Geese jellyfish search optimization trained deep learning for multiclass plant disease detection using leaf images

Abstract

Accurate and early detection of plant disease is significant for stable and proper agriculture and also for preventing the unwanted waste of financial and other possessions. Hence, a new technique is devised in this work, where geese jellyfish search optimization trained deep learning is used for multiclass detection of plant disease utilizing plant leaf images. At first, the input leaves of the plant image acquired from the database are pre-processed utilizing the Kalman filter. Then, the plant leaf segmentation is done by LinK-Net, where the training function of LinK-Net is processed by the proposed geese jellyfish search optimization, which is formed using wild geese migration optimization and jellyfish search optimizer. Then, image augmentation is carried out and then the feature extraction is done. Consequently, the classification of plant leaf type is processed, which is employed by Deep Q-Network (DQN), which is structurally adapted by the proposed geese jellyfish search optimization. At last, multi-label plant leaf disease is detected based on DQN. Moreover, the proposed geese jellyfish search optimization based DQN obtains an accuracy of 89.44%, true positive rate of 90.18%, and false positive rate of 10.56% respectively.

Keywords

Geese jellyfish search optimization wild geese migration optimization jellyfish search optimization LinK-Net deep Q-network

1. Introduction

Agriculture is only one of the main segments that contributes to the improvement of the Indian economy and is called the backbone of the Indian economy. At most 50% of Indian people’s occupation is based in the cultivation sector. To increase the outcome of agriculture, it is essential to identify the key information for plant disease so that the damage to harvest can be prevented [1]. The kinds of disease must be identified to avoid the spreading of the disease in the agricultural sector. In the early days, the disease is identified based on a visual examination of plant leaves’ symptoms, which creates a high degree of complexity in plant disease detection [2]. Misidentification of crop disease leads to the utilization of incorrect insecticides, which leads to a decrease in the nutrition efficiency in the agricultural sector [3]. The identification and solution of infections are necessary for enhancing the yield and growth of crops. An intellectual plant leaf disease identification technique is necessary to examine agricultural areas frequently [4]. Nowadays, some plant leaf disease identification techniques are devised for automatic plant disease identification utilizing Artificial Intelligence methods with low human participation [5, 6]. Based on the development of Artificial Intelligence, computer vision is required to integrate the facilities in the agriculture sector [7]. In agriculture management, the farmers are offered specialist knowledge depending upon a decision support system [8, 9]. The income and lifestyle of the farmers can be improved based on the quantity and superiority of the harvest. Both human and animal life depends on the plants [10].

The protection of plants from diseases plays an important part in satisfying the rising requirement for quality and quantity of food [11]. According to the Indian Council of Agricultural Research, more than 35% of crop production is affected by disease and pests [12]. Well-timed and correct identification of plant diseases is very important for sustainable and accurate agriculture [13]. The analysis of the disease of the plant leaf is not suitable for the revisions of the visibly perceptible outer layer supposed on the plant. Disease recognition and health observation on plants are truly serious for manageable cultivation [1]. The disease of the plant leaf is detected based on the image is a freshly investigated area by more researchers. Crop disease detection assists the farmers to learn about the cause of plant infections and get the prevention methods to treat the disease. Initial detection of plant disease depends on the color of the leaf, the size of the leaf, the growth of the pattern, and so on it helps farmers [7]. Some diseases require advanced examination techniques when the symptoms are not visible. However, a major amount of plant diseases portray visible symptoms and an experienced plant pathologist detects the disease via optical surveillance of disease-affected plant leaves [13]. However, the physical observation of plant disease is very difficult and it requires an unbelievable amount of endeavour plant disease detection requires too much computing period, so image-based treatment is used for the detection of plant disease [1].

Many conventional Machine Learning methods are devised for the classification and detection of plant diseases [13]. A large amount of research has been devised for crop disease identification and classification. The capability of machine learning techniques to solve real-time challenging issues is an incredible part of the intelligence of the human brain. The main problems of machine learning-based methods over the detection of crop disease are low efficiency and huge computation time as these frameworks create long lengthy codes, which increases the complexity of computation. In addition, the difference in brightness during the capturing of plant leaves image also complicates the identification method [14]. The utilization of deep learning methods in the detection and categorization of disease from images in medical is quite common. On another hand, deep learning-based studies for the classification and detection of plant diseases are utilized in numerous studies. But, still, it has some gaps in the examination regarding the utilization of particularly novel deep learning structures in the disease of plant leaves. Particularly, the necessity for effective methods with minimal parameters that can be trained quickly and without compromise on presentation is unavoidable [13]. Hyperparameters are the most important training parameters that can impact the process of deep learning methods. Random search and grid search are the most common hyperparameter tuning techniques in deep learning. High presentation processing power is required to train the deep learning algorithms with better efficiency and low execution time [15, 6].

The effective plant leaf disease classification is a primary contribution of this research paper and it is processed as follows. At first, the plant leaf image is taken from a plant village sampleset and it is preprocessed by the Kalman filter. Afterwards, the preprocessed image is subjected to the segmentation phase, where segmentation is employed by Link-Net, The weights and bias of the Link-Net are trained by the proposed geese jellyfish search optimization, which is the incorporation of geese migration optimization and jellyfish search optimizer. Consequently, image augmentation is performed and the augmentation applied image is allowed to feature extraction performance. Afterwards, plant-type classification is done by Deep Q-Network (DQN), which is trained by the proposed geese jellyfish search optimization. Finally, multiclass plant leaf disease is identified based on DQN, and it is trained by the proposed geese jellyfish search optimization.

The main key contribution of this paper is explained below, proposed geese jellyfish search optimization based DQN for disease detection of plant leaf: Here, the devised geese jellyfish search optimization based DQN is applied for multiclass plant disease detection, where geese jellyfish search optimization is the incorporation of geese migration optimization and jellyfish search optimizer. In this paper, Link-Net trained by geese jellyfish search optimization is employed for segmentation. Then, DQN trained by geese jellyfish search optimization is utilized for plant leaf type classification and detection.

The rest portion of this study is organized as below: Part 2 depicts a literature review of traditional techniques with their limitations and advantages. Part 3 indicates the proposed geese jellyfish search optimization based DQN and its performance, Part 4 signifies the proposed scheme’s outcome and experimental study and the work is concluded in Part 5.

