Multi-classification of disease induced in plant leaf using chronological Flamingo search optimization with transfer learning

Literature survey

Hussain et al. (2022) designed a model for the detection of diseases from cucumber leaves. The model was developed based on deep learning (DL) and includes fusion and collection of optimum features. Here, a visual geometry group (VGG) was used. The feature obtained was fused with a maximum fusion scheme and optimum features were selected with the Whale Optimization algorithm and classified with supervised learning. However, this technique was imperfect in handling other databases. Jadhav, Udup & Patil (2019) developed a model using k-nearest neighbors (KNN) classifiers, and a multiclass support vector machine for the detection and categorization of soybean diseases with color images. The thresholding was applied for extracting interesting regions. The Incremental K-means clustering was adapted for segmentation and finally, support vector machine (SVM) and KNN were utilized to classify disease. The method suffered from overfitting problems. Singh & Kaur (2021) devised a technique for detecting and classifying the disease that occurred in potato plants. Here, the consistent data set was adapted which was known as the Plant Village database. Here, the K-means scheme was considered for segmenting images, and the gray-level co-occurrence matrix concept with multi-class SVM was applied for classification. This method provided poor accuracy while dealing with huge data. Islam et al. (2017) devised a technique that unified machine learning and image processing for permitting disease diagnosis through leaf images. The automatic technique classified the disease of the potato plant. The SVM was utilized for classifying disease and provided better accuracy. It was not appropriate for handling other datasets. Tiwari, Joshi & Dutta (2021), designed a deep learning (DL)-based method for identifying and classifying the disease in plants with leaf images acquired through different resolutions. Here, the dense convolutional neural network (dense CNN) was trained using huge plant leaf images considering data from different countries. The method was not able to expand the plant leaf database for handling complex platforms. Roy & Bhaduri (2021) devised a deep learning-enabled object detection technique for classifying plant disease. The method helps to provide accurate discovery and fine-grained detection of disease. Moreover, the model was enhanced to optimize both speed and accuracy. However, this method was not suitable for handling real platforms. Atila et al. (2021), devised EfficientNet for classifying disease from plant leaves. Here, the PlantVillage dataset was utilized for training the models. Each model was trained with a different set of images. Here, the concept of transfer learning was used where all layers were trained for performing classification. However, this method failed to extend the disease database by elevating the diversity of plants considering different sets of classes. Lakshmi & Savarimuthu (2021) devised a deep learning framework for automatic plant disease detection and segmentation (DPD-DS) considering an enhanced pixel-wise mask-region-based convolution neural network (CNN). It utilized a region convolution neural network (R-CNN) to save memory and cost. It helps to elevate detection accuracy. However, it did not adopt an ensemble network for detecting disease in plant leaves. Bayram, Bingol & Alatas (2022) established an artificial intelligence technique for automatically detecting tomato leaf disease. Here, for the classification Inceptionv3, Resnet50, Efficientb0, Shufflenet, Googlenet, and Alexnet models were used. Then, the feature maps were gathered from the tomato images. The neighborhood component analysis (NCA) was used in the feature extraction. However, a single disease dataset was used in this research.

Image acquisition

Research for the automatic discovery of diseases in plants has been of great interest among researchers for several years. A model is developed for detecting several diseases in plants considering plant leaf images. Due to increasing plant images, less expertise and knowledge cannot fulfill the requirements of huge-scale image processing. Thus, an automated detection of plant leaf has gained more focus. Due to the design of imaging techniques, people can simply attain clear images of plants and the computer-assisted detection of plant images is a major hotspot. Imagine a database $F$ having a set of plant leaf images $e$ and is represented as

(1) $$F=\left\{{\chi }_1,{\chi }_2,\cdots ,{\chi }_v,\cdots ,{\chi }_e\right\}$$where, $e$ depicts total images and ${\chi }_v$ symbolizes $v^{th}$ image.

