Deep hybrid classification model for leaf disease classification of underground crops

Abstract

Underground crop leave disease classification is the most significant area in the agriculture sector as they are the significant source of carbohydrates for human food. However, a disease-ridden plant could threaten the availability of food for millions of people. Researchers tried to use computer vision (CV) to develop an image classification algorithm that might warn farmers by clicking the images of plant’s leaves to find if the crop is diseased or not. This work develops anew DHCLDC model for underground crop leave disease classification that considers the plants like cassava, potato and groundnut. Here, preprocessing is done by employing median filter, followed by segmentation using Improved U-net (U-Net with nested convolutional block). Further, the features extracted comprise of color features, shape features and improved multi text on (MT) features. Finally, Hybrid classifier (HC) model is developed for DHCLDC, which comprised CNN and LSTM models. The outputs from HC(CNN + LSTM) are then given for improved score level fusion (SLF) from which final detected e are attained. Finally, simulations are done with 3 datasets to show the betterment of HC (CNN + LSTM) based DHCLDC model. The specificity of HC (CNN + LSTM) is high, at 95.41, compared to DBN, NN, RF, KNN, CNN, LSTM, DCNN, and SVM.

Keywords

Leaf Disease improved multi texton modified U-Net hybrid classifier improved score level fusion

Table 1
Nomenclature

Abbreviation Description

AI Artificial Intelligence

CBB Cassava Bacterial Blight

CGM Cassava Green Mottle

CV Computer Vision

CMD Cassava Mosaic Disease

CBSD Cassava Brown Streak Disease

CNN Convolutional Neural Network

DCB Dense Convolution Blocks

DCNN Deep CNN

DL Deep Learning

DBN Deep Belief Network

ECNN Enhanced CNN

FAMP-Softmax Focal Angular Margin Penalty Softmax loss

FN False Negatives

FC Fully Connected

FP False Positives

HC Hybrid Classifier

H2K Harris, HOG, and KNN

HOG Histogram on Oriented Gradient

HSI Hyperspectral Imaging

KNN K-Nearest Neighbor

LSTM Long Short-Term Memory

ML Machine Learning

MF Median Filtering

MSI Multispectral Imaging

NN Neural Network

NDCB Nested Dense Convolution Blocks

LP Learning Percentage

SLF Score Level Fusion

SMOTE Synthetic Minority Over-sampling Technique

T-RNet Transformer Embedded ResNet

TTA Test-Time Augmentation

SVM Support Vector Machine

VGG Visual Geometry Group

RF Random Forest

Abbreviation	Description
AI	Artificial Intelligence
CBB	Cassava Bacterial Blight
CGM	Cassava Green Mottle
CV	Computer Vision
CMD	Cassava Mosaic Disease
CBSD	Cassava Brown Streak Disease
CNN	Convolutional Neural Network
DCB	Dense Convolution Blocks
DCNN	Deep CNN
DL	Deep Learning
DBN	Deep Belief Network
ECNN	Enhanced CNN
FAMP-Softmax	Focal Angular Margin Penalty Softmax loss
FN	False Negatives
FC	Fully Connected
FP	False Positives
HC	Hybrid Classifier
H2K	Harris, HOG, and KNN
HOG	Histogram on Oriented Gradient
HSI	Hyperspectral Imaging
KNN	K-Nearest Neighbor
LSTM	Long Short-Term Memory
ML	Machine Learning
MF	Median Filtering
MSI	Multispectral Imaging
NN	Neural Network
NDCB	Nested Dense Convolution Blocks
LP	Learning Percentage
SLF	Score Level Fusion
SMOTE	Synthetic Minority Over-sampling Technique
T-RNet	Transformer Embedded ResNet
TTA	Test-Time Augmentation
SVM	Support Vector Machine
VGG	Visual Geometry Group
RF	Random Forest

1. Introduction

There exist plants all around the planet. In India, 70% of the populace relies on the agricultural sector, which supports all of the country’s current civilizations. Agriculturalists can choose to produce an extensive variety of fruit and vegetable crops [5,12,19]. Agriculture has historically used techniques including crop rotation, herbicides, irrigation, and fertilizers. Farmers encounter a variety of issues while spotting and recognizing crop illnesses. The inclusion of soil, seed and chemical based methods to agriculture results in better farming system; nonetheless, the complex system of agriculture cannot survive without sensible and careful control of every input. Therefore, progress is needed to produce plants in the right manner [22,25,28]. The main problems that might influence the production of tomato, rice, pepper and potato plants include plant diseases and chemical fertilizers. This also burdens the development of underground crops. Therefore, a more accurate diagnosis and prompt correct crop care are required to avoid them from suffering severe losses as these crops are the major source of carbohydrates that is efficient for the human body. [6,17,18]. Early detection of plant diseases is essential because they can negatively affect animal and human health as well as change the quantity and quality of crops that are produced. Therefore, identifying and categorizing plant diseases is an essential effort [9,15,27]. Determining disease is thus a crucial stage in agriculture.