2. Literature review

Atila et al. [13] devised an EfficientNet deep learning structure for the identification of disease-affected plants. This technique achieved more effective outcomes by uniformly detecting width, resolution, and scaling depth, although this technique had maximum execution time. Ahmed and Reddy [3] developed a Convolution Neural Network (CNN) for the detection of crop disease symptoms and this algorithm performed superior results with less computation complexity, but this technique did not perform well on leaf images with complex backgrounds. Albattah et al. [14] introduced the DenseNet-77 for the detection of plant disease. Here, the identification of disease was more effective and the method identified the disease location of the plant leaf but this technique was easily affected by the overfitting problem. Vallabhajosyula et al. [16] devised the Deep Ensemble Neural Networks (DENN) for the identification of plant disease. This algorithm decreased training function and convergence time and tuned the huge amount of parameters effectively, although this technique failed to process some parameters, such as epochs, learning rate, activation functions, etc. to enhance efficiency. Pandian et al. [6] devised the Deep CNN for the detection of plant disease. Here, the overall efficiency of the technique was improved, because of the use of rotation, flipping, and cropping. However, this technique failed to identify the disease in the other portions of a plant like fruits, stems, and flowers. Chouhan et al. [10] developed the Internet of Things – Fuzzy Based Function Network for the detection of plant disease. Here, the computational cost and computational power of the architecture were minimal, but it failed to incorporate technologies like big data, the Internet of Things, GPS, Unmanned Aerial Vehicle, devices like sensors cameras, etc. Mahum et al. [17] developed an Efficient DenseNet for detection of plant leaf disease and this technique was flexible enough and can be simply fine-tuned. But this technique was unfeasible, because of high execution time. Pandian et al. [18] introduced the Deep CNN for plant disease detection and this technique was also performed in the real-time dataset, but this technique had a high computational cost.

3. Proposed geese jellyfish search optimization based Deep Q-Network for plant disease detection

The plant leaf type and disease classification is a major goal of this study, where the proposed geese jellyfish search optimization based Deep Q-Network (DQN) was introduced for the classification of plant leaves. The input leaf image is taken from the dataset and it is allowed to a preprocessing phase. In preprocessing, the Kalman filter is used to remove an error from the input. Then, the plant leaf segmentation is processed by Link-Net [19] which is trained by the proposed geese jellyfish search optimization and it is designed by the integration of geese migration optimization [20] and jellyfish search optimizer [21]. Afterwards, the segmented output is allowed for image augmentation, where the data sample is increased based on translation, resize, contrast, resize, hue, and saturation. Then, feature extraction is performed, wherein features are Complete Local Binary Pattern (CLBP) features [22] and statistical features [23] are processed. Here, classification is based on first-level and second level classification. At first, plant leaf type is classified using DQN [24], and DQN is trained by the proposed geese jellyfish search optimization developed by the combination of geese migration optimization [20] and jellyfish search optimizer [21]. Finally, plant leaf disease is identified, which is done by DQN, which is trained by the proposed geese jellyfish search optimization under the integration of geese migration optimization [20] and jellyfish search optimizer [21]. Figure 1 signifies the structural diagram of the proposed geese jellyfish search optimization based DQN.

Figure 1.

Structural diagram of the proposed geese jellyfish search optimization based deep Q-network for plant disease detection.

3.1 Plant leaf image acquisition

Consider the plant leaf image as input and it is accumulated from a dataset $M$ , which is designed as,

$\displaystyle M=\{m_{1},m_{2},\ldots m_{e},\ldots m_{h}\}$ (1)

where $m_{e}$ represents the $e^{\text{th}}$ image and $M$ denotes the database and the overall amount of image is detailed by $h$ , and the plant leaf image $m_{e}$ is allowed to consequent preprocessing phase.

3.2 Preprocessing of the plant leaf image

The input plant leaf image $m_{e}$ is subjected to a preprocessing step, where preprocessing is carried out by the Kalman filter [25]. The plant leaf image may have unwanted noise and errors, which are eliminated in this step. Moreover, this step improves the quality of the plant leaf image and the output of this step is denoted as $B$ .

3.2.1 Kalman filter

For using the Kalman filtering [25] technique, disregard the organized inputs and process noise, the state and its interrelated covariance matrix for the following iteration, which is achieved optimum solution. The advantage of the Kalman filter is preserved low power and it is computationally effective. The Kalman filter is designed by,

$\displaystyle a^{(b+1)}=C^{(b)}\hat{a}^{(b)}$ (2)

where $a^{(b)}$ indicates the $b^{\text{th}}$ iteration mean value, $C^{(b)}$ depicts the transition matrix, and $\hat{a}^{(b)}$ signifies the optimum mean value.

3.3 Plant leaf image segmentation

The pre-processed plant image $B$ is subjected to plant leaf segmentation, where leaf segmentation is carried out by Link-Net [19]. Using Link-Net, the segmentation process can be performed effectively in real-time applications. The Link-Net is trained by the proposed geese jellyfish search optimization, which is formed by the integration of geese migration optimization [20] and jellyfish search optimizer [21], and the segmentation process is employed to segregate the plant leaf area from the entire image. The outcome of plant leaf segmentation is depicted as $A_{d}$ .

3.3.1 Structure of Link-Net

The structure of Link-Net [19] is portrayed in Fig. 2, where the input of Link-Net is $B$ and here full conv indicates full convolution. The Link-Net has an encoder-decoder structure, where each convolution layer has Batch Normalization and it is continued by the ReLU non-linearity layer. The initial block of the encoder is a convolution layer and its kernel has a size 7X7 with a stride of two. The encoder block also processes spatial max-pooling in an area of 3X3 with a 2 stride. The same process is performed in the decoder section. Every input of an encoder is bypassed to a corresponding output of the decoder. The major goal of this network is to discover the loss the spatial information, where upsampling is used at the decoder. Moreover, the decoder distributes the information learned through the encoder at each layer, and fewer parameters are used in the decoder section. The output of the Link-Net is denoted as $A_{d}$ .

Figure 2.

Structure of Link-Net.

3.3.2 Structural optimization of geese jellyfish search optimization

The plant leaf segmentation is carried out by Link-Net, which is structurally adapted by geese jellyfish search optimization, which is designed by integration of geese migration optimization [20] and jellyfish search optimizer [21]. To improve the segmentation accuracy both the algorithm is combined and used to train the Link-Net. The geese migration optimization [20] is simple and a minimum number of parameters are used. It performed excellently in terms of computation cost and computational power. The jellyfish search optimizer [21] is utilized to solve the structural optimization issue efficiently. An efficient tradeoff between intensification and diversification assists in decreasing the cost of computation and employs effective optimization. The below steps explain the realization of the geese jellyfish search optimization.

1) Population initialization

Normally, the population is initialized randomly in any optimization algorithm, and it endures local optima. The population initiation is formulated as,

$\displaystyle D_{c+1}=\alpha D_{c}(1-D_{c}),0\leqslant D_{c}\leqslant 1$ (3)

where the position of $c^{\text{th}}$ jellyfish is indicated as $D_{c+1}$ , $D_{c}$ is represented as the map of logistic chaos of the position of the $c^{\text{th}}$ jellyfish, which is utilized to create the jellyfish’s initial population, and the value of $\alpha=4.0$ .

2) Evaluation of fitness solution

The fitness solution is estimated based on the Mean Square Error (MSE), which is designed as,

$\displaystyle\Delta=\frac{1}{f}\sum\limits_{d=1}^{f}[A_{d}^{\ast}-A_{d}]^{2}$ (4)

where the whole amount of training sample is denoted as $f$ , the expected outcome is signified as $A_{d}^{\ast}$ , and $A_{d}$ depicts the Link-Net output.

