or transmission) dependent on the wavelength of the light and the composition of the surface can occur. The following section provides a detailed description of this new segmentation method.

3.2 Segmentation

3.2.1 Description of the method

To highlight the area of interest (healthy or diseased leaves) in each of the images taken in the natural environment, we propose a new segmentation method in this section called CNDVI.

The intuition behind CNDVI is based on the fact that any object illuminated by a natural or artificial light source has a particular way of reflecting light. The camera captures the light reflected by the object to create its image. More precisely, when applied to leaves, this intuition is based on the following hypothesis: “Vegetation strongly absorbs the blue and red wavelengths of visible light and reflects the green wavelength of that light.”

Figure 4

Segmentation algorithm using a combination of two vegetation indices I1 and I2 calculated from the RGB components of visible light.

It is worth noting that compared to CNDVI, the segmentation method, normalized difference vegetation index (NDVI), uses near-infrared and red wavelengths to assess a single vegetation index.

(1) I = NIR − R NIR + R ,

where NIR is the grayscale level of the near-infrared wavelength, and R is the grayscale level of the red wavelength (Kosal [27]). The NDVI is not suitable for images taken by conventional color cameras (which do not record the near-red bands) and, therefore, does not take into account the three radiations – green, red, and blue – reflected by the leaves and recorded by these cameras. To address this limitation and propose a method suitable to Sub-Saharan Africa, it is easier and less expensive, in terms of cost, availability, and accessibility, to have a color camera, like a phone, for taking images than to have radar; therefore, we propose the CNDVI, which uses only visible light (RGB) recorded by these cameras. This leads to the calculation of two vegetation indices. The advantages of this new method are (i) the fact that the indices are calculated solely based on the reflections of visible light, making them applicable to images taken by any standard camera, and (ii) the use of a combination of two vegetation indices calculated for each pixel of the image to achieve efficient segmentation. Note that R , G , and B are the gray levels of the red, green, and blue wavelengths, and the first vegetation index I 1 is defined as follows:

(2) I 1 = V − R V + R ,

a normalized difference between the red and green bands. The second index, I 2, is defined as follows:

(3) I 2 = V − B V + B .

A combination of these two indices ( I 1 and I 2 ), calculated for each pixel of an image, allows for a clear distinction between the pixels corresponding to the vegetation foliage and the background of the image. This leads to the classification of the image points (pixels) based on the following criteria:

C1: if
(4) I 1 ≈ 1 and I 2 ≈ 1 .
In other words,
(5) R ≈ 0 , B ≈ 0 , and V ≥ 0 ,
so the image point corresponds to healthy vegetation.
C2: if
(6) 0 ≤ I 1 ≤ 1 and 0 ≤ I 2 ≤ 1
when
(7) R ≥ 0 , B ≥ 0 , V ≥ R , and V ≥ B ,
then the image point corresponds to a leaf that is beginning to turn yellow or red due to infection. These two criteria suggest that an image point will be considered either healthy or infected when
(8) I 1 ≥ 0 and I 2 ≥ 0 .

However, these two criteria alone are not sufficient for optimal and complete segmentation because points characterized by nearly equal values of R , G , and B may appear. We propose a modification of the two classification criteria by introducing a threshold ϕ (we assume its value is between 0 and 0.2, depending on the lighting levels during image capture) beyond which a pixel of the image will be considered healthy. More formally, a point is considered healthy if

(9) I 1 ≥ ϕ and I 2 ≥ ϕ .

This criterion is introduced into the algorithm to separate the green foliage from the rest of the image, as shown in Figure 4. This CNDVI segmentation algorithm masks (sets to black) all the noise in the image (any object that is not foliage) and keeps the foliage pixels unchanged. To evaluate the performance of this new CNDVI segmentation method, manual segmentation was performed using Photoshop on four randomly selected images from the image database, which is used as the ground truth. Figure 5 shows a comparison between CNDVI and ground truth. The Vinet dissimilarity measure was used to show that the segmentation obtained from CNDVI is similar to the ground truth, thus confirming the effectiveness of the new method.