Advanced methods should be used to solve agricultural issues in order to meet the need for food. The agriculture sectors are concentrating on AI techniques in this area [26,30,31]. Conventional ML techniques have been used to a different of agricultural tasks. Additionally, deep learning (DL) led to important advancements in the study of agriculture. This is because the DL algorithms have the capacity to automatically extract feature [2,8,40]. Among other agricultural issues, correctly classifying plant diseases is essential to increasing the amount and quality of agricultural output while minimizing the need for chemical sprayers like fungicides and herbicides. Therefore, advancing automation in agriculture is a growing area of research [11,37,39]. Due to the similarity in the prevalence of plant diseases, this agricultural duty is challenging. Numerous researches have been done in this area to better categorize plant diseases.

For classifying plant diseases, many typical ML schemes were used. Similar to this, the detection of plant/leaf diseases has also been accomplished using advanced HSI and MSI techniques [3,35,38]. But as DL progressed, a number of cutting-edge architectures, including as VGG, AlexNet, DenseNet, ResNet and Inception-v4, showed promise in classifying plant diseases. Numerous researches [16] have demonstrated the significance of DL-based approaches as opposed to conventional ML techniques. Plant diseases account for about 25% of crop losses, which means that a substantial amount of food that might feed 600 million people is lost. Plant leaf diseases are one of these illnesses that significantly lowers agricultural productivity. The potato crop is of great significance on a national and international scale, as it is a staple crop for both export and local consumption. As a result, treating plant diseases becomes essential for the agriculture industry and economy of our nation. We can increase yields, improve crop quality, and decrease the occurrence of damaged crops by successfully addressing this challenge. By taking this strategy, agriculture advances and the economy is strengthened as well. To overcome all these issues the proposed model is used.

Proposes anew DHCLDC model for underground crop disease classification, where, Improved U-Net model (U-Net with nested convolutional block) is employed for segmentation process.

Extracts color features, shape features and Improved multi texton features from segmented image, which are the essential information to train the hybrid model.

Deploys HC (combining CNN and LSTM) for classification purpose that determines the improved score level fusion for final classification result.

Section 2 describes the reviews of existing DHCLDC approaches. Section 3 describes proposed DHCLDC model & Sect. 4 describes ncolour feature, shape feature and improved MT features. Section 5 explains the hybrid classifiers with improved score level fusion. The outcomes are in 6 and 7 sections.

2. Literature review

2.1. Related works

In 2020, G. Sambasivam and Geoffrey Duncan Opiyo [33] sought to identify cassava infections. Due to class imbalance of dataset, the training of CNNs to attain high precision was very complicated. Class imbalance was an issue that was seen in several fields of ML. However, only little attention was paid to class imbalances. In order to attain an accuracy score of over 93%, this article concentrated on methods using focal loss with DCNN, SMOTE, and class weight. In order to effectively anticipate under represented classes, it was necessary to address high-class imbalance.

In 2022, Yiwei Zhong et al. [41] suggested T-RNet system that, by illustrating global data and reducing the influence of backdrop noise, improved the concern on targeted area for the task of classifying cassava leaf diseases. Additionally, a novel loss operation known as FAMP-Softmax was suggested to help the method obtain drigorous classification limits while battling the dataset’s imbalanced nature related to cassava leaf disease.

Umesh Kumar Lilhore et al. [23] in 2022 presented a thorough learning technique for real time Cassava leaf disease classification using ECNN models. The current CNN model has a high computational cost because to its comprehensive data processing features. The suggested ECNN model used a deep separable convolution layer to fix CNN problems. The number of features and computational overhead were reduced by this feature. The unbalanced images were processed using the suggested ECNN model using a unique block processing feature. The suggested ECNN model made use of a Gamma correction function to address the color segregation problem. The suggested ECNN model made use of worldwide mean election poll with batch normalization to reduce the variable election procedure and to boost computing efficiency.

