Deep Learning and Grey Wolf Optimization Technique for Plant Disease Detection: A Novel Methodology for Improved Agricultural Health

Plants are a crucial energy source and an excellent solution to address global warming. However, they are susceptible to numerous diseases that can have severe adverse effects on the environment, society, and the economy, potentially causing a significant decline in crop yield and quality [1]. Agriculture provides livelihood for 50% of the world's population, with the waste from harvesting serving as a source of income and sustenance, including products such as wax, oil, and jute [2].

In plant systems, emerging, re-emerging, and endemic diseases cause significant damage and can result in financial loss. Moreover, plant diseases indirectly and directly propagate infectious diseases that threaten humans and the environment. The spread of these diseases not only harms the plant's functionality but also the economy by limiting the amount of crop that can be grown [3]. Plant diseases can affect various plant components, including fruits, stems, and leaves. The most common bacterial, fungal, and viral diseases in plants are Alternaria, Anthracnose, bacterial spot, canker, and others.

Environmental changes often cause viral illnesses, bacterial diseases are brought on by germs in plant leaves, and fungal illnesses are caused by leaf fungi [4]. Plant diseases frequently threaten the output and quality of global agricultural products, contributing to a significant amount of production costs. It is estimated that plant disease losses have reduced global food output by at least 10% [5].

Awareness of the detrimental effects of extensive chemical pesticide usage on the environment and health is increasing. Consumers now prefer organically grown food. Additionally, regulatory bodies like the European Union (EU) are tightening restrictions on the use of chemicals in agricultural goods imported into their markets. These developments underline the need for early and accurate detection of pests and diseases, highlighting the importance of reducing pesticide use in agriculture [6].

Plant disease prevention and control have been the subject of intense discussion as plants, exposed to the outside environment, are extremely susceptible to infections. Accurate and rapid disease identification is crucial in effectively combating plant diseases, as it often leads to effective protective measures [7]. Plant diseases commonly reduce yields, quality, and economic growth. Therefore, professional observation with the naked eye is labor-intensive, and automation for identifying leaf diseases is a significant area of research [8].

It has now become necessary to identify and diagnose plant diseases using information technology. Techniques such as automated plant disease identification and categorization must be developed by employing leaf image processing methods. These techniques will serve as a beneficial strategy for farmers, providing early warnings before the disease spreads widely [9].

1.1 Type of plant diseases

Crop leaves are particularly susceptible to disease. Pathogens, including viruses, bacteria, fungi, and deficiencies, severely affect crops. Consequently, pathogens are categorized into two types: autotrophs, which feed on living tissue, and saprophytes, which feed on dead tissue. The symptoms of these diseases, which hinder crop development and growth, are clearly visible. The first sign of illness in plants is often discolored leaves [10].

In addition, the texture and shape of the leaves can be used to identify several diseases. Therefore, processing images of the leaves can help identify diseases such as mildew, rust, and powdery mildew [11]. There are three types of viruses and contagious plant diseases, as shown in Figure 1.

1.png

Figure 1. Types of viruses and contagious plant diseases

1.2 Fungal diseases

Fungi cause the majority of vegetable diseases. Fungal infections stress plants and destroy their cells. These infections can originate from various sources, including contaminated seeds, unsuitable soil, agricultural waste, neighboring crops, and weeds. Their focal points are the organic apertures, like the stomata in plants. Fungal growth can also occur inside plants when artificial wounds are created by pruning, harvesting, flooding, insects, or other diseases [12].

1.3 Bacteria diseases

Many parts of plants are susceptible to bacterial infections. Bacteria can affect plants internally without showing any outward symptoms. Cankers, leaf spots, overgrowth, scabs, and wilts are among the signs that a plant is infected by bacteria. An infected plant can serve as a source of infection for other nearby plants, rapidly spreading the disease [13].

1.4 Viral diseases

Viral infections affecting plant leaves are the most challenging to identify. Most viruses are latent, meaning they cannot be detected until they reach a certain level. It is common to confuse viral infections with nutritional deficiencies and pesticide damage. Carriers such as aphids, leafhoppers, whiteflies, cucumber beetles, and other insects can easily transmit viruses [14].

1.5 Related works

Given the importance of plants in providing food, reducing global warming, and acknowledging the serious negative environmental impact of plant diseases, it's necessary to find quick and accurate solutions for plant disease detection.