Figure 5

Columns 1 to 4 respectively present four natural images, then the renderings of their manual segmentations (ground truth), then the segmentation by CNDVI and finally the residuals between the two segmentations.

Figure 6

Infected plant showing a different texture from healthy plants.

3.2.2 Evaluation of CNDVI segmentation

The performance of the CNDVI segmentation algorithm is evaluated using the Vinet distance, which involves comparing the image segmented by CNDVI with the ground truth. This comparison counts the number of matching pixel pairs in both images. The Vinet distance is a measure of dissimilarity between two images segmented using different approaches, and it is computed as follows: Let X be the ground truth image with each X i a pixel in the image,

(10) X = ⋃ i = 1 n X i ,

and Y be the segmented image using CNDVI, where each Y i is a pixel in Y

(11) Y = ⋃ i = 1 m Y i .

The dissimilarity between segmentations X and Y is given as follows:

(12) Dvinet = N − ∑ i = 1 N C i N ,

where

(13) C i = card ( X i ∩ Y i ) , card = cardinality , i = 1 , ⋯ , n , j = 1 , ⋯ , m , N = min ( n , m ) .

The dissimilarity was calculated for four images from our dataset, and the results for each image, as well as the average of the four results, are recorded in Table 3.

Table 3

Results of dissimilarity between the two segmentations from each of the four images

Performance	M1S4CA2404-28	M1S5CA0205-15	M1S5CA0205-12	SS5CA0205-11	Average
Dvinet	0.0654	0.0492	0.0387	0.0789	0.05805

3.3 Features extraction

After the preprocessing and the segmentation of the image in the desired region of interest, the next stage of the work consists in extracting the characteristics of the segmented images, thus making it possible to reduce the complexity of the image to avoid the use of a large amount of computing power and memory. Therefore, feature extraction is a very important step in the image processing process because it helps to reduce the number of variables for further image processing (this is a dimension-reduction process). The extracted features can be, color, texture, grayscale, spectrum, or a combination of several elementary features. Careful observation of corn diseases (Figure 6) reveals a change in leaf texture when infected. Based on this observation, we chose second-order statistics that construct the GLCM and calculate 13 texture features. Of these features, we select four features (energy, entropy, contrast, and inverse moment of difference) as the most relevant to facilitate decision-making about the health status of corn plants.

Figure 7

Performance histogram for each classifier. (a) Precision of the four models. (b) Specificity of the four models. (c) AUC of the four models. (d) Sensibility of the four models.

3.4 Used classification methods

In this article, classification generally involves deciding whether a query image contains healthy or infected corn leaves based on the extracted feature set. The primary objective of this work was to introduce a novel segmentation approach for feature extraction. The features extracted through this segmentation process are employed to train traditional machine learning models. In contrast, deep learning models are trained directly on the data without the need for any preprocessing for feature extraction. This explains why we did not use deep learning models in this article. To achieve this, we opted for four classifiers, namely SVM, AdaBoost, k-nearest neighbor (KNN), and random forest, and applied them to our dataset. The performance of each model was assessed, and the best-performing model was selected. The performance of each algorithm is evaluated by the following four measures: accuracy (overall rate of good classification), specificity (rate of classification of healthy plants), sensitivity (rate of classification of infected plants), and area under the curve (AUC). All these measures are estimated from the confusion matrix generated by the model after the prediction phase. Note that, the most efficient model is the one whose AUC is closer to 1, and thus, specificity, sensitivity, and accuracy should be higher. However, since it is riskier for a model to predict that an infected plant is healthy than to predict that a healthy plant is infected, we require the model to have specificity greater than sensitivity in addition to large values of AUC and accuracy. For each of the classifiers trained on our dataset, the performance measures were estimated and plotted on the histograms in Figure 7 after computing the AUC from the receiver operating characteristic (ROC) curves in Figure 8. In the following sections, we describe the four classifiers.

3.4.1 KNN

KNN classifier makes the categorization of unidentified instances based on a similarity measure or distance function. It is a supervised machine learning, lazy learning, and nonparametric model. It is based on the principle of nearest-neighbor rule. For model generation, this classifier does not require any training pattern. All training patterns are used in the testing phase to classify the test pattern dependent on a similarity function. It behaves as a kind of instance-based learning where the functions are locally estimated and all the calculations have varied until the completion of the classification method.