Pre-processing with adaptive anisotropic diffusion

In the pre-processing phase, the inputted plant leaf image ${\chi }_v$ is considered. The pre-processing is adapted on the provided image in a sequence which makes it appropriate for better processing. The fundamental pre-processing phase is to resize the provided image. The initial image size is huge which makes it complex and takes more processing time. Hence, the application of pre-processing is done to eliminate noise and make it apt for improved processing. Thus, the adaptive anisotropic diffusion (Tang et al., 2007) is adapted for pre-processing. It is an effective smoothing procedure. Also, it is mainly used for noise removal and edge-preserving. The data loss and image blur problems are avoided in this model. Here, the evaluation of the edge map and the selection of k in diffusion coefficients are more imperative. The imperative selection of these two things will generate an improved anisotropic diffusion. If k is high, then the edge preservation will be best, but noise will not be discarded. If k is small, then noise will be discarded, but edges will be blurred. Hence, how to devise k is an imperative parameter in the diffusion technique. A suitable k is to divide the noise through the edges in an effective manner. The operator of gradient magnitude is sensitive to noise particularly whenever the edge strength is weak. Thus, the gradient magnitude is utilized as an edge map and is stated as

(2) $$k\left(q,r\right)=\sqrt{{\displaystyle \frac{{\left|\mathrm{\nabla }{B}_f\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_D\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_K\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_M\left(q,r\right)\right|}^2}{4}}}$$

(3) $$\mathrm{\nabla }{B}_D(q,r)=B\left(q,r+1\right)-B(q,r)$$

(4) $$\mathrm{\nabla }{B}_f(q,r)=B\left(q,r-1\right)-B(q,r)$$

(5) $$\mathrm{\nabla }{B}_K(q,r)=B\left(q+1,r\right)-B(q,r)$$

(6) $$\mathrm{\nabla }{B}_M(q,r)=B\left(q-1,r\right)-B(q,r).$$

The evaluation of diffusion coefficients is explored as

(7) $$d(k)=\exp \left(-{\left[{\displaystyle \frac{k(q,r)}{m}}\right]}^2\right)$$

(8) $$d(k)={\displaystyle \frac{1}{1+\left[k{(q,r)}^2-m^2\right]/\left[m^2\left(1+m^2\right)\right]}}.$$

Assume the distribution of noise in an image and consider a homogenous region with noise and can consider that $\mathrm{\nabla }{B}_D(q,r)$, $\mathrm{\nabla }{B}_f(q,r)$, $\mathrm{\nabla }{B}_K(q,r)$, and $\mathrm{\nabla }{B}_M(q,r)$ poses the same zero mean normal distribution like $\mathrm{\aleph }\left(0,{\sigma }^2\right)$ and hence equation becomes,

(9) $${\displaystyle \frac{4}{{\sigma }^2}\ast k(q,r{)}^2={\displaystyle \frac{1}{{\sigma }^2}\left({\left|\mathrm{\nabla }{B}_f\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_D\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_K\left(q,r\right)\right|}^2+{\left|\mathrm{\nabla }{B}_M\left(q,r\right)\right|}^2\right)}}$$

It poses a chi-squared distribution having four freedom degrees.

Once the noise gradient distribution is obtained, compute the choice of $m$. Consider that $q,r$ indicates the pixel coordinates of the image and one can utilize Eq. (9) to generate the gradient map. If the point belongs to the edge, then the gradient map becomes an outlier in four freedom degrees. Hence the threshold $m$ is chosen for rejecting those outliers and $m$ is generated as

(10) $$\rho \left(k{\left(q,r\right)}^2>m^2\right)\le \eta .$$

The aforesaid expression is rewritten as

(11) $$\rho \left({\displaystyle \frac{4}{{\sigma }^2}\ast k{(q,r)}^2>m^2\ast {\displaystyle \frac{4}{{\sigma }^2}}}\right)\le \eta$$where, $\eta$ stands for significance level. Hence, the $m$ is set as

(13) $$m=d\sigma$$where, ${\sigma }^2$ represents the variance of noise gradient and $d$ denotes constant. Hence, the pre-processed outcome generated through adaptive anisotropic diffusion is notified as $\mathrm{T}$.