3) Ocean current

Huge amounts of nutrients are contained in ocean current, so jellyfish are intent by it, where the current best position is determined based on the ocean current direction by all vectors averaging and it is formulated by,

$\displaystyle\overrightarrow{\textit{trend}}=D^{\ast}-\gamma\times\textit{rand% }^{\prime\prime\prime}(0,1)\times\vartheta$ (5)

where the current best position of the jellyfish swarm is represented as $D^{\ast}$ , jellyfish mean location is signified as $\vartheta$ , and $\gamma>0$ , and a new position of every jellyfish is formulated by,

$\displaystyle D_{c}^{(g+1)}=D_{c}^{g}+\textit{rand}^{\prime\prime\prime}(0,1)% \times\overrightarrow{\textit{trend}}$ (6)

Applying Eqs (5) and (6),

$\displaystyle D_{c}^{(g+1)}=D_{c}^{g}+\textit{rand}^{\prime\prime\prime}(0,1)% \times D^{\ast}-\gamma\times\textit{rand}^{\prime\prime\prime}(0,1)\times\vartheta$ (7)

where $D_{c}^{g}$ denotes the current location of the jellyfish, $g$ denotes the iteration number and $D_{c}^{(g+1)}$ indicates the new position of the jellyfish.

To improve convergence speed geese migration optimization [21] is added to jellyfish search optimizer, and from the geese migration optimization,

$\displaystyle D_{c}^{(g+1)}=D_{c}^{g}+E_{3}(D_{f}^{g}-D_{c}^{g}+F)+E_{4}(D_{% \textit{best}}^{g}-D_{f}^{g})$ (8) $\displaystyle D_{c}^{g}=\frac{D_{c}^{g+1}-(E_{3}-E_{4})D_{f}^{g}-E_{3}F-E_{4}D% _{\textit{best}}^{g}}{(1-E_{3})}$ (9)

Substituting Eq. (10) in Eq. (6),

$\displaystyle D_{c}^{g+1}=\frac{D_{c}^{g+1}-(E_{3}-E_{4})D_{f}^{g}-E_{3}F-E_{4% }D_{\textit{best}}^{g}}{(1-E_{3})}+\textit{rand}^{\prime\prime\prime}(0,1)\ast D% ^{\ast}-\gamma\textit{rand}^{\prime\prime\prime}(0,1)\ast\vartheta$ (10) $\displaystyle\frac{(1-E_{3}-1)D_{c}^{g+1}}{1-E_{3}}=\frac{\begin{array}[]{c}-(% E_{3}-E_{4})D_{f}^{g}-E_{3}F-E_{4}D_{\textit{best}}^{g}\\ {}+(\textit{rand}^{\prime\prime\prime}(0,1)\ast D^{\ast}-\gamma\textit{rand}^{% \prime\prime\prime}(0,1)\ast\vartheta)(1-E_{3})\end{array}}{(1-E_{3})}$ (11) $\displaystyle D_{c}^{g+1}=\frac{\begin{array}[]{c}(E_{3}-E_{4})D_{f}^{g}+E_{3}% F+E_{4}D_{\textit{best}}^{g}\\ {}-(\textit{rand}^{\prime\prime\prime}(0,1)\ast D^{\ast}-\gamma\textit{rand}^{% \prime\prime\prime}(0,1)\ast\vartheta)(1-E_{3})\end{array}}{E_{3}}$ (12)

where $\textit{rand}^{\prime\prime\prime}(0,1)$ denotes a random number, $E_{3}$ , $E_{4}$ indicates the random number (0, 1), and $F$ signifies radius of group range, $D_{f}^{g}$ indicates excellent head goose, and $D_{\textit{best}}^{g}$ signifies individual of global optimal.

4) Swarm of jellyfish

Jellyfish swarm undertakes active (Type A) and passive (Type B) motions, where mostly jellyfish signify type A motion, and the consequent upgraded position of every jellyfish is expressed as,

$\displaystyle D_{c}^{g+1}=D_{c}^{g}+\omega\times\textit{rand}^{\prime\prime% \prime}(0,1)\times(G-H)$ (13)

where $G$ represents an upper bound and $H$ denotes a lower bound of the search space, and $\omega=0.1$ .

Considering the Type B motion, where the displacement of efficient exploitation of a local search space is expressed as,

$\displaystyle\overrightarrow{\textit{Step}}=\textit{rand}^{\prime\prime\prime}% (0,1)\times\overrightarrow{\textit{Direction}}$ (14) $\displaystyle\overline{\textit{Direction}}=\left\{{\begin{array}[]{ll}D_{f}(g)% -D_{c}(g)&\text{if }l(D_{c})\geqslant l(D_{f})\\ D_{c}(g)-D_{f}(g)&\text{if }l(D_{c})<l(D_{f})\\ \end{array}}\right.$ (15)

Hence the updated position is given by,

$\displaystyle D_{c}(g+1)=D_{c}(g)+\overrightarrow{\textit{step}^{\prime\prime}}$ (16)

where the objective function of the position $D$ is represented as $l$ .

5) Mechanism of time control

This situation introduced the mechanism of time control, which normalizes the movement between the jellyfish and ocean current depend on inside a swarm of jellyfish, the time control function $n(g)$ is formulated as,

$\displaystyle n(g)=\left|{\left({1-\frac{g}{\textit{Maxiter}}}\right)\times(2% \times\textit{rand}^{\prime\prime\prime}(0,1)-1)}\right|$ (17)

where Maxiter represents the iteration of the maximum number, which is parameter initialization.

6) Boundary condition

Around the world, the ocean is located and the earth is roughly spherical, so jellyfish move outside the bounded search area and return to the opposite bound, and this is modelled as.

$\displaystyle X\left\{{\begin{array}[]{ll}{D}^{\prime\prime}_{c,h}=(D_{c,h}-G_% {h})+H(h)&\text{if }D_{c,h}>G_{h}\\ {D}^{\prime\prime}_{c,h}=(D_{c,h}-H_{h})+G(h)&\text{if }D_{c,h}<H_{h}\\ \end{array}}\right.$ (18)

$D_{c,h}$ is the position of $c^{\text{th}}$ jellyfish in $h^{\text{th}}$ dimension, ${D}^{\prime\prime}_{c,h}$ is the upgraded position after confirming boundary condition, $H(h)$ and $G(h)$ is an upper and lower bound of $h^{\text{th}}$ dimension in a search space.

7) Feasibility estimation

The position of every population is updated due to the roles of a proposed technique and the new position is evaluated.

8) Termination

The phases are repetitively executed till the highest iteration is achieved.

3.4 Augmentation of image

Image augmentation is a function of creating novel transformed versions of images from the acquired image dataset to raise its diversity, wherein augmentation techniques like resizing, shearing, translation, contrast, saturation, and hue are performed [26, 27]. Here, the outcome of segmentation $A_{d}$ is fed to an image augmentation and Q is denoted as output.

1) Resizing

Resize [26] is an image augmentation technique, which is used to adjust the longest length of the edge, which maintains the aspect ratio of the input image also the shortest edge is modified as per the required size, and the resized output is denoted as $q_{1}$ .

2) Shearing

Shearing [27] is a process of modifying the image direction along the x and y directions, which contains the two phases. The component in the x-axis is the first phase and the component in the y-axis is the second phase and the following expressions are explained in the x and y direction shearing.

