Early Detection of Cotton Plant Diseases Using Zebra Optimizer with Deep Learning Approach

The prevalence of plant diseases poses a major challenge to agriculture by reducing crop yield and deteriorating product quality. Hence, it is essential to adopt effective methods for early and accurate detection of diseases in agriculture. Among major commercial crops, cotton plays a vital role in the textile industry and contributes substantially to agricultural economies in many countries [1, 2]. However, cotton plants are susceptible to various leaf diseases such as bacterial blight, powdery mildew, target spot, aphids, and armyworm infestations. These diseases cause severe damage to leaves, reduce photosynthetic activity, and ultimately decrease crop yield and quality. Traditionally, farmers identify these diseases through manual visual inspection of plant leaves. Although this method is widely practiced, it is often time-consuming, subjective, and impractical for monitoring large-scale agricultural fields. Furthermore, incorrect diagnosis may lead to improper pesticide usage, which increases production costs and negatively impacts the environment.

With the rapid advancement of artificial intelligence and computer vision techniques, automated plant disease detection systems have emerged as an effective solution for improving agricultural monitoring [3, 4]. Machine learning (ML) and deep learning (DL) models have demonstrated strong capability in analyzing plant images and identifying disease symptoms [5-7]. In particular, Convolutional Neural Networks (CNNs) have been widely used for feature extraction and classification in plant disease detection tasks. Several studies have applied DL models such as VGG, Inception, ResNet, and EfficientNet for leaf disease classification with promising results [8]. Transfer learning techniques have also been utilized to improve model performance when training datasets are limited. Despite these developments, several challenges still remain in developing reliable automated disease detection systems [9-11]. Variations in illumination, background complexity, leaf orientation, and similarity between healthy and diseased leaf textures can reduce the accuracy of classification models. In addition, many existing approaches rely on manually selected hyperparameters, which may limit the adaptability and performance of the models [12-15].

To overcome these limitations, this research proposes an integrated DL framework for cotton plant disease detection that combines image preprocessing, feature extraction, optimization, and classification techniques. Initially, Wiener filtering is applied during the preprocessing stage to remove noise and enhance the quality of input images. This step improves image clarity and preserves important visual patterns that are necessary for reliable disease identification. After preprocessing, EfficientNet is employed as the feature extraction model. EfficientNet is known for its scalable architecture and balanced network design, which enables efficient learning of complex visual features from plant leaf images.

In order to further improve the performance of the DL model, hyperparameter tuning is performed using the ZOA. ZOA is a swarm-based optimization technique inspired by the natural behavior of zebras during foraging and defense against predators. The algorithm effectively explores the search space to identify optimal hyperparameter values, which enhances the learning capability of the model and reduces the dependency on manual parameter selection. Following the optimization stage, an attention-based ConvLSTM network is employed for the classification of cotton leaf diseases. ConvLSTM combines convolutional operations with recurrent neural networks to capture both spatial and contextual dependencies within the extracted features. The integration of an attention mechanism further improves the model’s ability to focus on important feature regions associated with disease patterns.

The novelty of this work lies in the integration of preprocessing, deep feature extraction, adaptive optimization, and attention-based spatiotemporal classification within a unified framework. Unlike conventional approaches that rely on standalone CNN models or manual parameter tuning, the proposed method combines image enhancement, optimization-driven learning, and contextual feature modeling to improve detection accuracy and robustness.

Thus, the main contributions of this work include the following:

• Development of an automated DL framework for accurate detection and classification of cotton plant leaf diseases to support precision agriculture.

• Application of Wiener filtering–based preprocessing to enhance image quality by reducing noise and improving disease-related feature visibility.

• Utilization of the EfficientNet architecture for robust feature extraction, enabling effective learning of complex visual patterns in cotton leaf images.

• Implementation of ZOA for adaptive hyperparameter tuning to improve model performance and training efficiency.

• Integration of an attention-based ConvLSTM classifier to capture spatial and contextual dependencies for improved multi-class disease classification.

The paper organization is as follows. Section 1 is the introduction regarding cotton plant leaf disease. Section 2 is a literature survey of existing works. Section 3 elaborates the proposed methodology with data collection, pre-processing using Wiener filter, feature extraction using EfficientNet, hyper parameter tuning using ZOA, and classification using attention based ConvLSTM. Section 4 is the result and analysis. Section 5 is the conclusion of the research work.