Segmentation of plant leaf with Moving Gorilla Remora algorithm (MGRA)-based U-Net++

Thus, the pre-processed outcome $\mathrm{T}$ is provided as segmentation input. The segmentation function aimed to mine the complete leaf region over the background. For improving the reliability and accuracy of mining, it is essential for the model to be in capacity for depicting the features contained in the image that are fine-grained and alteration in size and shape. The U-Net++ (Fan et al., 2022) is adapted for performing the segmentation of plant leaves with MGRA-based U-Net++. The structure of U-Net++ and its training module is stated in subsections.

Augmentation of image

The segmented outcome $E$ is provided to this phase. It is useful when one is provided a database with very less instances. Moreover, this process aids in fighting overfitting and enhances the efficiency of deep networks for dealing with several tasks. It aids in making data rich and adequate hence making the model perform better and precisely. It helps to minimize operational costs considering various kinds of transformations. Here, position and color augmentation methods are applied and are examined below:

Obtain crucial feature

Here, the augmented image vector $Z$ is given as input. Feature extraction indicates an imperative step in constructing the pattern categorization and its goal is to extract the pertinent data which helps to characterize each class. Here, the pertinent features are obtained through the images to form a feature vector. These feature vectors are further used by the classifier for recognizing the target. It is easy for the classifier to classify different classes by adapting the features as it permits easy to distinguish. Here, features like OCLBP-based DWT, SIFT, and LTP features are mined:

a) OCLBP-based DWT

The augmented image is fed to DWT and it is divided into LL, LH, HL, and HH bands wherein the HH band contains noise and thus it is prevented. Hence LL, LH, and HL bands are adapted with OCLBP (Vishnoi, Kumar & Kumar, 2022) and finally concatenated and applied with HoG to establish a feature vector.

OCLBP (Vishnoi, Kumar & Kumar, 2022) indicates a joint color-texture feature that helps to compare gray scale and color textures. Here, pairs of colors like red-green and yellow-blue are acquired by humans and called opponent colors. These opponent colors are determined by adapting LBP on center and adjacent pixels with opposite color channels. The texture $\tau$ is adapted as a distribution $\tau \approx \gamma \left(\alpha \left({\upsilon }_{\kappa }-{\upsilon }_{\delta }\right),\cdots ,\alpha \left({\upsilon }_{\kappa -1}-{\upsilon }_{\delta -1}\right)\right)$. The local texture considering the image around $\mathrm{\Re }\left({\vartheta }_{\delta },{\hslash }_{\delta }\right)$is described as

(17) $$LB{P}_{\zeta ,\mathrm{\Re }\left({\vartheta }_{\delta },{\hslash }_{\delta }\right)}=\sum _{\kappa =0}^{\kappa -1}\alpha \left({\upsilon }_{\kappa }-{\upsilon }_{\delta }\right)2^{\kappa }$$

(18) $$\alpha (j)=\{\begin{array}{c}1,\vartheta \ge 0\hfill \\ 0,\vartheta \le 0.\hfill \end{array}$$

Hence, the texture of the image is described approximately as

(19) $$\tau =\gamma \left(LB{P}_{\zeta ,\mathrm{\Re }\left({\vartheta }_{\delta },{\hslash }_{\delta }\right)}\right).$$

HoG (Vishnoi, Kumar & Kumar, 2022) offers information regarding the occurrence of orientation-related gradients in RoI or local regions. The evaluation of gradient $R$ and direction $\phi$ is modeled in a generalized way as

(20) $$\left|R\right|={\left(R_{\vartheta }^2+R_{\hslash }^2\right)}^{{\displaystyle \frac{1}{2}}}$$

(21) $$\phi ={\tan }^{-1}{\displaystyle \frac{R_{\vartheta }}{R_{\hslash }}}$$where $R_{\vartheta }$ and $R_{\hslash }$ represent gradient along $\vartheta$ and $\hslash$ directions. The image is split into various square cells or areas with specific sizes. This feature is explained as $V_1$.

b) SIFT

SIFT (Vishnoi, Kumar & Kumar, 2022) offers local key features considering the objects that are unchanged against the transformations of scale. It involves three steps namely determining the scale-space maxima and keypoints, orientation assignment, and key point descriptor.