$\displaystyle\left[{\begin{array}[]{l}p_{x}^{\prime}\\ p_{y}^{\prime}\\ \end{array}}\right]=\left[{\begin{array}[]{ll}1&\textit{sher }X^{\prime\prime}% \\ 0&1\\ \end{array}}\right]\left[{\begin{array}[]{l}{x}^{\prime}\\ {y}^{\prime}\\ \end{array}}\right]$ (19) $\displaystyle\left[{\begin{array}[]{l}p_{x}^{\prime}\\ p_{y}^{\prime}\\ \end{array}}\right]=\left[{\begin{array}[]{ll}1&0,\\ \textit{sher }Y^{\prime\prime}&1\\ \end{array}}\right]\left[{\begin{array}[]{l}{x}^{\prime}\\ {y}^{\prime}\\ \end{array}}\right]$ (20)

where after shearing new location of every pixel is denoted as $p_{x}^{\prime}$ , $p_{y}^{\prime}$ , ${x}^{\prime}$ and ${y}^{\prime}$ indicates the image coordinates, and $\textit{sher }{X}^{\prime\prime}$ , $\textit{sher }{Y}^{\prime\prime}$ is indicates the shearing of $X$ and $Y$ direction, and the outcome is denoted as $q_{2}$ .

3) Translation

The movement of an object from horizontal and vertical directions in the image is called translation, where a geometric image is translated based on black or white. The translation can be performed based on X and Y directions at the same time [27]. To avoid the locational weights in data, the picture can be translated in left, down, right, and up directions and the equation of translation is expressed as,

$\displaystyle\left[{\begin{array}[]{l}p_{x}^{\prime}\\ p_{y}^{\prime}\\ \end{array}}\right]=\left[{\begin{array}[]{l}{x}^{\prime}\\ {y}^{\prime}\\ \end{array}}\right]+\left[{\begin{array}[]{l}J_{x}^{\prime}\\ J_{y}^{\prime}\\ \end{array}}\right]$ (21)

where after translation, the new location of every pixel is denoted as $p_{x}^{\prime}$ , $p_{y}^{\prime}$ , ${x}^{\prime}$ and ${y}^{\prime}$ indicates the image’s original position, and $J_{x}^{\prime}$ , $J_{y}^{\prime}$ indicates the translation position of ${x}^{\prime}$ and ${y}^{\prime}$ direction, and the translated image is denoted as $q_{3}$ .

4) Contrast

The separation of dark and white color from an image is indicted contrast, which is used to control the amount of jitter based on the value 0 (no modification) and value 1 (gradually large modification). It specified only the prospective strength of the effect but did not include the higher and lower levels of the image [26]. The contrast adjusted image is denoted as $q_{4}$ .

5) Saturation

The random saturated jitter is added to the image to improve the color quality of the image, which is based on the amount of color in an image [26]. The parameter is utilized to manage the quantity of jitter in saturation based on the value 0 (no modification) and value 1 (gradually large modification), where saturation is specified only the prospective strength of the effect but does not include the higher and lower level of the image, and the outcome is denoted as $q_{5}$ .

6) Hue

Add random hue jitter from the image is hue applied image, which is based on the color shade of the image [26]. To control the quantity of jitter, the parameter hue is used with the value 0 and value 1, and the hue augmented image is indicated as $q_{6}$ .

The images generated by applying various approaches are combined and the overall image produced by the augmentation is expressed as, $Q=\{{q_{1},q_{2},\ldots q_{6}}\}$ . Here, $q_{1}$ indicates the resized image, $q_{2}$ depicts the sheared image, $q_{3}$ represents the translated image, $q_{4}$ signifies the contrast applied image, $q_{5}$ represents the saturation applied image, and $q_{6}$ represents the hue applied image.

3.5 Extraction of features

The image augmentation output $Q$ is subjected to the input of feature extraction, wherein features like Complete Local Binary Pattern (CLBP) [22] and statistical features like mean, variance, standard deviation, entropy, energy, contrast, correlation, and skewness [23]. In feature extraction, the dimension of the image is decreased and the output of feature extraction is denoted as $Z$ .

Complete Local Binary Pattern (CLBP)

To code the local image in a more absolute manner [22], the CLBP is devised for texture classification. The dissimilarity between the neighborhood pixels is denoted as $K_{r}$ .

$\displaystyle K_{r}=L_{r}\ast N_{r}$ (22)

where $L_{r}$ denotes a sign component, $N_{r}$ indicates a magnitude component, and CLBP is expressed as,

$\displaystyle\textit{CLBP}^{\prime\prime\prime}\_N_{O,P}=\sum\limits_{r=0}^{r-% 1}{t(N_{r}-u)}2^{r}$ (23)

where $t$ depicts the mean value of $N_{r}$ , $N_{O,P}$ depicts the neighborhood of $O$ points of the radius $P$ with admiration to the central pixel, and $u$ is represented as a random value, where $\kappa$ is indicted as CLBP.

Statistical features

Here, CLBP output $\kappa$ is fed to the input of statistical features, where in statistical features like mean, variance, standard deviation, entropy, energy, contrast, correlation, and skewness are used [23].

1) Mean

The mean represents the concentration of the data distribution [23] and is expressed as

$\displaystyle s_{1}=\sum\limits_{v=0}^{R-1}{v\ast I(v)}$ (24)

where $s_{1}$ depicts mean value, the whole quantity of grey level image is represented as $R$ , $I(v)$ signifies $v^{\text{th}}$ happening probability and $v$ signifies each image’s grey level.

2) Variance

The variance of the grey-level image is related to the deviation of the grey-level image from the mean and the variance [23] of the grey-level image is formulated by,

$\displaystyle s_{2}=\sum\limits_{v=0}^{R-1}{(v-s_{2})}^{2}\ast I(v)$ (25)

where $s_{2}$ indicates variance.

3) Standard deviation

The standard deviation [23] represents the distribution of data about the mean value, whereas a low standard deviation depicts the values that are many closeups to the mean, and a high standard deviation represents a large dispersion of values among the mean. The standard deviation is formulated as,

$\displaystyle s_{3}=\sqrt{\sum\limits_{v=0}^{R-1}{(v-s_{2})^{2}I(v)}}$ (26)

where $s_{3}$ indicates the standard deviation.

4) Entropy

Entropy [23] is a measurement of information between the segmented image and classified image and it is formulated as,

$\displaystyle s_{4}=\sum\limits_{v=0}^{R-1}{I(v)\ast\log_{2}}[I(v)]$ (27)

where $s_{4}$ depicts the entropy.

5) Skewness

The asymmetrical distribution [23] of a particular feature about the mean is called skewness, which has negative or positive values and is computed as

$\displaystyle s_{5}=s_{2}^{-3}\left[{\sum\limits_{v=0}^{R-1}{(v-s_{1})^{3}\ast I% (v)}}\right]$ (28)

where $s_{5}$ is represented as skewness.

6) Energy

The energy [23] is based on the second-order statistical features, which are formulated by Gray-Level Co-Occurrence Matrix (GLCM) squared components and it is expressed as

$\displaystyle s_{6}=\sum\limits_{v=0}^{R-1}{[I(v)]}^{2}$ (29)

where $s_{6}$ is signified as energy.