3.4.2 RF

It is an ensemble model of randomized decision tree classifiers. During training, multiple decision trees are constructed. Test dataset class labels are determined by voting of all classification trees, which becomes the result of this classifier. While building each individual tree, this classifier model uses bagging and random features. This model endeavors to make an unrelated forest of trees. The prediction of the performance of a forest of trees is more accurate than an individual tree since the result is an aggregation that reduces the variance in the result.

3.4.3 SVM

SVM is a supervised machine learning classifier model. This model finds the best hyperplane, which maximizes the distance among the nearest data points. This distance is referred to as a margin. SVM is of two types: linear and nonlinear. In a linear SVM classifier, the uniform distributions of data are allowed to draw a straight hyperplane among the classes. Whereas, in nonlinear SVM classifiers, data are spread in different directions and also have high dimensions. Most of the real-world applications are solved by nonlinear SVM classifiers. Kernel tricks are the property of SVM, which is helpful for nonlinear classification. SVM operates by transforming features with the help of several general functions, such as radial basis, polynomial, and linear functions. The training time of the classification process increases due to the transformation of features.

3.4.4 Adaptive boosting (AdaBoost or AB)

AB is an iterative selection of weak classifiers based on a distribution of training examples. Each example is weighted according to its difficulty with the current classifier. The outputs of these weak classifiers are combined into a weighted sum that represents the final output of the boosted classifier. AdaBoost is adaptive in the sense that subsequent weak classifiers are adjusted in favor of samples misclassified by previous classifiers. AdaBoost is notably sensitive to noisy or poorly correlated data. However, in some problems, it may be less prone to overfitting than other algorithms. The subclassifiers used can be weak as long as they perform at least slightly better than a random classifier, in which case the final model can be proven to converge to a strong classifier. AdaBoost (with decision trees as weak classifiers) is often referred to as the best turnkey workbook. The hyperparameters actually used for each of the four models are clearly provided in Section 4.2.1.

4 Results

4.1 Results on pretreatment and segmentation

By implementing the Gaussian filter on the original images from our database using Python, we obtain a new database of filtered images. Figure 9 shows an example of a filtered image generated from the coded image M1S4CA2404-11.

Figure 8

ROC curves for evaluating the performance of the different models on which the dataset was tested. (a) ROC of the RF model, (b) ROC of the AdaBoost model, (c) ROC of the SVM model optimized by a four-block cross-validation and a polynomial kernel of degree 4, and (d) ROC of the KNN model.

Once the database of filtered images is obtained, the next task is the segmentation of these images using the CNDVI algorithm applied to the visible light spectrum. The implementation of this new technique on the dataset allows us to have a third dataset of filtered and segmented images. The CNDVI is applied to all the pixels of the image and automatically masks out all the background pixels or those that do not correspond to the foliage. It retains, without alteration, all the pixels that correspond to the foliage. Figure 10 presents an example of a filtered and segmented image generated from the original image SS5CB1804-13.

Figure 9

The original image M1S4CA2404-11 (on the left side) and the filtered image by the Gaussian filter (on the right side).

4.2 Results on characteristic extraction and implementation of machine learning technical on characteristic database

From each of the harvested, preprocessed and segmented images, a quartet of features (entropy, contrast, energy, and inverse moment of difference) was extracted, and a label was assigned to each quartet of features in the images. All these features obtained on the 130 images are recorded in a table with five columns (one column for each of the four features and one column for the label). The feature set was divided into two: a training database containing 78% of the dataset and the rest for the test database. To identify infected leaves in a plantation, we applied four machine learning techniques, following a four-step approach.

Training each technique using the training database to generate a model for each algorithm,
Using the models obtained to evaluate the predictive ability of each algorithm on the test database,
Assessing the performance of each model, and
Selecting the best-performing model.

However, all other leaf diseases affecting corn can be detected by our work as infected plants because any leaf infection alters its texture, making the extracted texture features different from those of healthy leaves. Therefore, it will be sufficient to extract the features in the new images, label them, and then retrain the machine learning model.