A very effective CNN design that was ideal for detecting potato illness was proposed by Trong et al. in 2021 [21]. Diseases of plants have a big impact on output, hence crop diseases need to be found and understood. The automated identification of diseased crops is made possible by smart farming employing ML. Disease management and control in real time improved output and lower agricultural losses. The training set’s database was built via image processing. Cross-entropy was employed for model analysis, while Adam was employed as the optimizer. The testing results show that the proposed model has an average reduction in parameter usage of 99.3% and a 99% accuracy rate for diagnosing plant sickness.

Alberta et al. [4] in 2023 presented the ResNet-9 model, which recognized the blight disease status in tomato and potato leaf images that farmers could deploy for their usage. A total of 3,990 data sets were used for training samples. The algorithm was trained with these parameters and tested on the test set, which consisted of 1,331 images. Saliency maps were used to explain the proposed model’s predictions and offer justification for its predictions, allowing for a thorough understanding of the model. In making its predictions, the ResNet-9 model was found to take into account the leaf shape, any diseased regions that were present, and the leaf’s overall green spots. This observation helped us better understand the prediction of the model.

A unique CNN architecture called Efficient Net was suggested in 2021 by Chih Chen et al. [7] employing low bandwidth imaging sensors. For this research, the researcher used negligible-bandwidth, a smaller scale IoT imaging sensors on a farm to periodically notice the cassava leaf images. To correctly categorize and assess the pathology of cassava plants, data enhancement techniques such TTA and cutout, cutmix, and k-fold were used. This study used many simulated trials to categorize and assess the illnesses identified in 5 datasets of cassava leaf samples. Despite some variations in the test set images, this framework was able to produce generally correct classification findings. This article obtained accuracy in classification of 89%, which was better than that of prior research.

Suganya et al. [10] introduced an image processing based method in 2020 that automatically recognized and classified the illnesses affecting groundnut leaves. The H2K approach was recommended for groundnut leaf diseases in order to reliably identify and classify them. The previous literature focused on diseases of leaves that were prevalent to all crops, but this research offered a novel, strong, and ideal technique for identifying and categorizing the diseases affecting groundnut crop. As a result, the H2K technique helped to increase crop output and yield.

In 2022, Mohd Ghazalli et al. [13] deployed image processing techniques to categorize cassava leaves into mosaics and bacteria related ailments. Here, SVM was utilized to identify the cassava leaf based on features like color and texture that were derived. It was discovered that SVM provides accuracy of 87.5%. SVM was therefore shown to be helpful in determining the presence of diseases in cassava leaves.

In 2022, Sabari et al. [32] have deployed a very efficient framework that may be used in the feature extraction stage to categorize various plant and fruit leaf diseases. It employs a modified deep transfer learning method for this purpose. To summarize, we extract characteristics using model engineering (ME). Several support vector machine (SVM) models are used to improve processing speed and feature discrimination. In the training step, the chosen model informs the determination of the radial basis function (RBF) kernel parameters.

In 2022, Kaur et al. [20] have looks into grapevine plant diseases. Leaf blight, black rot, stable, and black measles are the four diseases that grape plants are prone to. Despite several prior study suggestions using machine learning algorithms to identify one or two diseases in grape plant leaves, no one has presented a comprehensive detection of all four illnesses in grape plant leaves. The photos are from the plant village dataset, which is used to retrain the EfficientNet B7 deep architecture via transfer learning. After the transfer learning, a logistic regression technique is used to downsample the gathered features.

2.2. Research gaps

The existing research on Cassava disease faces numerous drawbacks, such as reduced recognition rate, high process time, and reduced precision. Some of the difficulties faced in reviewed existing works were as follows: Over fitting issue occurs while employing CNN model. Deep networks typically needs long training time, thus makes itnot capable for applications in real world. In, ECNN, higher processing time is challenging issue. In order to confirm symptoms and complement images, PCR tests can be required. InceptionV3 takes a long time and a lot of processing power. CNN model has poor generalization capability. Separable Convolutions UNet model requires ineffective ensemble of schemes of varied depths. In SVM model, there was a long training time for large datasets. Table 2 shows the review on the existing models.

Table 2
Reviews on existing models on plant disease classification

Author Methods Features Limitations

G. Sambasivam and Geoffrey Duncan Opiyo [33] DCNN model ∙ Encounter high-class imbalance Over fitting issue

∙ Accurately predict underrepresented classes.