7) Contrast

Here, the intensity difference between each pixel and its neighborhood pixel is calculated [23] and is termed contrast and it is expressed as,

$\displaystyle s_{7}=\sum\limits_{x=0}^{R-1}\sum\limits_{w=0}^{R-1}{(x-w)}^{2}$ (30)

where contrast is represented as, $s_{7}$ , the different location of a grayscale image is $x$ and $w$ , summation of kernel indicates as $({x-w})$ , which is equal to 1.

8) Correlation

The pixel value of the image is connected with its neighbourhood [23], which is called correlation and it is expressed as

$\displaystyle s_{8}=\sum\limits_{x=0}^{S-1}\sum\limits_{w=0}^{S-1}\frac{(x-s_{% 1x})(w-s_{1w})}{\sqrt{(s_{2x})(s_{2w})}}$ (31)

where $s_{8}$ is depicted as correlation.

Then, the features are concatenated to form the feature vector and the expression is

$\displaystyle\wp=\{s_{1},s_{2},\ldots s_{8}\}$ (32)

where $s_{1}$ depicts the mean, $s_{2}$ signifies the variance, $s_{3}$ indicates the standard deviation, $s_{4}$ depicts the entropy, $s_{5}$ signifies the skewness, $s_{6}$ represents energy, $s_{7}$ signifies the contrast, $s_{8}$ indicates the correlation, and $\wp$ denotes the overall feature vector.

3.6 Plant leaf type classification

The type of plant leaf is classified by DQN, where feature extraction output $\wp$ is subjected to the input of plant leaf type classification. DQN [24] is trained by geese jellyfish search optimization, which is a designed combination of geese migration optimization [20] and jellyfish search optimizer [21]. The DQN is performed for real-time application and a minimum number of parameters is used for performance. Here, a variety of leaves is examined and detected as, apple, cherry, strawberry, tomato, and potato. The classification of plant leaf outcome is represented as $Z_{y}$ .

3.6.1 Architecture of Deep Q-Network (DQN)

The most commonly used reinforcement learning algorithm is called Q-learning, and itis used as the DQN [24] function also Q-function’s active value function is approximately calculated by CNN. Sometimes Reinforcement learning is represented as an unstable or nonlinear function that approximates divergence and an active value function is used for Neural Network representation. Moreover, DQN consists of convolution, dropout, max-pooling, flattening, and dense layers.

The convolution layer is a simple application of an input filter and has an activation. Here, the frequent application of an identical filter to an input outcome in an activation map is denoted as a feature map, demonstrating the positions and potency of an identified feature in an input, such as an image. Then, the function of the pooling layer is to reduce the dimension of the hidden layer by combining the outputs of neuron clusters at the before layer into a single neuron in the consequent layer. A flattened layer can be utilized to exchange semi-structured information for a representation of relationships. A dense layer is utilized to categorize images based on the outcome from convolution layers. Figure 3 portrays the structure of DQN.

Figure 3.

Structure of Deep Q-Network (DQN).

3.6.2 Training function of Deep Q-Network (DQN) by geese jellyfish search optimization

The type of plant leaf is classified using DQN, trained by geese jellyfish search optimization, and it is designed combination of geese migration optimization [20] and jellyfish search optimizer [21]. A structural adaptation process is expanded in section 3.3.2 and $Z_{y}$ represents the classified outcome of plant leaf disease. The solution of the fitness process is represented below.

$\displaystyle\Delta_{1}=\frac{1}{f}\sum\limits_{y=1}^{f}{[Z_{y}^{\ast}}-Z_{y}]% ^{2}$ (33)

where the accepted outcome is indicated as $Z_{y}^{\ast}$ , and the DQN output is denoted as $Z_{y}$ .

After classifying the plant leaf, the output $Z_{y}$ is fed to the disease detection phase to discover the type of disease affecting the plants.

3.7 Plant disease detection

The plant disease detection is carried out by DQN, where leaf-type classified output $Z_{y}$ is allowed to the input of plant disease detection. DQN [24] is trained by geese jellyfish search optimization, which is a designed combination of geese migration optimization [20] and jellyfish search optimizer [21].

3.7.1 Architecture of Deep Q-Network (DQN)

DQN [24] is utilized for the identification of plant leaf disease and it is detailed in section 3.6.1. Here, plant leaf type classification outcome $Z_{y}$ is allowed to the input of plant disease detection, and output produced is $S_{z}$ . The DQN is briefly explained in section 3.6.1.

3.7.2 Training function of Deep Q-Network (DQN) by geese jellyfish search

The plant disease detection is carried out by DQN, where plant disease classification outcome $S_{z}$ is subjected to an input of plant disease detection. DQN is trained using geese jellyfish search optimization. This process is briefly explained in section 3.6.2 and outcome of multiclass plant leaves disease is signified as $\vec{V}$ . The fitness solution is explained as follows,

$\displaystyle\Delta_{2}=\frac{1}{f}\sum\limits_{y=1}^{f}{[S_{z}^{\ast}}-S_{z}]% ^{2}$ (34)

where $S_{z}^{\ast}$ is the anticipated outcome, and the DQN outcome is indicated as $S_{z}$ .

4. Results and discussion

The results of the proposed geese jellyfish search optimization based DQN for plant leaf disease classification and detection are illustrated in this portion. The evaluation of performance metrics, examination and dataset description, experimental arrangement, and comparative techniques are detailed in this part. Table 1 show the nomenclature used in the proposed work.

Table 1
Nomenclature

Abbreviation	Definition
GJSO	Geese Jellyfish Search Optimization
GMO	Wild Geese Migration Optimization
JSO	Jellyfish Search optimizer
DQN	Deep Q-Network
TPR	True Positive Rate
FPR	False Positive Rate
CNN	Convolution Neural Network
DENN	Deep Ensemble Neural Networks
DCNN	Deep Convolutional Neural Network
IoT_FBFN	Internet of Things- Fuzzy-Based Function Network
GJSO-DQN	Geese Jellyfish Search Optimization based Deep Q-Network
CLBP	Complete Local Binary Pattern

4.1 Dataset description

The proposed geese jellyfish search optimization based DQN for the detection of plant leaf disease is implemented with the Python tool employing the Plant Village dataset [28]. The multilevel classification of plant leaf disease is carried out using images from the Plant Village dataset [28]. In the plant village image, a different version of an image, such as the raw image’s greyscale version, RGB images with colour accustomed, and leaves segmented images are obtained. The total images in the dataset are 13294. Here, plant leaf images contained the total images of plants, like cherry with total image 1906, total apple images 3171, total tomato images of 4500, total image of potato is 2152, and total image of strawberry is 1565.

4.2 Evaluation measures

The efficacy of the proposed geese jellyfish search optimization based DQN is evaluated using metrics like accuracy, True Positive Rate (TPR), and False Positive Rate (FPR).

1) Accuracy

The effectiveness of plant disease detection and classification is examined utilizing accuracy, which is calculated based on the proportion between the correctly detected image to the whole quantity of taken image and it is formulated by,

$\displaystyle\Im=\frac{T_{y}+T_{z}}{T_{y}+T_{z}+U_{y}+U_{z}}$ (35)

where accuracy represents $\Im,T_{y},T_{z},U_{y},U_{z}$ depicts true positive, true negative, false positive, and false negative rate.

