Plant Leaf Disease Detection Using Metaheuristic Optimization Algorithms and Deep Learning

Plant diseases adversely affect the agricultural productivity. Food insecurity may increase, if the plant diseases are not diagnosed promptly [1]. Earlier identification is vital for effective control and prevention of the plant diseases and they serve a crucial role in managing the agricultural productivity as well [2]. At present, plant disease detection remains a huge concern. Precise and prompt analysis of the plant diseases is critical for correct and sustainable agricultural production [3]. Certain plant diseases have invisible indications; in such cases, the advanced analyses techniques are used. However, most plant diseases start showing symptoms whereas skilled plant pathologists can sense the conditions and detect the disease by examining the affected plant leaves visually [4]. A pathologist should have good observation skills to find the characteristic symptoms in an accurate manner. However, the variations in the context of plant disease due to climatic changes, a large variety of plants and the faster spread of disease lead even skilled pathologists to misdiagnose some conditions [5].

In this background, the exploitation of intelligent and expert systems to automatically detect the plant disease precisely offers valuable contributions to agriculturists [6]. In addition, when these systems are presented as simple mobile apps, which non-expert farmers can easily use, it would be a great achievement and it helps the farmers to make decisions effectively. The recent developments in Artificial Intelligence (AI) technology have made a huge contribution to the advancement of automatic systems that can give precise and faster outcomes in disease diagnosis [7]. Recently, in the study of plant disease detection, Deep Learning (DL) technology has been further developed [8]. DL has expressed original image characteristics that contain numerous end-to-end features. Such characteristics make the DL technology obtain more attention in plant disease detection domain [9]. CNN, a type of DL method, is the preferable approach in disease detectio processes. It is well known for its potential in image processing and classification applications [10]. DL methods were first introduced in plant image detection based on leaf vein patterns.

1.1 Objective

The main goal behind early detection of plant diseases is to control it as it reduces the crop yields. Plant diseases have historically reduced the crop yields and endangered the food supply process. The aim of this research work is to develop a dynamic system that can accurately diagnose and classify the plant leaf diseases based on the symptoms observed in leaf pictures. This sort of identification is required for early illness detection and mitigation. The aim of the APLDD-ESOSDL is to classify the leaf diseases from plant pictures in an accurate manner. This classification is essential for decision-making in agriculture. The APLDD-ESOSDL method extracts the features using the Inception ResNet-v2 model. In general, image-based disease diagnosis requires feature extraction. The research uses Stacked Long Short-Term Memory (SLSTM) for disease classification. Sequential data analysis is SLSTM's forte. To improve the performance, the SLSTM’s hyperparameters are fine-tuned utilising the Enhanced Symbiotic Organism Search (ESOS). The study compares the performance of the APLDD-ESOSDL algorithm against the existing state-of-the-art systems to prove the former’s superiority. The method improves disease diagnosis accuracy and effectiveness.

In the current study, a novel APLDD-ESOSDL technique has been presented for accurate detection and classification of plant leaf diseases. The aim of the given technique is to categorize the presence of leaf diseases properly. The proposed technique encompasses several stages of operations. The SLSTMs are ideal for sequential data problems like time series and image sequences. Plant leaf disease diagnosis involves a series of changes in leaf symptoms. The SLSTMs can capture temporal relationships, which helps in disease detection process. LSTMs and other deep learning models excel at image classification. The capacity to automatically learn and extract the features from image data can help in identifying tiny leaf disease changes. Traditional machine learning methods may struggle to capture complicated leaf disease patterns and linkages. As deep learning models, the SLSTMs can represent complex data relationships. ESOS is a metaheuristic optimisation technique that functions on the basis of natural symbioses. This technique is specialized in efficient exploration of the parameter spaces and the detection of optimal or near-optimal solutions. When used for hyperparameter adjustment, ESOS can improve the parameters of the model. Preventing overfitting in deep learning models requires hyperparameter adjustment. ESOS helps in identifying the hyperparameters that balance the model’s complexity with generalisation. The customization of ESOS towards specific optimisation aims makes it suitable for plant leaf disease diagnostics. For noise reduction and image edge preservation, bilateral filtering is an efficient option. BF can improve the plant leaf picture data for analysis, even with noise or flaws. BF enhances the essential visual elements and hides the unimportant ones. This can assist the model focus on disease-related features and reduce unnecessary image artefacts. Pre-processing methods like BF can improve the model’s performance by cleaning and informing the input data. These methods are chosen because they can handle sequential data, optimise the model’s performance, and preprocess the images for plant leaf disease identification. The researchers must provide detailed arguments in their study to confirm these design decisions and establish its efficacy compared to other methodologies. Figure 1 illustrates the workflow of the APLDD-ESOSDL approach.