1.1 Research motivation

Cotton crops are referred to as "cash crops." The growth of Cotton plants may be affected by a wide range of diseases that impair production through the leaf, including target spot, leaf spot etc. Traditional disease identification methods rely on manual inspection is a time consuming one and often inaccurate for large farming areas. Recent deep learning techniques have improved plant disease detection, but many models still face challenges such as noise in images, complex backgrounds, and improper hyperparameter selection. Therefore, this work aims to develop an efficient detection framework by integrating image preprocessing, deep feature extraction, optimization-based hyperparameter tuning, and attention-based classification. The proposed approach enhances disease recognition accuracy and supports early diagnosis for effective crop management in precision agriculture.

3.1 Data collection

The dataset used for the proposed cotton plant disease classification is obtained from a publicly available Kaggle repository. It is gathered from the link, “https://www.kaggle.com/datasets/dhamur/cotton-plant-disease”. The dataset mainly contains leaf images representing both healthy and diseased cotton plants. The disease categories included in the dataset are depicted in Figure 2. All images are captured under natural field conditions, which introduces variations in lighting, background, orientation, and leaf texture. The dataset focuses only on leaf-level symptoms and does not include images of stems, buds, flowers, or bolls. This class diversity enables multi-class disease classification, while the presence of healthy samples supports reliable discrimination between infected and non-infected leaves. The dataset provides sufficient variability in color patterns, lesion shapes, and infection severity to evaluate the robustness of the proposed detection framework. To improve generalization and handle variability, data augmentation techniques (rotation, flipping, scaling) are applied to the training images.

Figure 2. Sample images gathered from the Kaggle dataset

The entire dataset has been divided in the ratio of 80:20 for training and testing process. The details about the samples used in the experimentation under different classes are presented in Table 2.

Table 2. Experimental dataset details

3.2 Pre-processing

After collecting the images from the Kaggle dataset, the pre-processing is accomplished using the Wiener filter. This helps in the removal of noise and also improves the quality of the collected input images. An ideal trade-off between noise smoothing and inverse filtering is carried out via Wiener filtering. Meanwhile, it reverses the blurring as well as eliminates the additional noise. The Mean Square Error (MSE) among the intended procedure as well as the approximated procedure is reduced using the Wiener filter. During the noise smoothing and inverse filtering processes, it reduces the total MSE. A linear estimate associated with the original image is what the Wiener filter does. Wiener filters are useful for reducing noise in images by comparing the image's noise level to an estimate of the desired noiseless signal. It is predicated on statistical analysis. Three crucial characteristics define Wiener filters.

Performance criterion: Minimum Mean-Square Error (MMSE). This filter is commonly employed during the deconvolution procedure.

Requirement: The filter needs to be causal and physically realizable.

Assumption: Stationary linear stochastic procedures associated with noise and image having known autocorrelation or known spectral characteristics and cross correlation.

The Wiener filter is widely used due to its quickness as well as simplicity. It is considered simple because it determines optimal filter weights by solving linear equations to reduce noise in the received signal. These weights are computed using covariance and cross-correlation of the noisy signal, enabling accurate estimation of the underlying signal in the presence of Gaussian noise. The signal's deterministic component is further evaluated by processing a fresh input signal having suitable filter weights and comparable noise features. When the noise distribution is Gaussian, this technique works well. Moreover, a small number of quick computing processes are needed to execute it. Fourier Domain Wiener Filter is shown below.

$H(v, w)=\frac{I *(v, w) Q_t(v, w)}{|I(v, w)|^2 Q_t(v, w)+Q_o(v, w)}$     (1)

Dividing by $Q_t$ facilitates the explanation of its behavior:

$H(v, w)=\frac{I *(v, w)}{|I(v, w)|^2+\frac{Q_o(v, w)}{Q_t(v, w)}}$     (2)

Here, the Power Spectral Density (PSD) associated with the un-degraded image is shown by $Q_t(v, w)$, PSD associated with the noise is shown by $Q_o(v, w)$, complex conjugate related to the degradation function is given by $I *(v, w)$, and the degradation function is given by $I(v, w)$ respectively.

The phrase $\frac{Q_o}{Q_t}$ may be seen as the Signal-to-Noise Ratio (SNR) reciprocal. Using statistics obtained from the local neighborhood associated with every pixel, the Wiener filter is employed to eliminate noise from a distorted image. The noise strength of an image, or the variance in noise, determines how this filter works. A big variation results in minimal smoothing, whereas a small variance results in more smoothing from the filter.