4.2.1 Models training

Given that there is no way to know the best values for hyperparameters in advance, ideally, we should try all possible values to determine the optimal values. GridSearchCV is the process of automatically tuning hyperparameters to find optimal values for a given model. It is a function provided in the model-selection package of Scikit-learn (or SK-learn) that helps select the best parameters from the listed and tested hyperparameters. The “cv” parameter in this GridSearchCV function allows training k-folds to try the selected hyperparameters in cross-validation. For each of the classifiers, the hyperparameters that provided the best accuracy were:

SVM: Kernel = RBF (radial basis function), accuracy regularization C= 1 and influence of the example γ = 1.
KNN: Number of neighbors n _ neighbors = 6 .
RF: among about 20 hyperparameters used by this classifier, the one that best optimizes the model is the number of features used by the trees. In this case, max features = 4 .
AdaBoost: n _ estimators = 50 , learning _ rate = 1 .

4.2.2 Evaluation of model performance

After the test phase, a confusion matrix is generated to calculate the values of the following four performance measures of each classifier: specificity, sensitivity, precision, and AUC (Area Under the ROC Curve). Figures 7 and 8 present the results of these metrics, and Table 4 presents a comparison between the performances calculated on each classifier.

Figure 10

Original image SS5CB1804-13 (left side) and filtered and segmented image (right side) by CNDVI.

Table 4

Performance results for each of the four classifiers on which the data was tested

Performance model	Accuracy	Sensitivities	Specificities	AUC of ROC	Dominance rate
RF	0.97	1	0.95	0.9789473684210526	3/4
AD	0.97	1	0.95	0.9473684210526316	3/4
KNN	0.62	0.47	0.79	0.6710526315789473	0/4
SVM	0.97	1	0.95	0.9947368421052633	4/4

The analysis of the performance results of the four machine learning methods, summarized in Table 4, justifies that SVM is the most efficient classifier for our database as it has the largest values on the performance measures. The confusion matrix of the SVM model along with the accuracy results are shown in Tables 5 and 6, respectively.

Table 5

SVM model confusion matrix

Confusion matrix	Negative: 0	Positive: 1
Negative: 0	True negative = 19	False negative = 0
Positive: 1	False positive = 1	True positive = 9

Table 6

SVM model accuracy

Measure	Precision	Recall	f1-score	Support
Negative: 0	0.95	1.00	0.97	9
Positive: 1	1.00	0.90	0.95	10
Accuracy			0.97	29

5 Concluding remarks

Recent machine and deep learning techniques have demonstrated their capacity to address computer vision problems that were traditionally challenging for conventional methods. In this article, we present a model for the early detection of certain maize diseases from RGB leaf images captured in the field under natural lighting and environmental conditions. One of the major challenges when processing images taken in natural environments is separating the region of interest from the complex background of the image. Our novel approach involves a segmentation technique based on a CNDVI applied to the light spectrum. This new segmentation approach efficiently separates the corn leaves from the rest of the image, enabling feature extraction and classification. These techniques can be integrated into low-resource embedded systems.

The model proposed for identifying corn-plant diseases is based entirely on traditional machine learning methods. This choice is made because deep learning methods do not necessitate a preprocessing step for feature identification. In the proposed approach, the proposed segmentation method extracts features, which are then forwarded to a traditional machine learning method for classification. However, in our future work, we intend to collect a larger dataset comprising various types of corn-plant diseases, which can be used to train a deep-learning model for disease detection without requiring segmentation.

Acknowledgments

We appreciate the reviewers for their valuable comments, which have significantly improved the quality of our paper. All images included in this paper are originally produced by the authors.