1.png

Figure 1. Workflow of the APLDD-ESOSDL approach

Many task-specific factors have been studied for Inception ResNet-v2 model based feature extraction. This application benefits from unique architecture and computer vision. The essential reasons to use the inception ResNet-v2 model are listed below.

1. Beginning deep residual link: Architecture Inception from GoogLeNet helps in ResNet-v2 model’s reuse connections. The deep hybrid model captures both low- as well as high-level traits. This depth is needed to extract the correct patterns and structures from complex multi-scale medical images.

2. Multi-scale feature extraction: Inception ResNet-v2 model have varied convolutional filter sizes. The multi-scale approach accounts for medical imaging variability by collecting the characteristics at multiple granularities. The model should have adaptability since it should work with healthcare data images of various resolutions, orientations, and textures.

3. Inception in computer vision: ResNet-v2 governs the classification process. Classification, object recognition, and segmentation are handled for varied image datasets. The proposed model extracts task-relevant information and executes differentiation.

4. Transfer learning inception: Transfers the learning outcomes from the ImagesNet ResNet-v2. The training data and representations are used with pre-trained models. When updated with the medical imaging dataset, the model responds quickly and saves data from processing.

5. Inception: efficiency and scalability: Model complexity and computational efficiency are balanced by ResNet-v2 model while the computationally-light healthcare software benefits from the deep design of the model. Scalability lets us to add features or alter the models.

3.1 Image pre-processing

To improve the quality of plant leaf disease images, Bilateral Filtering (BF) technique is used. BF is a noise removal technique inspired by Gaussian filtering (GF) technique [19]. But, by effectually preserving the denoising ability of GF, it efficiently reduces the additive noise without ending the specifics and edges of the images. The edge data of the images are maintained though the impulse noise cannot be separated.

3.2 Feature extraction process

For effectual extraction of the features, the Inception ResNet-v2 model is used. Transfer Learning (TL) is a branch of ML that proposes data transmission in a source model for objective function by utilising the correlations in parts, models, or outcomes [20]. For the TL model, the Inception‐Resnet‐V2 structure with pre-training weighted is utilised. The network model can learn rich element representations for different images. Different sizes of convolution filters and residual networks can be combined in the Inception‐Resnet block. The stimulus to choose this structure was based on the experimental outcomes and comparison study with other DL approaches. As the model is considered to be used in edge platform, other resource‐extensive, bulky methods cannot be selected.

The residual networks offer model shortcuts, thus permitting this structure to achieve even more superior outcomes. It is a hybrid of residual and inception blocks, which enhances the efficiency. In order to improve the efficiency of the trained model, the Inception accomplishes optimum utilisation of the computational resources and removes further features with similar computation counts. The resultant of the previous layer is integrated with the network after which the computation of the 5×5 convolutional kernel gets large. The 1×1 convolutional is utilised in the Inception element for two reasons: the primary reason is to overlap more convolutions on receptive domains of similar scale to derive rich features in the constellation diagram. The second reason is to reduce the computational cost and measurement. The 1×1 convolutional layer of the network that derives before 3×3 and 5×5 convolutional layers is utilized for reducing the dimensionality issues. This method is determined to strike a right balance between overfitting and underfitting. The dense layers utilize the ReLU activation function, which is mathematically determined utilizing the subsequent formula:

$R(Z)=\max (0, Z)$        (1)

The FC layer is exchanged with global average pooling layer excluding the inception model’s elements. It is complete to reduce the variable counts. Batch Normalization (BN) also forms a part of the network procedure. The BN layer creates the entire mini‐batch constellation maps that are similar to it moving forward to NN layer, thus avoiding the gradients from disappearing problem.

It is a group of constellations that assists the trained ground to provide some of the constellation. The BP technique also computes the Jacobians. These are easily partial derivations of norms for the variables $a$ and $\chi$.