3.3 Feature extraction

The features are extracted from the pre-processed images using the EfficientNet method. Its efficiency as well as scalability helps in capturing intricate patterns as well as variations related to distinct cotton plant diseases. The Google Brain Team created CNN method EfficientNet. According to this network scaling, performance may be increased by optimizing the depth, breadth, as well as resolution of the network. Scaling a Neural Network (NN) to build additional DL models, which provide far greater effectiveness and accuracy than the earlier employed CNNs, is one way to generate a novel method. EfficientNet consistently and accurately completed large-scale image recognition tasks for the Image Net. A network's depth and count of layers are correlated. The breadth of the convolutional layer corresponds to the count of filters it has. The resolution is determined by the supplied image's height as well as breadth.

$e=\alpha \varphi$     (3)

$x=\beta \varphi$      (4)

$s=\gamma \varphi$     (5)

$\alpha \geq 1, \beta \geq 1, \gamma \geq 1$     (6)

Here, $e$, $x$, and $s$ stand for the depth, width, as well as resolution of the network, correspondingly, and the grid search hyperparameter tuning method was used to obtain the constant terms $\alpha$, $\beta$, and $\gamma$. All model scaling resources are managed by a user-defined variable called the coefficient. On the basis of available resources, this method modifies the network's depth, breadth, and resolution to maximize memory use and accuracy. EfficientNet confirmed its usefulness outside of the dataset and yielded exceptional outcomes even with the use of the transfer learning approach. Scales ranging from 0 to 7 were included in the method's release, signifying an improvement in parameter size and accuracy.

In this case, the chosen classifier is the sigmoid. The mathematical function associated with a sigmoid classifier having a recognizable S-shaped curve is displayed in Eq. (7). A logistic function that conducts a binary classification is called sigmoid. It establishes a threshold value of 0.5, assigning values to either 0 or 1. These classes are represented by the neuron at the last dense layer.

$(y)=\operatorname{Sigmoid}(y)=\frac{1}{1+e^{-y}}$     (7)

3.4 Hyper parameter tuning using zebra optimization algorithm (ZOA)

Following the feature extraction process, the hyperparameter tuning is performed for the automatic selection of optimal hyper parameters, thereby leading to attaining peak performance on the particular task of cotton plant disease detection. here, the hyperparameter tuning is accomplished by the ZOA. As a result, fine-tuning was necessary because EfficientNet's present structure could not be used for the task that is selected. After that, the training data is used to fine-tune the suggested end layers while freezing every layer associated with the basic method. By using this technique, it is capable of maintaining the feature extraction capacity in the weights that are acquired during training with the dataset inside the extraction layers and stop the training iterations from overriding them. Next, all of the layers in the network are unfrozen using weights taken from the dataset after training the suggested layers. This allows us to integrate and build the final method. Next, the finished model is verified using the test data.

ZOA replicates the way zebra that eat and defend themselves from predators. Zebras are part of the population that uses the population related optimizer ZOA. Mathematically, the plain with zebras on it represent the search space for the problem and each zebra represents a potential solution.

The ZOA is employed to optimize key hyperparameters of the proposed model to enhance classification performance. The parameters selected for optimization include learning rate, batch size, number of training epochs, and dropout rate, as these significantly influence model convergence and generalization.

The details of simulation hyperparameters are presented in Table 3.

The population size and iteration count of the Zebra Optimization process are fixed to ensure stable convergence during hyperparameter tuning.

The objective function of ZOA is to maximize classification accuracy (or minimize validation loss). The algorithm iteratively updates candidate solutions until convergence is achieved, which occurs either when the maximum number of iterations is reached or when there is no significant improvement in performance over successive iterations.

Table 3. Simulation hyperparameters

Decision variable values are based on where each zebra is located in the search space. The elements associated with a vector that depict each zebra as an individual inside the ZOA may thus be used to describe the values associated with the issue variables. The zebra population can be quantitatively designed using a matrix. Initially, the zebras are placed at random around the search area. The ZOA population matrix can be found in Eq. (8).