Determination of scale-space maxima and key points: The operation utilized to find key points is termed scale-space and is expressed as $A\left(a,c,\sigma \right)$. It is termed as a convolution of Gaussian kernel $C\left(a,c,\sigma \right)$ and image $H(a,c)$ such that

(22) $$A(a,c,\sigma )=C\left(a,c,\sigma \right)\ast H(a,c)$$where, $\ast$ denotes convolution amid $a$ and $c$, and $\sigma$ stands for the standard deviation of various scales such that

(23) $$C(a,c,\sigma )={\displaystyle \frac{1}{2\pi {\sigma }^2}{\exp }^{-{\displaystyle \frac{a^2+c^2}{2{\sigma }^2}}}}.$$

The difference of Gaussians (DOG) $S\left(a,c,\sigma \right)$ is evaluated as

(24) $$S(a,c,\sigma )=A\left(a,c,T\sigma \right)-A\left(a,c,\sigma \right)$$where $T$ is constant which divides two successive smooth images.

Orientation assignment: Hence, the gradient magnitude $V(a,c)$ and orientation $W(a,c)$considering smoothed image at scale $\sigma$ is given by

(25) $$V(a,c)={\left\{{\left(A(a+1,c)-A(a-1,c)\right)}^2+{\left(A(a,c+1)-A(a,c-1)\right)}^2\right\}}^{{\displaystyle \frac{1}{2}}}$$

(26) $$W(a,c)={\tan }^{-1}{\displaystyle \frac{A(a,c+1)-A(a,c-1)}{A(a+1,c)-A(a-1,c)}}.$$

Key point descriptor: It is established for every key that poses a stable orientation, scale, and position. This SIFT feature is signified by $V_2$.

c) LTP

LTP (Vishnoi, Kumar & Kumar, 2022) is an expansion of LBP in which the center pixel and its adjacent pixels are done in three unique zones. Considering three zones, the LTP histogram is generated. The LTP operator is stated as

(27) $$LT{P}_{\zeta ,\mathrm{\Re }}\left(s,t\right)=\sum _{\kappa =0}^{\kappa -1}\alpha \left({\upsilon }_{\kappa }-{\upsilon }_{\delta }\right)3^{\kappa }$$where

(28) $$\alpha \left(j\right)=\{\begin{array}{c}-1,j\le s_{\delta }-\gamma \hfill \\ 0-\gamma <j<s_{\delta }+\gamma \hfill \\ 1j\ge s_{\delta }+\gamma \hfill \end{array}$$where $\gamma$ denotes threshold. The LTP feature is notified by $V_3$. Hence, the feature vector formed is stated by,

(29) $$V=\{V_1,V_2,V_3\}.$$

First-level classification to identify plant leaf type using TL with LeNet

After completing the feature extraction, the obtained feature vectors are applied for further inspection before being grouped into specific classes. The plant leaf type classification is an essential step in discovering the type of plant. In previous works, the researchers struggle and spend a huge time establishing the dataset by accumulating several leaf samples as raw databases. Here, the first level classification is considered which helps to recognize the type of plant leaf using TL with LeNet. The outlook of TL with LeNet is described along with CFSA steps.

Outlook of TL with LeNet

The LeNet is simple and a minimum number of layers only required. Also, it needs less training time. TL is merged into a LeNet model to enhance the accuracy and model reliability when trained using a small quantity of data. Here, the Outlook of TL with LeNet is described herewith.

CNN gives enhanced efficacy with large datasets in contrast to smaller ones. It is useful in the applications of CNN wherein data tends to be small. Hence, the idea behind the TL is that it comprises a trained model with huge databases that are utilized for applications that contain small databases. Figure 2 provides LeNet with TL.