Yiwei Zhong et al. [41] Transformer-Embedded ResNet model ∙ Extract robust features Needs long training time.

∙ Effective classification of disease.

Umesh Kumar Lilhore et al. [23] ECNN ∙ Achieves high accuracy Higher processing time

Trong et al. [21] CNN model ∙ High Accuracy. Real time analysis has to be considered.

∙ Average reduction in parameter consumption

Alberta Chen et al. [4] ResNet-9 model ∙ Achieve higher recall This model is not deployed with an interface.

∙ Higher F1-score

Chih-Cheng Chen et al. [7] CNN model ∙ Higher accuracy Poor generalization

Suganya et al. [10] H2K model ∙ High accuracy Requires extensive analysis on classification Time

∙ Highly robust.

Mohd Ghazalli et al. [13] SVM model ∙ More effective in high dimensional spaces Long training time for large datasets.

∙ Relatively memory efficient.

Author	Methods	Features	Limitations
G. Sambasivam and Geoffrey Duncan Opiyo [33]	DCNN model	∙ Encounter high-class imbalance	Over fitting issue
∙ Accurately predict underrepresented classes.
Yiwei Zhong et al. [41]	Transformer-Embedded ResNet model	∙ Extract robust features	Needs long training time.
∙ Effective classification of disease.
Umesh Kumar Lilhore et al. [23]	ECNN	∙ Achieves high accuracy	Higher processing time
Trong et al. [21]	CNN model	∙ High Accuracy.	Real time analysis has to be considered.
∙ Average reduction in parameter consumption
Alberta Chen et al. [4]	ResNet-9 model	∙ Achieve higher recall	This model is not deployed with an interface.
∙ Higher F1-score
Chih-Cheng Chen et al. [7]	CNN model	∙ Higher accuracy	Poor generalization
Suganya et al. [10]	H2K model	∙ High accuracy	Requires extensive analysis on classification Time
∙ Highly robust.
Mohd Ghazalli et al. [13]	SVM model	∙ More effective in high dimensional spaces	Long training time for large datasets.
∙ Relatively memory efficient.

3. Proposed DHCLDC model for underground crops

The developed DHCLDC model for underground crop disease classificationis given below. This work considers three crops like cassava, potato and groundnut. The initial step is pre-processing is carried out by deploying median filter that eliminates the unwanted noise. Then, deploys Improved U-net based segmentation model, that is, U-Net with nested convolutional block that including the convolutional and dropout layer in the nested architecture. Features like colour feature, shape feature and improved multi text on features are extracted. Finally, HC (CNN and LSTM) is deployed for classification of underground crop disease by training the extracted feature set.The classification outcome is obtained by the improved SLF. The adopted DHCLDC model for underground crops is represented in Fig. 1.

Fig. 1.

Diagrammatic demonstration of DHCLDC model for underground crops.

3.1. Pre-processing: Noise removal by MF

Pre-processing aims to improve the image data by reducing distortions and helps to enhance certain components that are essential for further processing.

MF is used for reducing noise and it effectively preserves edges. ‘Salt and pepper’ type noise is particularly well-admitted by this method. The term “Window” in MF refers to the neighborhood pattern, which moves pixel by pixel over the whole image. The median is evaluated by placing the pixel in the middle (median) of every value of a pixel from the window and then sorting them all into numerical order. This work deploys CV2. Median Blur () to compute the median of all pixels under the kernel window and then the central value is replaced with median value.

Then, improved U-Net scheme is used to segment the processed image, ${Im}^{mf}$ .

3.2. Improved U-Net (U-Net with nested convolutional block) for segmentation

U-Net is a network with training technique that makes better use of labeled data through data augmentation, enabling it to provide effective results with fewer samples. The loss function of U-Net model, however, has a number of drawbacks [34]. The U-Net model has advantages were carried over into an enhanced U-Net, which also made several additional improvements. Improved U Net with a novel loss function combines new elements including skip route and deep supervision. It corrects the data imbalance issue in image segmentation and produced effective segmentation outcomes [34]. Also, the architecture is improved by adding the nested convolutional block with the combination of convolutional layer and dropout layer, and is illustrated in Fig. 2.