2) True Positive Rate (TPR)

TPR determines the number of positive images, which are perfectly obtained out of a large quantity of positive image samples, and is given by,

$\displaystyle\Re=\frac{T_{y}}{T_{y}+U_{z}}$ (36)

where TPR is represented as $\Re$ .

Figure 4.

Experimental outcome of geese jellyfish search optimization based Deep Q-Network.

3) False Positive Rate (FPR)

FPR is formulated by the total amount of normal images classified as positive and the whole amount of normal images, which is expressed by,

$\displaystyle\aleph=\frac{U_{y}}{U_{y}+T_{y}}$ (37)

where $\aleph$ indicates the FPR.

Figure 5.

Comparative examination of geese jellyfish search optimization based Deep Q-Network based on first-level classification.

Figure 6.

Comparative examination of geese jellyfish search optimization based Deep Q-Network based on first-level classification.

Figure 7.

Performance examination of proposed geese jellyfish search optimization based Deep Q-Network based on first-level classification.

Figure 8.

Performance examination of geese jellyfish search optimization based Deep Q-Network based on first-level classification.

4.3 Experimental outcome

Figure 4 indicates the experimental outcome of the proposed geese jellyfish search optimization based DQN for plant disease detection. Figure 4a indicates the input leaf image, Fig. 4b represents the pre-processed image, Fig. 4c shows the plant leaf segmented image, consequently, Fig. 4d–i, show the resized image, sheared image, translated image, contrast applied image, saturated image, and hue applied images, and also Fig. 4j–o indicates the CLBP based resized image, CLBP based sheared image, CLBP based translated image, CLBP based contrast applied image, CLBP based saturated image, and CLBP based hue processed images.

4.4 Comparative analysis

Various comparative methods, like EfficientNet Depp Learning (DL) [13], CNN [3], DenseNet-77 [14], and Deep Ensemble Neural Networks (DENN) [16] are used to examine the effectiveness of the proposed technique for the detection of plant leaf disease. In a comparative analysis, analysis with a training set is employed to reveal the efficacy of multilevel plant leaf classification done by the proposed geese jellyfish search optimization based DQN (GJSO-DQN).

4.4.1 Comparative assessment with training set in the first level of classification

Figure 5 depicts a graph drawn among various presentation metrics and shifting training set percentages depending on first level classification. Figure 5a represents the comparative assessment of the proposed GJSO-DQN for accuracy. When an 80% training set is considered for examination, the accuracy of the proposed GJSO-DQN is 88.94%, and the accuracy attained by the conventional method, like EfficientNet DL is 76.77%, CNN is 78.07%, DenseNet-77 is 83.48%, and DENN is 85.05%. The proposed GJSO-DQN is found to be 4.37% greater than the DENN technique. Figure 5b depicts the FPR value evaluation. By taking a 90% training set, the proposed GJSO-DQN’s FPR value is 10.52% and the FPR value of other techniques, like EfficientNet DL is 20.82%, CNN is 17.50%, DenseNet-77 is 15.16%, and DENN is 12.72%. Here, the FPR of proposed GJSO-DQN is 6.63% lower than CNN. Figure 5c represents the TPR value. While considering 80% of the training set, the TPR value of proposed GJSO-DQN is 88.32% and 76.01% for EfficientNet DL, 79.15% for CNN, 81.86% for DenseNet-77, and 85.84% for DENN. The process improvement of the proposed GJSO-DQN is 7.31% better than that of DenseNet-77.

4.4.2 Comparative assessment with training set in the classification of the second level

Figure 6 portrays the comparative examination of multiple performance metrics and shifting training set percentages depending on second-level classification. Figure 6a shows the estimation of accuracy for the GJSO-DQN. When taking training data is 90%, the accuracy value of 90.18% is achieved by the proposed GJSO-DQN, and the accuracy of other conventional techniques is 79.87% for EfficientNet DL, 83.13% for CNN, 84.79% for DenseNet-77, and 87.58% for DENN. Here, the proposed GJSO-DQN has 7.81% better accuracy than CNN Fig. 6b represents the analysis of the FPR technique. While considering a training set is 80%, the FPR value of proposed GJSO-DQN is 12.75% and other conventional techniques like, EfficientNet DL is 24.02%, CNN is 19.61%, DenseNet-77 is 16.11%, and DENN is 14.97%. Here, the proposed GJSO-DQN is 17.41% lower compared to the DENN method. Figure 6c shows an examination of the TPR value. For 70% of the training set, the TPR value is 85.15% for proposed GJSO-DQN, 73.79% for EfficientNet DL, 75.69% for CNN, 78.65% for DenseNet-77, and 82.66% for DENN. The performance development of proposed GJSO-DQN is 11.10% superior to DenseNet-77.

4.4.3 Performance assessment with training set in classification of first-level

Figure 7 signifies the presentation examination of the proposed GJSO-DQN based on different performance metrics in terms of the training set with several iterations. Figure 7a indicates the performance examination based on accuracy. By taking the 80% training set percentage, the accuracy of the proposed GJSO-DQN with iteration 20 is 76.78%, iteration 40 is 79.06%, iteration 60 is 83.21%, iteration 80 is 85.56%, and iteration 100 is 87.50%. Figure 7b shows the FPR value of the proposed GJSO-DQN with several iterations. For analyzing the 90% training set, 21.97% for 20 iterations, 16.24% for 40 iterations, 14.67% for 60 iterations, 12.77% for 80 iterations, and 10.33% for 100 iterations. Figure 7c portrays the TPR value of the proposed GJSO-DQN with multiple iterations. While taking a 90% training set, the TPR value of the proposed GJSO-DQN with iteration 20 is 76.46%, iteration 40 is 78.84%, iteration 60 is 81.40%, iteration 80 is 84.70%, and iteration 100 is 87.99%.

4.4.4 Performance assessment with training set in the classification of second-level

Figure 8 indicates the performance examination of the proposed GJSO-DQN based on various performance measures in terms of the training set with various iterations. Figure 8a depicts the accuracy value of the proposed GJSO-DQN with several iterations. When taking the 90% training set, the accuracy of the proposed GJSO-DQN is 78.72% with 20 iterations, 82.57% with 40 iterations, 85.45% with 60 iterations, 87.72% with 80 iterations, and 90.26% with 100 iterations. Figure 8b signifies the FPR value of the proposed GJSO-DQN with multiple iterations. When considering the training set is 80%, the FPR of the proposed GJSO-DQN is 24.11% for iteration 20, 19.55% for 40 iterations, 18.11% for 60 iterations, 14.40% for 80 iterations, and 12.34% for the 100 iterations. Figure 8c depicts the TPR value based on the proposed GJSO-DQN with multiple iterations. By analysing training data is 90%, the TPR of proposed GJSO-DQN with iterations 20, 40,60,80, and 100 is 80.74%, 82.55%, 85.77%, 86.77%, and 89.24% respectively.