$Z=\left[\begin{array}{c}Z_1 \\ \vdots \\ Z_k \\ \vdots \\ Z_P\end{array}\right]_{P \times o}=\left[\begin{array}{ccccc}z_{1,1} & \cdots & z_{1, l} & \cdots & z_{1, o} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ z_{k, 1} & \cdots & z_{j, l} & \cdots & z_{k, o} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ z_{P, 1} & \cdots & z_{P, l} & \cdots & z_{P, o}\end{array}\right]_{P \times o}$    (8)

In this case, $P$ denotes the number of population members (zebras), $O$ represents number of decision variables, $Z$ denotes the population of zebras, $Z_k$ denotes the $k^{t h}$ zebra, and $z_{j, l}$ denotes the solution to the $l^{\text {th}}$ problem variable that the $k^{\text {th}}$ zebra supplied. Each zebra represents a possible fix for the optimization problem. Hence, the objective function can be assessed by referring to the values that each zebra suggests for the problem variables. A vector representing the values acquired for the objective function is provided by Eq. (9).

$H=\left[\begin{array}{c}H_1 \\ \vdots \\ H_k \\ \vdots \\ H_P\end{array}\right]_{P \times 1}=\left[\begin{array}{c}H\left(Z_1\right) \\ \vdots \\ H\left(Z_k\right) \\ \vdots \\ H\left(Z_P\right)\end{array}\right]_{P \times 1}$     (9)

H shows the vector of objective function values in this situation and $H_k$ shows the value of the objective function for the $k^{t h}$ zebra. One way to find the best possible solution to a problem is to compare the values of the objective function, which measures how well different solutions perform.

The best zebra in the group is called the pioneer zebra in ZOA; it leads the population's other zebras to its position in the search space. Therefore, it is possible to replicate numerically how zebras' locations are changed throughout the foraging phase using Eqs. (10) and (11).

$z_{k, l}^{\text {new }, R 1}=z_{k, l}+t \cdot\left(R B_l-K \cdot z_{k, l}\right)$    (10)

$Z_k=\left\{\begin{array}{c}Z_k^{\text {new }, R 1}, \quad H_k^{\text {new }, R 1}<H_k \\ Z_k, \quad \text { else }\end{array}\right.$     (11)

Here, the values associated with the objective function $H_k^{{new,R1}}$, $Z_k^{{new,R1}}$ represents the novel status related to the $k^{t h}$ zebra based on the first phase, the $l^{\text {th}}$ dimension value $z_{k, l}^{\text {new, } R 1}$, the $l^{\text {th}}$ dimension $R B_l$, and the random count $t$ in the interval [0, 1] are displayed. Moreover, $R B$ displays the pioneer zebra, which characterizes the ideal person. $K$ is equal to round(1 + Rand), where Rand represents random count with interval of [0, 1]. If K belongs to {1, 2} and if K = 2, then the mobility of population altered more significantly. According to the ZOA method, there exists an equal chance of either among the following two scenarios happening: In the initial approach, zebras are ambushed by lions and attempt to escape by congregating in the vicinity of their location. This means that this procedure can be quantitatively represented using the mode $U_1$ in (12). In an effort to mislead and frighten the attacker, the remaining zebras in the herd move toward the attacked one, attempting to create a defense pattern. The second approach is this one. This zebra approach is mathematically represented by the mode $U_2$ in (12). If changing a zebra's location results in a higher value for the objective function there, then the new position can be allowed. Using the Eq. (13) update scenario for model is shown below.

$z_{k, l}^{\text {new}, R 2}=\left\{\begin{array}{c}U_1: z_{k, l}+T \cdot(2 t-1) \cdot\left(1-\frac{v}{V}\right) \cdot z_{k, l}, \quad R_u \leq 0.5 \\ U_2: z_{k, l}+t \cdot\left(C B_l-K \cdot z_{k, l}\right), \quad {otherwise}\end{array}\right.$     (12)

$Z_k=\left\{\begin{array}{c}Z_k^{\text {new}, R 2}, \quad H_k^{\text {new}, R 2}<H_k \\ Z_k, \quad \text { otherwise }\end{array}\right.$     (13)

The novel status related to the $k^{t h}$ zebra based on the second phase is shown by $Z_k^{\text {new, R2 }}$; $z_{k, l}^{\text {new, } R 2}$ represents its $l^{\text {th}}$ dimension value, the objective function value is shown by $H_k^{\text {new, } R 2}$; the iteration contour is shown by $v ;$; the maximum count of iterations is shown by V. T represents the constant count, which is equal to value of 0.01. $R_u$ represents the probability of selecting one of two schemes produced at random from the interval [0, 1]. The state of attacked zebra is represented by $C B$, and the $l^{{th}}$ dimension value is shown by $C B_l$. The final step of each ZOA cycle is to update the population members using data from the first and second stages. Until the method is fully completed, the population is updated according to phases (10) through (13) of the process. After multiple iterations, the optimal candidate solution is refined and stored. After it has been fully developed, ZOA stands out as the best option to solve the given challenge. The ZOA phases are shown as pseudocode in Algorithm 1.