Training of LeNet

LeNet (Bouti et al., 2020; Wahlang et al., 2020) indicates a current convolutional network that is devised for detecting plant leaf disease. It expresses a CNN having adequate input to generate several objects and various outputs. It can determine strings that lack prior segmentation. Hence, max pooling and layers of sparse convolution indicate the LeNet center. Thus, the lower layer contains the max pool and convolution layers. The output generated by LeNet is signified by ${\ell }_{out}$. Figure 3 illustrates a preview of the LeNet structure.

CNN model

CNN (Aslam & Cui, 2020) is well-known because of its enhanced efficiency in classifying the data. Here, the set of convolution layers and filters helps in extracting spatial and temporal features through data. The layer comprises a weight-sharing method which helps in reducing evaluations. The CNN expresses a feed forward ANN which comprises two problems.

Second-level classification to classify multi-class plant disease

Earlier detection and aversion of disease in plants are imperative factors in harvesting crops as they can effectually minimize any disorders of growth and hence reduce the application of pesticides for attaining better crop production. Hence, automatic classification of plant disease is an effective technique for attaining precision agriculture. Here, the second level classification involves the classification of multi-class plant diseases using TL with LeNet. Here, the tuning of TL with LeNet is done by CSFA.

Set-up of experiment

CFSA+TL-based CNN+LeNet is programmed in Python. The CFSA+TL-based CNN+LeNet parameters are presented in Table 1.

Dataset description

The technique evaluation is performed with the Plant Village Dataset (Mohanty, 2022; https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw/color). It comprises 54,303 healthy and unhealthy images of the leaf which is split into 38 classes by species as well as disease. It is an open-access image repository that evaluates plant health to enable the design of mobile disease diagnosis. It is a dataset containing images of diseased plant leaf and their labels. There are 14 crop species available in the dataset, such as tomato, strawberry, squash, soy, raspberry, potato, pepper, peach, orange, grape, cherry, blueberry, and apple. There are 17 fundamental diseases, such as mold disease 2, mite disease 1, viral disease 2, and bacterial disease 4 in this dataset.

Experimental upshots

Figure 4 shows experimental outputs of CFSA+TL-based CNN+LeNet using different images. The input image is explicated in Fig. 4A. The pre-processed image produced by adaptive anisotropic filtering is exposed in Fig. 4B. The segmented image obtained by U-Net++ is endowed in Fig. 4C. The contrast-enhanced image is depicted in Fig. 4D. The hue image is exposed in Fig. 4E. The saturated image obtained is depicted in Fig. 4F. The Affine transformed image is displayed in Fig. 4G. The padded image is depicted in Fig. 4H. The Rotated image is endowed in Fig. 4I. The translated image produced is explicated in Fig. 4J.

Figure 5 depicts the outputs of CFSA+TL-based CNN+LeNet with different image augmentation techniques along with extracted features. The affine transformation of DWT-OCLBP, LTP, and SIFT is explicated in Fig. 5A. The contrast of DWT-OCLBP image, LTP, and SIFT is displayed in Fig. 5B. The hue of the DWT-OCLBP image, LTP, and SIFT is induced in Fig. 5C. The Padding of DWT-OCLBP image, LTP, and SIFT is notified in Fig. 5D. The Rotation of the DWT-OCLBP image, LTP, and SIFT is noted in Fig. 5E. The Saturation of the DWT-OCLBP image, LTP, and SIFT is noted in Fig. 5F. The Translation of DWT-OCLBP image, LTP, and SIFT is noted in Fig. 5G.

Metrics used

Proficiency of each scheme ability is observed by inspecting CFSA+TL-based CNN+LeNet with different performance parameters which is explained below.

(a) Accuracy

It is considered as a metric of closeness degree to its true value and can be expressed as

(44) $$\mathrm{M}={\displaystyle \frac{B+V}{B+V+F+D}}.$$

Hence, $V$states true positive, $B$ offers true negative, $D$ gives false positive, and displays false negative.