4.5 Comparative discussion

The comparative discussion of the proposed GJSO-DQN for the classification of plant leaves is explained in Table 2. By considering the 90% training set, plant leaf classification is identified based on the evaluation metrics. Here, the first and second levels of classification are performed. The accuracy of proposed GJSO-DQN is 90.61% and other techniques such as, EfficientNet DL is 80.91%, CNN is 83.13%, DenseNet-77 is 85.62%, and DENN is 86.55% during first-level classification. Second level classification accuracy of proposed GJSO-DQN is 90.18%, EfficientNet DL is 79.87%, CNN is 83.13%, DenseNet-77 is 84.79%, DENN is 87.58%. Here, the GJSO-DQN attained a high accuracy because of the execution of segmentation by Link-Net. The first level classification’s TPR of EfficientNet DL is 79.18%, CNN is 82.50%, DenseNet-77 is 84.84%, DENN is 87.28%, and proposed GJSO-DQN is 89.48%, the second level classification’s TPR of EfficientNet DL is 78.85%, CNN is 83.04%, DenseNet-77 is 84.54%, DENN is 86.95%, and proposed GJSO-DQN is 89.44% and proposed technique achieved superior TPR due to usage of DQN. The first level classification’s FPR value of proposed GJSO-DQN is 10.52%, and another technique has EfficientNet DL has 20.82%, CNN has 17.50%, DenseNet-77 has 15.16%, DENN has 12.72%. The second level classification’s FPR value of proposed GJSO-DQN is 10.56%, and another technique has EfficientNet DL has 21.15%, CNN has 16.96%, DenseNet-77 has 15.46%, DENN has 13.05%. In the proposed GJSO-DQN method, various augmentation techniques are performed, which results in a minimum FPR rate.

Table 2
Comparative discussion

Classification types	Evaluation parameters	EfficientNet DL	CNN	DenseNet-77	DENN	Proposed GJSO-DQN
First level classification	Accuracy (%)	80.91	83.13	85.62	86.55	90.61
	TPR (%)	79.18	82.50	84.84	87.28	89.48
	FPR (%)	20.82	17.50	15.16	12.72	10.52
Second level classification	Accuracy (%)	79.87	83.13	84.79	87.58	90.18
	TPR (%)	78.85	83.04	84.54	86.95	89.44
	FPR (%)	21.15	16.96	15.46	13.05	10.56

4.6 Convergence graph

The convergence graph is used to represents how the variations of the results take place in each iteration. The difference between the last two iterations and the target convergence are provided in the graph. The methods used for analysis are Tunicate Swarm Algorithm (TSA) [29], Bald Eagle Search (BES) [30], Harris Hawks Optimization (HHO) [31] and the proposed Geese Jellyfish Search Optimization (GJSO) for analysing the convergence behaviour. Figure 9 illustrate the fitness of the proposed model with the other models. The TSA, BES, HHO, and the proposed GJSO attain the fitness value of 10.608, 5.720, 3.101, and 0.012 for the iteration 100. Thus, the above analysis shows that the proposed method converges easily compared to other methods.

Figure 9.

Convergence plot. BES – Bald Eagle Search, GJSO – Geese Jellyfish Search Optimization, HHO – Harris Hawks Optimization, TSA – Tunicate Swarm Algorithm.

5. Conclusion

Effective plant disease detection is proposed using the geese jellyfish search optimization based DQN in this work using the leaf image acquired from the Plant Village dataset. The input leaf image is pre-processed by the Kalman filter and it is fed to the segmentation, which is done by Link-Net. The Link-Net is trained by the proposed geese jellyfish search optimization and it is of geese migration optimization and JS optimizer. Further, segmented image is allowed to the image augmentation function and following it and feature extraction is performed. Consequently, the extracted image is allowed for the classification of the type of plant leaf, which is done by DQN trained by the proposed geese jellyfish search optimization. geese jellyfish search optimization is a competitive algorithm and can handle complex practical problems. Also, it has an efficient performance in solving a range of large-scale optimization problems. It has a faster convergence rate with less time. In large-scale functions, it can find the optimal values. It achieves better results and is also used for solving various engineering optimization problems. The proposed geese jellyfish search optimization based DQN in first level classification achieved an accuracy of 90.61%, True Positive Rate (TPR) of 89.48%, and False Positive Rate (FPR) of 10.52% and in second-level classification achieved an accuracy of 90.18%, True Positive Rate (TPR) of 89.44%, and False Positive Rate (FPR) of 10.56%. The further dimension of this study will be to use a large-scale dataset for the improvement of performance parameters.

Footnotes

Appendix

Algorithm 1: Pseudo-code of the proposed geese jellyfish search optimization based Deep Q-Network
1. Initialize the population (weights of DQN)
2. Start
3. For $c=1:e$ Pop do
4. If $n(g)\geqslant 0.5$ :
5. Jellyfish go around the ocean’s current
6. Else: Jellyfish shift within the swarm
7. If $\textit{rand}^{\prime\prime\prime}(0,1)>(1-n(g))$
8. Jellyfish displays type A motion
9. Else:
10. Jellyfish displays type B motion
11. End if
12. End if
13. End for
14. End

Author’s Bios

Bandi Ranjitha is pursuing PhD in Presidency University in the Area of Machine Learning. She has published papers in Journals and International conferences. She worked as an Assistant Professor in different colleges, Hyderabad. She has more than 10 years of teaching experience. Her research interest includes machine Learning, Artificial intelligence and deep learning. ORCID: 0009-0002-5659-1375.

Sampath A K has completed his research in the Area of Machine Learning under Veltech University, Chennai in the Year of 2019. He has published more than 8 papers in SCI and Scopus index journals. He has published 3 patents including one Australian Patent. Currently he is working as an Associate Professor in Presidency University, Bangalore. He has more than 17 years of teaching. His research interest includes machine Learning, Artificial intelligence and System security.

References

Rajiv

Rajasekhar

Prasanna Lakshmi

Srinivasa Rao

and Sabitha Reddy

, Accuracy evaluation of plant leaf disease detection and classification using GLCM and multiclass SVM classifier, in: Proceedings of Intelligent Learning for Computer Vision: Proceedings of Congress on Intelligent Systems 2020, Singapore: Springer Singapore, Vol. 61, May 2021, pp. 41–54.

Jiang

Chen

Liu

and Liang

, Real-time detection of apple leaf diseases using deep learning approach based on improved convolutional neural networks, IEEE Access 7 (May 2019), 59069–59080.

Ahmed

A.A.

and Reddy

G.H.

, A mobile-based system for detecting plant leaf diseases using deep learning, Agri Engineering 3(3) (July 2021), 478–493.

Geetharamani

and Pandian

, Identification of plant leaf diseases using a nine-layer deep convolutional neural network, Computers & Electrical Engineering 76 (June 2019), 323–338.

Rumpf

Mahlein

A.K.

Steiner

Oerke

E.C.

Dehne

H.W.

and Plümer

, Early detection and classification of plant diseases with support vector machines based on hyperspectral reflectance, Computers and Electronics in Agriculture 74(1) (October 2010), 91–99.

Pandian

J.A.

Kumar

V.D.

Geman

Hnatiuc

Arif

and Kanchanadevi

, Plant disease detection using deep convolutional neural network, Applied Sciences 12(14) (July 2022).