• In 2020, Sindhu and Indirani [1], developed a cloud-based IoT model for diagnosing wheat leaf illnesses using Optimal Deep Neural Network (ODNN) and Grey Wolf Optimization (GWO) Algorithm, achieving 95.92% precision, 96.41% recall, and 96.96% accuracy.
• In 2022, Harakannanavar et al. [12], developed a machine learning method for detecting leaf disease on tomato plants using SVM, CNN, and K-nearest neighbors, with SVM (88%), K-NN (97%), and CNN (99.6%) being the most accurate classifiers.
• In 2019, Sullca et al. [15], combined CNN, SVM, ANN, and RF to create a model for identifying objects. They used HOG and LBP for feature extraction, and the model achieved 84% accuracy in classifying plants as infected.
• In 2020 Anagnostis et al. [16]. developed a CNN model for classifying leaf pictures with anthracnose contamination, achieving 98.7% accuracy in grayscale and RGB formats.
• In 2020, Kaur and Devendran [17], A convolutional low mask was used to extract texture features with grey wolves’ method for optimizing segmentation. and an SVM classifier was used to categorize plant diseases.
• In 2019 Geetharamani and Pandian [18]. A novel model was created to diagnose a plant leaf ailment developed a novel model for diagnosing plant leaf ailment using 39 distinct classes and unique photos, achieving 96.4% accuracy using image flipping, gamma management, and principal component analysis.
• In 2019 Balakrishna and Rao [19]. Use PNN and KNN techniques to detect early tomato leaf health, improving accuracy to 91.88%.
• In 2019 Zhang et al. [20]. Developed a three-tone model of a three-channel convolutional neural network (CNN) to detect plant leaf disease. Each CNN learns diminishing light and protects one of the three darkened regions of sick foliage. The SoftMax layer grouping information photos into pre-established categories achieved 94.2% accuracy.
• In 2017 Lu et al. [21]. used CNNs to detect rice illness using 500 photos from a test field, achieving 95.48% accuracy.

The objective of GWO is to identify the subset of characteristics that improves the model's performance. By combining the results of Local binary pattern (LBP), Histogram of oriented gradients (HOG), and Haralick texture features, the next step will be applying GWO algorithm steps to these features to select the most efficient features. These are the steps of the algorithm that was followed to build the optimization algorithm in this research.

Step 1. Load the LBP, HOG, and Haralick features

Step 2. Concatenate the features

             X=[LBP HOG Haralick];

Step 3. Define the search space.

            d=size(X, 2);

            searchSpace=zeros(d, d);

Step 4. Define the objective function.

             objFcn=@(subset) modelPerformance(X(:, subset), y);

Step 5. Define the stopping criteria.

            maxIter=100; //Max Iteration

Step 6. Initialize the positions of the alpha, beta, and delta wolves.

            alpha=randi(d);

            beta=randi(d);

            delta=randi(d);

Step 7. Initialize the best solution.

            bestSol=zeros(d, 1);

            bestVal=-inf;

Step 8. Run the GWO algorithm.

            For iter: $\leftarrow$ 1 to maxIter

            //Calculate the objective function value for each wolf

            alphaVal=objFcn(alpha);

            betaVal=objFcn(beta);

            deltaVal=objFcn(delta);

            //Update the positions of the wolves

            alpha=updatePosition (alpha, alphaVal, betaVal, deltaVal, searchSpace);

            beta=updatePosition (beta, alphaVal, betaVal, deltaVal, searchSpace);

            delta=updatePosition (delta, alphaVal, betaVal, deltaVal, searchSpace);

            //Update the best solution

            if alphaVal>bestVal

            bestSol=alpha;

            bestVal=alphaVal;

            end

            if betaVal>bestVal

            bestSol=beta;

            bestVal=betaVal;

            end

            if deltaVal>bestVal

            bestSol=delta;

            bestVal=deltaVal;

            end

            GWOFeatures$\leftarrow$bestVal

            end For

            Return GWOFeatures

3.1 Convolution neural networks

Convolutional neural network (CNN) is a multilayer perceptron (MLP) network. The human brain serves as an inspiration for CNN's computers. A CNN needless pre-processing and feature extraction work since it combines the feature extraction and classification operations [33]. The convolutional layer, the pooling layer, and the output layer make form a fundamental convolutional neural network. Sometimes a pooling layer is not necessary with three convolutional layers in Figure 5. The input layer, many hidden layers (repetition of convolutional, normalizing, and pooling), a fully connected layer, and an output layer make up the system [34].

5.png

Figure 5. Typical convolutional neural network architecture

Although CNN does not need pre-processing and feature extraction, in this paper we used three feature descriptors to extract features and passed them to the GWO algorithm to improve and reduce these features. Then we used a two-dimensional neural network to build the classification model.

Data preprocessing is a step-in data analysis that involves image resize, and data augmentation.

5.1 Image resize

All images in the dataset will be resized to 256 pixels in width and 256 pixels in height during the loading process. Figure 7 showed resized images.

7.png

Figure 7. Dataset samples for images resizing

5.2 Data augmentation

To increase the amount of data each class in dataset was augmented using three transformation functions X-reflection, Y-reflection, and Rotation. Table 1 shows the augmentation results for each class. Table showed the augmented data.

Table 1. Augmentation results

For three types of plant leaves, pepper, potato, and tomato, and for the two classes, three classes, and nine classes, results were obtained by using different population samples (125, 150, 175 and 200), 1900 instants for the pepper ,2200 instants for the potato and 12000 instants for the tomato, 0.3 threshold, 50 iterations, and a 70:30% training-to-testing ratio for the three plant types: pepper, potato, and tomato. The number of features that were chosen for five rounds is shown in Table 3.

Table 3. Feature selection results for different population in pepper, potato and tomato