Skip pathway: A set of NDCB connects decoder and encoder that make up U-Net. Closing the semantic gap between the encoder and decoder feature maps is the main objective of UNet before fusion. By simply substituting the proposed nested dense skip paths for the simple skip connections, UNet may be quickly deployed. The decoder receives the encoder’s feature maps directly in UNet, but UNet subjects them through a DCB. For instance, the skip path ways among nodes $X_{0, 0}$ and $X_{1, 3}$ includes a DCB with 3 layers, in which every convolution layer is leaded by a concatenated layer, which fused the output from the preceding convolution layer of similar dense block with the related up-sampled outcome of the low dense block [43].

Deep supervision: U Net, enables the approach to function in 2 modes: 1) precise mode in which the outputs from every segmentation twigs were averaged; 2) faster mode in which the last segmentation map gets elected from a segmentation branch, the selection of which portray the extent of speed gain and model pruning [43].

In traditional U Net, BCE is deployed as loss function as exposed in Eq. (1), in which, N indicate sample count, $x_{i}$ implies sample of truth label, $P_{i}$ signify predicted probability. $\begin{matrix} (1) & {BCE}_{loss} = \frac{- 1}{N} \sum_{i = 1}^{N} [x_{i} log ({\hat{x}}_{i}) + (1 - x_{i}) * log (1 - {\hat{x}}_{i})] \end{matrix}$

In enhanced UNet, rather than BCE loss, we have combined the weighted BCE loss with dice loss as provided in Eq. (2), where, ψ indicates tuning parameter, ψ can be used to tune FP and FN, e.g. If we set $ψ ⩾ 1$ , we can reduce the count of FN, similarly, if $ψ < 1$ , we can reduce the count of FP. $x_{i}$ implies target label, ${\hat{x}}_{i}$ implies predicted label and N implies training sample count. $\begin{matrix} (2) & Loss (x_{i}, {\hat{x}}_{i}) = \frac{- 1}{N} \sum_{i = 1}^{N} [\frac{ψ}{2} . x_{i} log ({\hat{x}}_{i}) + \frac{2 . x_{i} . {\hat{x}}_{i}}{x_{i} + {\hat{x}}_{i}}] \end{matrix}$

The enhanced segmented image is implied by ${Im}^{IU}$ .

Fig. 2.

Illustration of improved U-Net (U-Net with nested convolutional block) for segmentation.

4. Deriving color feature, shape feature and improved multi texton features

Extracting features enhances the efficacy of ML by eliminating unwanted and superfluous data by reducing the noise. From ${Im}^{IU}$ , thecolour feature, shape feature and improved MT featuresare drawn.

4.1. Improved MT features

In MTH oriented image retrieving method [24], orientation of texture should be recognized for text on analysis. Orientation of texture could also be deployed for estimating the textured image shape. Using specific gradient operators in both vertical and horizontal orientations on a greyscale image, 2 gradient images, signified as $h_{χ}$ and $h_{κ}$ are obtained. A gradient map $h (χ, κ)$ is attained, with the gradient orientation and magnitude defined as in Eq. (4) and (3). $\begin{array}{c} (3) & | h (χ, κ) | = \sqrt{h_{χ}^{2} + h_{κ}^{2}} \\ (4) & θ (χ, κ) = arctan (\frac{h_{κ}}{h_{χ}}) \end{array}$

The gradients in χ and κ orientations are signified by 2 vectors $b (r_{χ}, g_{χ}, b_{χ})$ and $a (r_{κ}, g_{κ}, b_{κ})$ , in which, $r_{χ}$ represents gradient in r channel at parallel direction. Their dot product and norm are given in Eq. (5)–Eq. (7). $\begin{array}{c} (5) & | b | = \sqrt{{(r_{χ})}^{2} + {(g_{χ})}^{2} + {(b_{χ})}^{2}} \\ (6) & | a | = \sqrt{{(r_{κ})}^{2} + {(g_{κ})}^{2} + {(b_{κ})}^{2}} \\ (7) & b . a = r_{χ} . r_{κ} + g_{χ} . g_{κ} + b_{χ} . b_{κ} \end{array}$

The angle among b and a is modeled as in Eq. (8). $\begin{matrix} (8) & cos (χ, κ) = (\frac{b . a}{| b | . | a |}) \end{matrix}$

An alternative orientation strategy is used for colored images, as shown in Eq. (9). $\begin{matrix} (9) & MT = θ (χ, κ) = arccos (\frac{b . a}{| b | . | a |}) \end{matrix}$

As per improved MT, the orientation is modeled as in Eq. (10), where $Rd$ indicates radian value, ${Dis}_{ba}$ refers to distance. $\begin{array}{c} (10) & IMT = θ (χ, κ) = arccos (\frac{b . a}{| b | . | a | + {Dis}_{ba}}) * Rd \\ (11) & Rd = \frac{180}{π} \\ (12) & {Dis}_{ba} = \frac{cos (b . a) + (| b | - | a |)}{1 + | b | + | a |} \end{array}$

4.2. Shape features

The shape features considered are area, perimeter, approximation and convexity. The input image is transformed to grey scale image prior to the extraction of shape features. The shape features are implied as $SF$ .