(b) Sensitivity

It depicts the proportion of true positives number to the total number of positives, and is presented by

(45) $$\mathrm{N}={\displaystyle \frac{V}{V+F}}.$$

(c) Specificity

It defines the ratio of negatives and is accurately identified and it is notated by

(46) $$\mathrm{K}={\displaystyle \frac{B}{B+D}}.$$

(d) F-measure

It is the compromise between precision and recall and it is noted

(47) $$FM={\displaystyle \frac{V}{V+{\displaystyle \frac{1}{2}\left(D+F\right)}}}.$$

Algorithm methods

The algorithm efficiency estimation considered for analysis is PSO+TL-based CNN+LeNet (Wang, Tan & Liu, 2018; Bouti et al., 2020), CSO+TL-based CNN+LeNet (Cheng & Jin, 2014; Bouti et al., 2020), ROA+TL-based CNN+LeNet (Jia, Peng & Lang, 2021; Bouti et al., 2020), GTO+TL-based CNN+LeNet (Xiao et al., 2022; Bouti et al., 2020), FSA+TL-based CNN+LeNet (Zhiheng & Jianhua, 2021; Bouti et al., 2020), and proposed CFSA+TL-based CNN+LeNet.

Algorithmic analysis

The evaluation of scheme efficacy with first-level and second-level classifications is described with different metrics by altering swarm size along the x-axis.

a) Graphical estimation of algorithm efficacy with first-level classification

Figure 6 gives an evaluation of algorithm efficacy with first-level classification considering different metrics. The graph depicting accuracy is explicated in Fig. 6A. When the swarm’s size is 20, the accuracy produced by PSO+TL-based CNN+LeNet is 0.837, CSO+TL-based CNN+LeNet is 0.865, ROA+TL-based CNN+LeNet is 0.887, GTO+ TL-based CNN+LeNet is 0.898, FSA+TL-based CNN+LeNet is 0.918, and CFSA+TL-based CNN+LeNet is 0.946. The graph denoting analysis considering sensitivity is explicated in Fig. 6B. For swarms size is 20, the elevated sensitivity of 0.958 is produced by CFSA+TL-based CNN+LeNet whilst sensitivity of PSO+TL-based CNN+LeNet, CSO+TL-based CNN+LeNet, ROA+TL-based CNN+LeNet, GTO+TL-based CNN+LeNet, FSA+TL-based CNN+LeNet are 0.837, 0.858, 0.877, 0.887, and 0.908. The graph regarding specificity analysis is considered in Fig. 6C. Considering swarms size as 20, the specificity generated is 0.827 for PSO+TL-based CNN+LeNet, 0.848 for CSO+TL-based CNN+LeNet, 0.859 for ROA+TL-based CNN+LeNet, 0.865 for GTO+TL-based CNN+LeNet, 0.877 for FSA+TL-based CNN+LeNet, and 0.936 for CFSA+TL-based CNN+LeNet. The graph regarding F-Measure analysis is considered in Fig. 6D. Considering swarms size as 10, the F-measure generated is 0.808 for PSO+TL-based CNN+LeNet, 0.822 for CSO+TL-based CNN+LeNet, 0.842 for ROA+TL-based CNN+LeNet, 0.857 for GTO+TL-based CNN+LeNet, 0.872 for FSA+TL-based CNN+LeNet, and 0.907 for CFSA+TL-based CNN+LeNet.