Funding information: This work was carried out under the research grant FR 21-333 RG/MATHS/AF/AC_G-FR 3240319514 offered by UNESCO-TWAS and the Swedish International Development Cooperation Agency to the Laboratory of Mathematics of the University of Douala-Cameroon. The authors of the University of Douala sincerely thank TWAS. M.A. thanks the National Research Foundation of South Africa for support through project number CSRP23040990793.
Author contributions: Conceptualization: Thierry Tchokogoué, Auguste Vigny Noumsi, and Louis Aimé Fono; methodology: Thierry Tchokogoué, Auguste Vigny Noumsi, Marcellin Atemkeng, Michèle Flore Yimga Fonkou, and Louis Aimé Fono; software: Thierry Tchokogoué, Auguste Vigny Noumsi, and Marcellin Atemkeng; validation: Thierry Tchokogoué, Auguste Vigny Noumsi, Marcellin Atemkeng, Michèle Flore Yimga Fonkou, and Louis Aimé Fono; writing – original draft preparation: Thierry Tchokogoué and Michèle Flore Yimga Fonkou; writing – review and editing: Auguste Vigny Noumsi, Marcellin Atemkeng, and Louis Aimé Fono; supervision: Auguste Vigny Noumsi, and Louis Aimé Fono; funding: Marcellin Atemkeng and Louis Aimé Fono.
Conflict of interest: The authors have no conflicts of interest to declare. All coauthors have read and agree with the contents of the manuscript, and there is no financial interest to report.
Ethical approval: Not applicable, as the study did not involve the use of identifiable humans or animals, and thus did not require ethical clearance.
Data availability statement: The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

References

[1] Hanssen IM, Lapidot M. Major Tomato viruses in the Mediterranean basin. Adv Virus Res. 2012;84:31–66. 10.1016/B978-0-12-394314-9.00002-6Search in Google Scholar PubMed

[2] Tagne A, Tankou C, Ghoghomou TR, Ngoko Z, Tekeu H, Ngonkeu EM, et al. Fusarium diseases of maize (Fusarium verticilloides ex Fusarium moniliforme). Biol Biochem Sci. 2021;29(2):88–99. Search in Google Scholar

[3] Haleem A, Javaid M, Khan IH. Current status and applications of Artificial Intelligence (AI) in medical field: An overview. Current Med Res Practice. 2019;9(6):231–7. 10.1016/j.cmrp.2019.11.005Search in Google Scholar

[4] Becker A. Artificial intelligence in medicine: What is it doing for us today?. Health Policy Technology. 2019;8(2):198–205. 10.1016/j.hlpt.2019.03.004Search in Google Scholar

[5] Abduljabbar R, Dia H, Liyanage S, Bagloee SA. Applications of artificial intelligence in transport: An overview. Sustainability. 2019;11(1):189. 10.3390/su11010189Search in Google Scholar

[6] Rykiel Jr EJ. Artificial intelligence and expert systems in ecology and natural resource management. Ecologic Model. 1989;46(1–2):3–8. 10.1016/0304-3800(89)90066-5Search in Google Scholar

[7] Nandutu I, Atemkeng M, Okouma P. Intelligent systems using sensors and/or machine learning to mitigate wildlife-vehicle collisions: A review, challenges, and new perspectives. Sensors. 2022;22(7):2478. 10.3390/s22072478Search in Google Scholar PubMed PubMed Central

[8] Dharmaraj V, Vijayanand C. Artificial intelligence (AI) in agriculture. Int J Current Microbiol Appl Sci. 2018;7(12):2122–8. 10.20546/ijcmas.2018.712.241Search in Google Scholar

[9] Jha K, Doshi A, Patel P, Shah M. A comprehensive review on automation in agriculture using artificial intelligence. Artif Intell Agriculture 2019;2:1–12. 10.1016/j.aiia.2019.05.004Search in Google Scholar

[10] de la Cruz AJL. Potential of multispectral images acquired by drone in the detection of areas infected by the Golden Flavescence of the vine. Ph.D. thesis. France: University of Toulouse, 3 Paul Sabatier; 2018. Search in Google Scholar

[11] Fusellier JL. The corn sector in Cameroon (The corn sector in Cameroon, What prospects production development corn). Department of Annual Crops. 1993. Search in Google Scholar

[12] Salathé M, Mohanty SP, Hughes DP. Using deep learning for image-based plant disease detection. Front Plant Sci. 2016;7:14–9. 10.3389/fpls.2016.01419Search in Google Scholar PubMed PubMed Central