$\frac{\sigma \operatorname{Norm}(a, \chi)}{\partial a}$ and $\frac{\sigma \operatorname{Norm}(a, \chi)}{\partial \chi}$     (2)

In this network, Adam optimization algorithm is utilized for maximizing the network parameter and minimizing the loss. This model is highly effective if the functioning of the model has huge issues with several data or parameters. It is effectual and needs minimal memory.

$\theta_x:=\theta_{x-1}-\alpha \cdot \frac{\widehat{m}_x}{\sqrt{v} x+\varepsilon}$    (3)

At this point, $\alpha \in R$ and $\theta, \widehat{m}_t, \hat{v}_t, e \in R^n$ for some $n$.

To this day, dropout's regularisation feature is useful for avoiding overfitting issues. In the majority of instances, the dropout may be seen as a sign that a neuron of NN is disabled during training with a certain probability p. The following formula demonstrates the following method for calculating the dropout for the probability $p_i(1 \leq i \leq t)$.

$E_R=\frac{1}{2}\left(x-\sum_{i=1}^t p_i w_i I_i\right)^2+\sum_{i=1}^t p_i\left(1-p_i\right) w_i^2 I_i^2 a$      (4)

3.3 Plant leaf disease classification model

In order to accurately identify and categorise diseases affecting plant leaf tissue, the SLSTM model is employed. The SLSTM model is an LSTM stack, which consists of many stacked layers. Hence, the subsequent layer takes its input from the previous layer's LSTM output [21]. The LSTM model makes use of three gates. Each of the three types of gates input, output, and forget has a single cell state. The LSTM receives as inputs in each time-step t the hidden state $h_{t-1}$ and the cell state $c_{t-1}$, which were created in the time-step before t-1. The input t-1, is obtained in the following way at the current time t:

$h_t, c_t=f_{L S T M}\left(u_t, h_{t-1}, c_{t-1}\right)$        (5)

An explanation of the internal structure of the LSTM model's memory cells and gates is provided below.

$i_t=\sigma\left(W_i u_t+R_i h_{t-1}+b_i\right)$    (6)

$f_t=\sigma\left(W_f u_t+R_f h_{t-1}+b_f\right)$        (7)

$o_t=\sigma\left(W_o u_t+R_o h_{t-1}+b_o\right)$           (8)

$z_t=\tanh \left(W_z u_t+R_z h_{t-1}+b_z\right)$          (9)

$c_t=i_t \odot+f_t \odot c_{t-1}$        (10)

$h_t=0_t \odot \tanh \left(c_t\right)$      (11)

A bias vector is represented by the letter b, the input weight matrix that was learnt during training by the letter W, and the recurrent weight matrix by the letter R in this context. Furthermore, the sigmoid function, denoted by the equation $(x)=1 / 1+\exp (x)$, is also represented by the symbol σ. On top of that, it compresses the input to a range of values between zero and one, thanks to its squashing effect. To avoid a drastic rise in value over time, the hyperbolic tangent function Tanh generates values between -1 and 1, thereby limiting the range of possible values. Finding the values of these two functions requires a computation that takes each component into account. The symbol $i_t$ represents the input gate, $o_t$ represents the output gate, and $f_t$ represents the forget gate. The output of the sigmoid function is used to calculate the linear projections. These projections are then added to the linear projections of $u_t$ and $h_{t-1}$. The forget, input, and output gates are utilised to modulate the cell value of the prior state $c_{t-1}$, the input conversion $z_t$, and the component-wise multiplication output $\odot$, respectively. By utilising a three-layer stacked LSTM architecture, the output of the third layer is directly fed as input to the second layer, while the output of the first layer serves as the input for the second layer. In order to obtain the concealed states and initial cell, we compute the mean of the collections of the vector p. For the first LSTM layer, they are $c_0^1$ and $h_0^1$. For the second LSTM layer, they are $c_0^2$ and $h_0^2$. Finally, for the third LSTM layer, they are $c_0^3$ and $h_0^3$. Two different Multi-Layer Perceptrons (MLP) are used to process these evaluations; they are represented as $f_{i n i t_{-C}}^i$ and $f_{i n i t_{-} h}^i, i$ respectively, where $i=1,2,3$:

$h_0^i=f_{i n i t_{-} h}^i\left(\frac{1}{L} \sum_i^L a_i\right)$          (12)

$c_0^i=f_{i n i t_{-C}}^i\left(\frac{1}{L} \sum_i^L a_i\right)$         (13)

The annotation vectors $a_i$ where i ranges from 1 to L, represent the characteristics associated with certain sub-regions of the picture. In addition, an initial layer of LSTM uses a multi-layer LSTM stack. Figure 2 depicts the structure of the SLSTM technology.