4.3. Colour features

The color features like mean, histogram, standard deviation, and median, frequency of maximal color and frequency of minimal color are extracted. These features are extracted for each B, G and R channels separately. The color features are implied as $CF$ .

The modified colour feature, shape feature and improved MT features are entirely implied by $FS$ .

5. HybridClassifiers with improved score level fusion

Hybrid Model: The method of combining two different classifiers for classification purpose is known as a hybrid mode of classification. This HC model enhances the accuracy of classification over individual classification models. Here, we combine the CNN and LSTM models to present an HC model.

5.1. CNN

The features $FS$ are provided to CNN for DHCLDC of underground crops. A CNN’s principal function is to transmit local features from higher-layered inputs to lower-layered inputs for more complicated information. CNNs are a subset of machine learning techniques used in signal classification. CNN does not require human assistance to identify the necessary elements. Equation (13), which describes the operation of the convolution layer, in which $FS$ indicates the input retrieved feature set, M represents the 2D filter size $m \times n$ , and $C^{out}$ signifies the predicted CNN score. The operating function of the convolutional layer is provided as $FS * M$ . ReLU layer is carried out for feature maps to improve nonlinearity, and Eq. (14) provides the CNN mathematical formula. $\begin{array}{c} (13) & C^{out} (i, j) = (FS * M) (i, j) = \sum \sum C^{out} (i + m, j + n) M (m, n) \\ (14) & f (x) = max (0, x) \end{array}$

The pooling lowers the number of parameters by stacking down samples of a specified input size. In an input area, max pooling – the most popular method – yields the highest value.

Advantages of CNN

No need for human oversight is necessary.

Incredibly precise in both image identification and categorization.

Sharing of weight.

Cuts down on computation.

Applies the same knowledge to every location in an image.

The capacity to manage big datasets.

Learning in hierarchies.

5.2. LSTM

The features $FS$ are provided to LSTM for DHCLDC of underground crops. LSTM include “input, output, and forget gates” [42]. The neurons in LSTM are evaluated as exposed in Eq. (15)–(19), and input layer is implied by $({FS}_{1}, {FS}_{2}, \dots, {FS}_{3})$ . $\begin{array}{c} (15) & ℓ_{t} = K (w_{ℓ} \times [H_{t - 1}, {FS}_{t}] + o_{ℓ}) \\ (16) & κ_{t} = K (w_{κ} \times [H_{t - 1}, {FS}_{t}] + o_{κ}) \\ (17) & ρ_{t} = K (w_{ρ} \times [c_{t}, H_{t - 1}, {FS}_{t}] + o_{ρ}) \\ (18) & c_{t} = κ_{t} \times c_{t - 1} + ℓ_{t} \times tan H (w_{c} \times [H_{t - 1}, {FS}_{t}]) \\ (19) & H_{t} = ρ_{t} \times tan H (c_{t}) \end{array}$

Here,

t = time

$ℓ_{t}$ , $κ_{t}$ , $ρ_{t} =$ input of input and output gate

$K =$ sigmoid

${FS}_{t} =$ input neuron

$o_{l}$ , $o_{m}$ , $o_{n} =$ offset vector

$H_{t} =$ output of hidden layer

$w_{l}$ , $w_{m}$ , $w_{n} =$ weight matrices of input and output cell.

Advantages of LSTM

For time series applications, LSTM are superior than basic RNNs in a number of ways.

They are superior to RNNs at capturing long-term dependencies since RNNs have a tendency to forget about remote previous inputs.

For time series with long-term cycles or patterns, like meteorological data or economic indicators, this is crucial. Table 3 represents the classifier hyperparameteres.