b) Graphical evaluation of algorithm efficacy with second-level classification

Figure 7 gives the evaluation of algorithm efficacy with second-level classification considering different metrics. The graph depicting accuracy analysis is explicated in Fig. 7A. When swarms size is 20, the accuracy noted by PSO+TL-based CNN+LeNet is 0.877, CSO+ TL-based CNN+LeNet is 0.887, ROA+TL-based CNN+LeNet is 0.898, GTO+TL-based CNN+LeNet is 0.908, FSA+TL-based CNN+LeNet is 0.927, and CFSA+TL-based CNN+LeNet is 0.957. The graph denoting analysis considering sensitivity is explicated in Fig. 7B. When swarms size is 20, the elevated sensitivity of 0.965 is produced by CFSA+TL-based CNN+LeNet whilst sensitivity of PSO+TL-based CNN+LeNet, CSO+TL-based CNN+LeNet, ROA+TL-based CNN+LeNet, GTO+TL-based CNN+LeNet, FSA+ TL-based CNN+LeNet are 0.848, 0.858, 0.865, 0.898, 0.908. The graph regarding specificity analysis is considered in Fig. 7C. For swarms size is 20, the specificity produced is 0.858 for PSO+ TL-based CNN+LeNet, 0.877 for CSO+TL-based CNN+LeNet, 0.887 for ROA+TL-based CNN+LeNet, 0.898 for GTO+TL-based CNN+LeNet, 0.928 for FSA+TL-based CNN+LeNet, and 0.947 for CFSA+TL-based CNN+LeNet. The graph regarding F-Measure analysis is considered in Fig. 7D. For swarms size is 10, the F-measure produced is 0.826 for PSO+ TL-based CNN+LeNet, 0.842 for CSO+TL-based CNN+LeNet, 0.856 for ROA+TL-based CNN+LeNet, 0.876 for GTO+TL-based CNN+LeNet, 0.892 for FSA+TL-based CNN+LeNet, and 0.927 for CFSA+TL-based CNN+LeNet.

Comparative methods

The schemes efficiency estimation adapted for assessment are dense CNN (Tiwari, Joshi & Dutta, 2021), deep learning (Roy & Bhaduri, 2021), EfficientNet deep learning (Atila et al., 2021), DPD-DS (Lakshmi & Savarimuthu, 2021), DbneAlexnet-MGRA, and CFSA+TL-based CNN+LeNet.

Comparative analysis

The valuation of technique efficacy with first and second-level categorization is defined with various measures by altering the learning epoch along the x-axis.

a) Graphical estimation of scheme efficacy with first-level categorization

Figure 8 gives an evaluation of scheme efficacy with first-level classification considering different metrics. The graph depicting accuracy analysis is noted in Fig. 8A. When the learning epoch is 90%, the accuracy produced by dense CNN is 0.809, Deep learning is 0.827, EfficientNet deep learning is 0.848, DPD-DS is 0.877, DbneAlexnet-MGRA is 0.898, and CFSA+TL-based CNN+LeNet is 0.946. The graph denoting analysis considering sensitivity is noted in Fig. 8B. When the learning epoch is 90%, the sensitivity produced by dense CNN is 0.837, Deep learning is 0.859, EfficientNet deep learning is 0.877, DPD-DS is 0.898, DbneAlexnet-MGRA is 0.908, and CFSA+TL-based CNN+LeNet is 0.958. The graph regarding specificity analysis is considered in Fig. 8C. When the learning epoch is 90%, the specificity produced is 0.798 for dense CNN, 0.817 for deep learning, 0.848 for EfficientNet deep learning, 0.877 for DPD-DS, 0.909 for DbneAlexnet-MGRA, and 0.936 for CFSA+TL-based CNN+LeNet. The graph regarding F-measure analysis is considered in Fig. 8D. When the learning epoch is 90%, the specificity produced is 0.817 for dense CNN, 0.837 for deep learning, 0.862 for EfficientNet deep learning, 0.887 for DPD-DS, 0.908 for DbneAlexnet-MGRA, and 0.947 for CFSA+TL-based CNN+LeNet.