Panchal

A.V.

Patel

S.C.

Bagyalakshmi

Kumar

Khan

I.R.

and Soni

, Image-based plant diseases detection using deep learning, Materials Today: Proceedings 80(3) (April 2023), 3500–3506.

Popovic

Latinovic

Pesic

Zecevic

Krstajic

and Djukanovic

, Architecting an IoT-enabled platform for precision agriculture and ecological monitoring: A case study, Computers and Electronics in Agriculture 140 (August 2017), 255–265.

Talavera

J.M.

Tobon

L.E.

Gomez

J.A.

Culman

M.A.

Aranda

J.M.

Parra

D.T.

Quiroz

L.A.

Hoyos

and Garreta

L.E.

, Review of IoT applications in agro-industrial and environmental fields, Computers and Electronics in Agriculture 142(a) (November 2017), 283–297.

10.

Chouhan

S.S.

Singh

U.P.

and Jain

, Automated plant leaf disease detection and classification using fuzzy based function network, Wireless Personal Communications 121 (July 2021), 1757–1779.

11.

Strange

R.N.

and Scott

P.R.

, Plant disease: a threat to global food security, Annu. Rev. Phytopathol 43(1) (February 2005), 83–116.

12.

Ashwinkumar

Rajagopal

Manimaran

and Jegajothi

, Automated plant leaf disease detection and classification using optimal MobileNet based convolutional neural networks, Materials Today: Proceedings 51(1) (2022), 480–487.

13.

Atila

Ü.

Uçar

Akyol

and Uçar

, Plant leaf disease classification using EfficientNet deep learning model, Ecological Informatics 61 (March 2021).

14.

Albattah

Nawaz

Javed

Masood

and Albahli

, A novel deep learning method for detection and classification of plant diseases, Complex& Intelligent Systems 8 (September 2021), 507–524.

15.

Agarwal

Gupta

S.K.

and Biswas

K.K.

, Development of Efficient CNN model for Tomato crop disease identification, Sustainable Computing: Informatics and Systems 28 (December 2020).

16.

Vallabhajosyula

Sistla

and Kolli

V.K.K.

, Transfer learning-based deep ensemble neural network for plant leaf disease detection, Journal of Plant Diseases and Protection 129(3) (May 2021), 545–558.

17.

Mahum

Munir

Mughal

Z.U.N.

Awais

Sher Khan

Saqlain

Mahamad

and Tlili

, A novel framework for potato leaf disease detection using an efficient deep learning model, Human and Ecological Risk Assessment: An International Journal 29(2) (April 2022), 303–326.

18.

Pandian

J.A.

Kanchanadevi

Kumar

V.D.

Jasińska

Gono

Leonowicz

and Jasiński

, A five convolutional layer deep convolutional neural network for plant leaf disease detection, Electronics 11(8) (April 2022).

19.

Chaurasia

and Culurciello

, Linknet: Exploiting encoder representations for efficient semantic segmentation, in: Proceedings of 2017 IEEE Visual Communications and Image Processing (VCIP), IEEE, December 2017, pp. 1–4.

20.

Zhang

Song

Zhang

and Zhao

, Wild Geese Migration Optimization Algorithm: A New Meta-Heuristic Algorithm for Solving Inverse Kinematics of Robot, Computational Intelligence and Neuroscience 2022, September 2022.

21.

Chou

J.S.

and Truong

D.N.

, A novel metaheuristic optimizer inspired by behavior of jellyfish in ocean, Applied Mathematics and Computation 389 (January 2021).

22.

Guermoui

and Mekhalfi

M.L.

, A Sparse Representation of Complete Local Binary Pattern Histogram for Human Face Recognition, arXiv preprint arXiv: 1605.09584, May 2016.

23.

Lessa

and Marengoni

, Applying artificial neural network for the classification of breast cancer using infrared thermographic images, in: Proceedings of Computer Vision and Graphics: International Conference, ICCVG, Vol. 8, September 2016, pp. 429–438. Springer International Publishing.

24.

Sasaki

Horiuchi

and Kato

, A study on vision-based mobile robot learning by deep Q-network, in: Proceedings of 2017 56th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE), IEEE, September 2017, pp. 799–804.

25.

Pakrashi

and Chaudhuri

B.B.

, A Kalman filtering induced heuristic optimization based partitional data clustering, Information Sciences 369 (July 2016), 704–717.

26.

Data documentation is taken from, https://mxnet.apache.org/versions/1.3.1/tutorials/python/types_of_data_augmentation.html, accessed on March 2023.

27.

Khalifa

N.E.

Loey

and Mirjalili

, A comprehensive survey of recent trends in deep learning for digital image augmentation, Artificial Intelligence Review 55 (September 2021), 2351–2377.

28.

PlantVillage – Dataset is taken from, https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw/color, accessed on March 2023.

29.

Kaur

Awasthi

L.K.

Sangal

A.L.

and Dhiman

, Tunicate Swarm Algorithm: A new bio-inspired based metaheuristic paradigm for global optimization, Engineering Applications of Artificial Intelligence 90 (April 2020).

30.

Alsattar

H.A.

Zaidan

A.A.

and Zaidan

B.B.

, Novel meta-heuristic bald eagle search optimisation algorithm, Artificial Intelligence Review 53 (July 2019), 2237–2264.

31.

Heidari

A.A.

Mirjalili

Faris

Aljarah

Mafarja

and Chen

, Harris hawks optimization: Algorithm and applications, Future Generation Computer Systems 97 (August 2019), 849–872.

Geese jellyfish search optimization trained deep learning for multiclass plant disease detection using leaf images

Abstract

Keywords

1. Introduction

2. Literature review

3. Proposed geese jellyfish search optimization based Deep Q-Network for plant disease detection

3.2.1 Kalman filter

3.3.1 Structure of Link-Net

1) Population initialization

2) Evaluation of fitness solution

3) Ocean current

4) Swarm of jellyfish

5) Mechanism of time control

6) Boundary condition

7) Feasibility estimation

8) Termination

1) Resizing

2) Shearing

3) Translation

4) Contrast

5) Saturation

6) Hue

Complete Local Binary Pattern (CLBP)

Statistical features

1) Mean

2) Variance

3) Standard deviation

4) Entropy

5) Skewness

6) Energy

7) Contrast

8) Correlation

3.6.1 Architecture of Deep Q-Network (DQN)

3.7.1 Architecture of Deep Q-Network (DQN)

3.7.2 Training function of Deep Q-Network (DQN) by geese jellyfish search

Table 1 Nomenclature

4.2 Evaluation measures

1) Accuracy

2) True Positive Rate (TPR)

3) False Positive Rate (FPR)

4.4 Comparative analysis

4.4.1 Comparative assessment with training set in the first level of classification

4.4.2 Comparative assessment with training set in the classification of the second level

4.4.3 Performance assessment with training set in classification of first-level

4.4.4 Performance assessment with training set in the classification of second-level

4.5 Comparative discussion

Table 2 Comparative discussion

Footnotes

Appendix

Author’s Bios

References

Table 1
Nomenclature

Table 2
Comparative discussion