Table 3
Hyper parameters of classifiers

Methods Hyperparameters

CNN Number of input layer:1

Number of hidden layers:1

Number of convolution layer:1,

filter size:32, activation=‘relu’

Number of pooling layer:1

optimizer: adam

activation function: softmax

batch size:100

loss=‘sparse_categorical_crossentropy’

epoch=25

LSTM Number of input layers:1

Number of LSTM layer:1, LSTM unit=16

Number of output layer:1

optimizer: sgd

activation function: sigmoid

batch size:128

loss=‘categorical_crossentropy’

epoch=25

Methods	Hyperparameters
CNN	Number of input layer:1
	Number of hidden layers:1
	Number of convolution layer:1,
	filter size:32, activation=‘relu’
	Number of pooling layer:1
	optimizer: adam
	activation function: softmax
	batch size:100
	loss=‘sparse_categorical_crossentropy’
	epoch=25
LSTM	Number of input layers:1
	Number of LSTM layer:1, LSTM unit=16
	Number of output layer:1
	optimizer: sgd
	activation function: sigmoid
	batch size:128
	loss=‘categorical_crossentropy’
	epoch=25

5.3. Improved SLF

The improved SLF is performed by getting MSE from CNN and LSTM classifiers. The similarity between the input and template biometric feature vectors is gauged by the match score. Fusion is referred to as occurring at the match score level when the combined match scores from various biometric matchers lead to a final recognition determination. This is sometimes referred to as measurement-or confidence-level-level fusion. In addition to the unprocessed data and feature vectors, the match scores hold the most comprehensive information regarding the input pattern. Additionally, combining and accessing the scores produced by several biometric matchers is not too difficult. $\begin{matrix} (20) & Err = [{CNN}_{Err}, {LSTM}_{Err}] \end{matrix}$

The predicted scores of CNN denoted by $S_{i, CNN}$ and predicted scores of LSTM $S_{i, LSTM}$ are normalized as shown in Eq. (21) and (22). $\begin{array}{c} (21) & S_{i, CNN} = S_{i, CNN} / max (S_{i, CNN}) \\ (22) & S_{i, LSTM} = S_{i, LSTM} / max (S_{i, LSTM}) \end{array}$

Then, the mean of CNN and LSTM are determined as shown in Eq. (23) and (24). $\begin{array}{c} (23) & μ_{CNN} = \frac{1}{N} \sum_{i = 1}^{N} S_{i, CNN} \\ (24) & μ_{LSTM} = \frac{1}{N} \sum_{i = 1}^{N} S_{i, LSTM} \end{array}$

The deviation of CNN and LSTM are computed as in Eq. (25) and (26). $\begin{array}{c} (25) & S_{CNN} = \frac{1}{N} \sum_{i = 1}^{N} {(S_{i, CNN} - μ_{CNN})}^{2} \\ (26) & S_{LSTM} = \frac{1}{N} \sum_{i = 1}^{N} {(S_{i, LSTM} - μ_{LSTM})}^{2} \end{array}$

Thus, list of S and MSE error is obtained as shown in Eq. (27) and (28). $\begin{array}{c} (27) & S list = [S_{CNN}, S_{LSTM}] \\ (28) & MSE error list = [{CNN}_{MSE}, {LSTM}_{MSE}] \end{array}$

Subsequently, get the minimum MSE error from both CNN and LSTM [36]. Based on minimum index, get the weight value, e.g. if ${CNN}_{MSE}$ is minimum, the SLF is modeled as in Eq. (29). If ${LSTM}_{MSE}$ is minimum, the SLF is modeled as in Eq. (30). $\begin{array}{c} (29) & ISLF = W_{1} * S_{i, CNN} + (1 - W_{1}) S_{i, LSTM} \\ (30) & ISLF = W_{1} * S_{i, LSTM} + (1 - W_{1}) S_{i, CNN} \end{array}$

Thus, the cassava dataset outcomes are classified as Healthy, CBSD, CBB, CMD and CGM. The potato dataset outcomes are classified as healthy, Early blight, Late blight. The groundnut dataset outcomes are classified as early leaf spot, early rust, healthy leaf, late leaf spot, nutrition deficiency and rust.

6. Results and discussions

6.1. Experimental set up

The HC (CNN + LSTM) model for DHCLDC of underground crops was done in Python. Here, the extant schemes like DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13] models were compared over presented HC (CNN + LSTM) method. The following performance measures are used to analyze plant disease detection.