3.5 Classification using attention-based ConvLSTM

Although ConvLSTM is commonly used for temporal or sequential data, it can also be effectively applied to capture spatial and contextual dependencies within feature representations. In this work, the feature maps extracted from EfficientNet are treated as structured spatial sequences. The ConvLSTM layer models the relationships among these features, enabling better representation of complex disease patterns.

Unlike conventional CNN classifiers that process features independently, ConvLSTM captures inter-feature dependencies, which is particularly beneficial for distinguishing visually similar disease classes. Additionally, the integration of an attention mechanism allows the model to focus on the most relevant regions of the feature maps, further enhancing classification performance.

The optimal features obtained from the feature extraction and optimization stages are then fed into the attention-based ConvLSTM model for final classification. ConvLSTM combines convolutional operations with recurrent learning, enabling it to preserve spatial structure while modeling contextual relationships among features. Compared to fully connected LSTM (FC-LSTM), ConvLSTM applies convolution operations within the gating mechanisms, making it more suitable for structured feature maps derived from image data. The mathematical formulation of ConvLSTM is described as follows. Initially, figure out the input gate:

$j_u=\sigma\left(X_{y j} * y_u+X_{i j} * i_{u-1}+X_{d j}{ }^{\circ} d_{u-1}+c_j\right)$     (14)

The forget gate is calculated as below.

$g_u=\sigma\left(X_{y g} * y_u+X_{i g} * i_{u-1}+X_{d g}{ }^{\circ} d_{u-1}+c_g\right)$     (15)

The cell state is computed as follows.

$d_u=g_u{ }^{\circ} d_{u-1}+j_u{ }^{\circ} \tanh \left(X_{y d} * y_u+X_{i d} * i_{u-1}+c_d\right)$     (16)

The output gate is calculated as below.

$p_u=\sigma\left(X_{y p} * y_u+X_{i p} * i_{u-1}+X_{d p}{ }^{\circ} d_u+c_p\right)$     (17)

The hidden state is measured as below.

$i_u=p_u{ }^{\circ} \tanh \left(d_u\right)$      (18)

Here, the convolution operator is shown by *, the Hadamard product is shown by $\circ$, and the sigmoid function is represented by $\sigma . X_{y j}, X_{y g}, X_{y d}$ and $X_{y p}$ describes the matrices of weight joining the inputs $y_1, \cdots, y_u$ to three gates and input of the cell; $X_{i j}, X_{i g}, X_{i d}$ and $X_{i p}$ describes the matrices of weight joining the hidden states $i_1, \cdots, i_{u-1}$ to three gates and input of the cell; $X_{d j}, X_{d g}$ and $X_{d p}$ describes the weight matrices joining the $d_1, \cdots, d_u$ to three gates; and $c_j, c_g, c_d$ and $c_p$ represents the bias terms of three gates as well as the cell state.

A number of recent sequence-to-sequence learning experiments have shown that the attention mechanism is effective. The attention method concentrates on the key problem having the LSTM-oriented classification method, which favors choosing short-term data that has a strong correlation with the future. The fundamental ConvLSTM method that creates the hidden state description $i_u$ serves as the encoder in the experiments. A self-attention mechanism is employed to process the inputs following the functions of Eqs. (14)-(18).

$n_{u, u^{\prime}}=\tanh \left(X_n i_u+X_{n^{\prime}} i_{u^{\prime}}+c_n\right)$     (19)

$f_{u, u^{\prime}}=\sigma\left(X_b n_{u, u^{\prime}}+c_b\right)$     (20)

$b_u=\operatorname{softmax}\left(f_u\right)$    (21)

$m_u=\sum_{u^{\prime}=1}^o b_{u, u^{\prime}} \cdot i_{u^{\prime}}$    (22)

where, $c_n$ and $c_b$ represent bias terms; $X_n$ and $X_{n^{\prime}}$ describe weight matrices respective to the hidden states $i_u, i_{u^{\prime}}$; and $b_{u, u^{\prime}}$ describes an attention matrix. Last but not least, $m_u$ denotes a weighted total of $i_{u^{\prime}}$. As inputs, many optimal features are initially generated. After that, the model is trained to identify distinct cases of cotton plant leaf disease such as target spot, powdery mildew, bacterial blight, army worm, aphids, and healthy leaf respectively.