b) Graphical evaluation of scheme efficacy with second-level categorization

Figure 9 gives the evaluation of scheme efficacy using different metrics. The graph depicting accuracy is noted in Fig. 9A. When the learning epoch is 90%, the accuracy produced by dense CNN is 0.837, Deep learning is 0.858, EfficientNet deep learning is 0.877, DPD-DS is 0.898, DbneAlexnet-MGRA is 0.927, and CFSA+TL-based CNN+LeNet is 0.957. The graph denoting analysis considering sensitivity is explicated in Fig. 9B. When the learning epoch is 90%, the elevated sensitivity of 0.965 is produced by CFSA+TL-based CNN+LeNet whilst the sensitivity of the remaining schemes is 0.865, 0.887, 0.908, 0.916, and 0.936. The graph regarding specificity analysis is considered in Fig. 9C. When the learning epoch is 90%, the specificity produced is 0.859 for dense CNN, 0.865 for deep learning, 0.887 for EfficientNet deep learning, 0.908 for DPD-DS, 0.918 for DbneAlexnet-MGRA, and 0.947 for CFSA+TL-based CNN+LeNet. The graph regarding F-measure analysis is considered in Fig. 9D. When the learning epoch is 90%, the F-measure produced is 0.862 for dense CNN, 0.876 for deep learning, 0.897 for EfficientNet deep learning, 0.912 for DPD-DS, 0.927 for DbneAlexnet-MGRA, and 0.956 for CFSA+TL-based CNN+LeNet.

Analysis using segmentation accuracy

Figure 10 provides the estimation with the segmentation accuracy metric. For the 60% learning epoch, the segmentation accuracy noted by SegCNN is 0.808, U-Net is 0.827, Region-based segmentation is 0.837, K-mean algorithm is 0.858, and U-Net++ is 0.898. Considering the 90% learning epoch, the high segmentation accuracy of 0.949 is noted by U-Net++ whilst the segmentation accuracy of the remaining schemes is 0.838, 0.859, 0.865, 0.887.

Comparative estimate

The efficiency of schemes and algorithms considering different metrics are described below.

Algorithm estimate

Table 2 defines algorithm efficiency assessment considering diverse performance criteria. Considering first-level classification, the augmented accuracy of 94.6% is observed by CFSA+TL-based CNN+LeNet while the accuracy of enduring schemes is 83.7%, 86.5%, 88.7%, 89.8%, and 91.8%. The high sensitivity of 95.8% is produced by CFSA+TL-based CNN+LeNet while the sensitivity of enduring schemes is 83.7%, 85.8%, 87.7%, 88.7%, and 90.8%. The high specificity of 93.6% is produced by CFSA+TL-based CNN+LeNet while the specificity of enduring schemes is 82.7%, 84.8%, 85.9%, 86.5%, and 87.7%. The highest F-measure is 94.7%. Considering second-level classification, the augmented accuracy of 95.7%, sensitivity of 96.5%, specificity of 94.7%, and F-measure of 95.6% is noted by CFSA+TL-based CNN+LeNet. From the evaluation, it is observed that the CFSA+TL-based CNN+LeNet has improved its ability to provide a better classification of plant leaf disease.

Scheme evaluation

Table 3 describes the scheme efficiency computation with various evaluation criteria. Considering first-level classification, the augmented accuracy of 94.6% is observed by CFSA+TL-based CNN+LeNet while the accuracy of enduring schemes is 80.9%, 82.7%, 84.8%, 87.7%, and 89.8%. The high sensitivity of 95.8% is noted by CFSA+TL-based CNN+LeNet whilst the sensitivity of enduring schemes is 83.7%, 85.9%, 87.7%, 89.8%, and 90.8%. The high specificity of 93.6% is produced by CFSA+TL-based CNN+LeNet whilst the specificity of enduring schemes is 79.8%, 81.7%, 84.8%, 87.7%, and 90.9%. The high F-measure of 94.7% is produced by CFSA+TL-based CNN+LeNet whilst the F-measure of enduring schemes are 81.7%, 83.7%, 86.2%, 88.7%, and 90.8%. Considering second-level classification, the augmented accuracy of 95.7%, sensitivity of 96.5%, specificity of 94.7%, and F-measure of 95.6% is noted by CFSA+TL-based CNN+LeNet. From the analysis, it is revealed that the CFSA+TL-based CNN+LeNet is effective in performing the classification to find and classify diseases present in plant leaves with improved accuracy.

Statistical analysis

Table 4 describes the statistical analysis of the CFSA+TL-based CNN+LeNet. It is based on the best, mean, and variance of the model using various evaluation metrics.