3.4 Parameter tuning process

For the optimal selection of the hyperparameters related to the SLSTM model, the ESOS algorithm is used. SOS creates a population initialization of the organisms in which all the organisms are created randomly in a predetermined search space [22]. During the iteration or generation process, this algorithm exploits parasitism, mutualism, and commensalism to find the optimum global solution.

2.png

Figure 2. Structure of SLSTM

3.4.1 Mutualism phase

The term "mutualism" is used to describe symbiotic partnerships where two species work together and both parties benefit from the association. From the population, two distinct organisms, $X_i^g$ and $X_{R_1}^g$, are chosen at random. Here we have the tangible illustration. Two new species are born from the mating of the two animals, as seen below.

$X_{\text {inew }}^g=X_i^g+\operatorname{rand}[0,1]\left(X_{\text {best }}^g-B P 1 * M v\right)$      (14)

$X_{R_1 \text { new }}^g=X_{R_1}^g+\operatorname{rand}[0,1]\left(X_{\text {bes }}^g-B F 2 * M v\right)$      (15)   

$M v=\frac{X_i^g+X_{R_1}^g}{2}$       (16)

In this context, $X_{\text {best }}^g$ represents the most superior individual organism in the current population. The term rand [0,1] refers to a randomly generated value within the range of 0 to 1. BF1 and BF2 are variables that represent the benefit factors of all organisms. These factors are randomly generated as either 1 or 2, depending on whether the benefits during the interaction are full or partial. Afterwards, $X_{\text {inew }}^g$ and $X_{R_1 \text { new }}^g$ are calculated using the primary function. These values are then compared against $X_i^g$, and $X_{R_1}^g$ in order to identify the most optimal organisms in each pair.

3.4.2 Parasitism phase

One kind of symbiotic interaction occurs when one organism gains an advantage over another, while the other organism has negative consequences. The present population or environment is used to randomly choose $X_i^g$ and $X_{R_1}^g$ to start. Furthermore, $X_{\text {parasite }}^g$ stands for an organism that is present in the population or environment in the form of a copy. The subsequent creation of a parasitic organism is the responsibility of $X_{\text {parasite }}^g$. The host organism for the parasite is represented by $X_{R_1}^g$. The $X_{\text {parasite }}^g$. modifies some of its components at random so that it can affect the host.

$X_{\text {parasite }}^g=L+\operatorname{rand}[0,1](U-L)$         (17)

If $X_{\text {parasite }}^g$ has a higher objective function value than the host organism $\left(X_{R_1}^g\right)$, then $X_{\text {parasite }}^g$ can replace the host organism in the present environment. Algorithm 1 depicts the common architecture of the SOS algorithm.

In the ESOS algorithm, the chaotic optimizer technique is utilized for initializing the population. The chaotic logistic mapping system decides the initialized population. Chaotic modification support the prevention of local minima by presenting arbitrary perturbation as the search model. This perturbation supports the model for exploring the searching space regions, which it could not be searched. Chaos is established as the searching procedure to make this technique highly possible to determine the global minima and less possibly to get stuck in local minima. The mathematical equation of the chaotic logistic mapping is as follows:

$y(1)=\operatorname{rand}, y(i+1)=r y(i)(1-y(i))$         (18)

where, rand denotes the random integer that lies in the range of 0 and $1, r$ signifies the specific variable of the logistic map with a value of 4. Next, the population is initialized using the subsequent formula, assuming that $(L B)$ lower bound and $(U B)$ upper bound of the optimizer parameter.

$P(i)=y(i) \cdot(U B-L B)+L B$        (19)

In Eq. (19), the counter $i$ counts from 1 to $N$ and $P$ signifies the place.

Fitness Function is a vital feature in the ESOS algorithm. An encoder result can be employed to assess the aptitude of the candidate result. At this point, accuracy is the major factor exploited to scheme a fitness function.

Fitness $=\max (P)$          (20)

$P=\frac{T P}{T P+F P}$     (21)