[13] Wäldchen J, Rzanny M, Seeland M, Mäder P. Identification of rice diseases using deep convolutional neural networks. Neurocomputing. 2017;267(4):378–84. 10.1016/j.neucom.2017.06.023Search in Google Scholar

[14] Aït EK, Boukouri S, Belbrick M, Hajji H, Chtaina N. Experimentation with a model for early detection of tomato diseases by deep learning. Crop Production 2020;14:19–30. Search in Google Scholar

[15] Ferentinos KP. Deep learning models for plant disease detection and diagnosis. Comput Electron Agric. 2018;145:311–8. 10.1016/j.compag.2018.01.009Search in Google Scholar

[16] Goel L, Nagpal A. A systematic review of recent machine learning techniques for plant disease identification and classification. IETE Tech Rev. 2022;40:423–39. 10.1080/02564602.2022.2121772. Search in Google Scholar

[17] Yu H, Liu J, Chen C, Heidari A, Zhang Q, Chen H, et al. Corn leaf diseases diagnosis based on K-means clustering and deep learning. Digital Object Identifier. 2021;9:143824–35. 10.1109/ACCESS.2021.3120379. Search in Google Scholar

[18] Mousavi SA, Hanifeloo Z, Sumari P, Arshad S. Enhancing the diagnosis of corn pests using Gabor wavelet features and SVM classification. J Scientific Industrial Res. 2016;75:349–54. Search in Google Scholar

[19] Pawan E, Thamrin R, Widodo W, Bei S, Luanmasa J. Implementation of forward chaining method in expert system to detect diseases in corn plants in Muara Tami district. Int J Comput Inform Syst (IJCIS). 2022;75:349–54. 10.29040/ijcis.v3i1.59Search in Google Scholar

[20] Seyed M, Ahmad B, Keyvan A, Yiannis A. Diagnosis of grape leaf diseases using automatic K-means clustering and machine learning. Smart Agricultural Technology. 2022;3:1000815. 10.1016/j.atech.2022.100081Search in Google Scholar

[21] Deng R, Tao1 M, Xing H, Yang X, Liu C, Liu C, et al. Automatic diagnosis of rice diseases using deep learning. Frontiers Plant Sci. 2021;12:701–38. 10.3389/fpls.2021.701038Search in Google Scholar PubMed PubMed Central

[22] Sumita M, Rishabh S, Diksha R. Deep convolutional neural network based detection system for real-time corn-plant recognition. Proc Comput Sci. 2020;167:2003–10. 10.1016/j.procs.2020.03.236Search in Google Scholar

[23] Pan S, Qiao J, Wang R, Yu H, Wang C, Taylor K, et al. Intelligent diagnosis of Northern corn leaf blight with deep learning model. J Integrative Agriculture. 2022;21(4):1094–105. 10.1016/S2095-3119(21)63707-3Search in Google Scholar

[24] Wäldchen J, Mäder P. Using deep learning for image-based plant disease detection. Archives Comput Methods Eng. 2018;25(2):507–43. 10.1007/s11831-016-9206-zSearch in Google Scholar PubMed PubMed Central

[25] Prajapati HB, Shah JP, Dabhi VK. Detection and classification of rice plant diseases. Intelligent Decision Technologies. 2017;11(3):357–73. 10.3233/IDT-170301Search in Google Scholar

[26] Wäldchen J, Rzanny M, Seeland M, Mäder P. Automated plant species identification-trends and future directions. PLoS Comput Biol. 2018;14(4):1–19. 10.1371/journal.pcbi.1005993Search in Google Scholar PubMed PubMed Central

[27] Kosal K. Contribution de l’imagerie dronique pour la caractérisation des paramètres biophysiques des cultures agricoles. PhD thesis. Canada: University of Montreal; 2020. Search in Google Scholar

Received: 2023-04-13

Accepted: 2023-11-28

Published Online: 2024-05-31

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/jisys-2023-0049

Keywords for this article

phytosanitary; diagnosis; maize; segmentation; Adaboost; random forest; k-nearest neighbors; support vector machine

Creative Commons

BY 4.0