Accuracy

Accuracy is measured using a set of measurements and their real results. Equation (31) establishes the computed accuracy formula, in which $TP$ denotes true positive, $TN$ indicates true negative, $FP$ is signifies as false positive, and $FN$ denotes false negative. $\begin{matrix} (31) & Accuracy = \frac{TN + TP}{FP + FN + TP + TN} \end{matrix}$

Precision

The percentage of correctly classified instances serves as a proxy for precision. It states this in Eq. (32). $\begin{matrix} (32) & Precision = \frac{TP}{FP + TP} \end{matrix}$

Sensitivity

The probability of a positive test result is described by sensitivity, which only includes positive true values. The expression for it is Eq. (33). $\begin{matrix} (33) & Sensitivity = \frac{TP}{FN + TP} \times 100 \end{matrix}$

Specificity

Specificity is the probability of just having actual negative values. Equation (34) displays this equation $\begin{matrix} (34) & Specificity = \frac{TN}{FP + TN} \times 100 \end{matrix}$

FNR

It’s defined as the possibility that a test will overlook a real positive. Equation (35) is explained. $\begin{matrix} (35) & FNR = FN (FN + TP) \end{matrix}$

FPR

Regardless of whether the test is a machine learning model, its false positive rate (FPR) serves as a gauge of its effectiveness in Eq. (36) $\begin{matrix} (36) & FPR = \frac{FP}{FP + TN} \end{matrix}$

FDR

The predicted ratio of the total number of positive classifiers to the total number of erroneous positive classifications is known as the false positive rate, or FDR. It is described in Eq. (37). $\begin{matrix} (37) & FDR = \frac{FP}{FP + TP} \end{matrix}$

F-measure

To evaluate the accuracy results, the F-measure is employed. The Eq. (38) represents it. $\begin{matrix} (38) & F - measure = \frac{TP}{1 / 2 (FP + FN) + TP} \end{matrix}$

MCC

Using a contingency matrices technique, the Matthews correlation coefficient is used to calculate the Pearson coefficient for the product-moment correlation between real and expected data. This alternative metric is unrelated to the problem of imbalanced datasets. It can be seen in Eq. (39). $\begin{matrix} (39) & MCC = \frac{TP . TN - FP . FN}{\sqrt{(TP + FP) . (FN + TP) . (TN + FP) . (TN + FN)}} \end{matrix}$

NPV

Equation (40) denotes negative predictive value. $\begin{matrix} (40) & NPV = \frac{TN}{TN + FN} \end{matrix}$ Table 4 shows the simulation setup. The device and window specification are given below.

Table 4
Simulation setup

Device and Window specification

Installed RAM 16.0 GB (13.9 GB usable)

Product ID 00330-80000-00000-AA360

System Type 64-bit operating system, x64-based processor

Pen and touch No pen or touch input is available for this device

Edition Windows 11 Pro

Version 22H2

OS 22621.2283

Experience Windows Feature Experience Pack 1000.22

Device and Window specification
Installed RAM	16.0 GB (13.9 GB usable)
Product ID	00330-80000-00000-AA360
System Type	64-bit operating system, x64-based processor
Pen and touch	No pen or touch input is available for this device
Edition	Windows 11 Pro
Version	22H2
OS	22621.2283
Experience	Windows Feature Experience Pack 1000.22

6.2. Dataset description

The data was gathered from [1,29] and [14] for Cassava, Potato and Groundnut crops that were described here as dataset 1, 2 and 3 respectively. Cassava Dataset: Only when the test is turned in for scoring will your notebook get access to the entire collection of test images. The test set should contain about fifteen thousand photos. Farmers may find a quick automated turnaround to be particularly helpful in certain situations where burning the contaminated plants is the primary treatment to stop the disease from spreading. Potato dataset: 1500 picture files from three distinct classes – healthy, late blight, and early blight – make up this dataset. The groundnut dataset allowed us to count the number of leaves in a particular picture. The dataset is a component of a full-featured mobile machine learning tool that assesses the degree of groundnut plant disease. The dataset includes XML files with labels for the photos. Figure 3 exposes the sample images.

Fig. 3.

Sample images for datasets a) 1 b) 2 and c) 3.

6.3. Performance analysis

The evaluation of the HC (CNN + LSTM) over DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13] is exposed in Fig. 4–12 for three datasets (Cassava, Potato and Groundnut datasets). For every metric taken into consideration, the developed HC (CNN + LSTM) has shown improved values. Regarding the Cassava dataset, the HC (CNN + LSTM) accuracy is increasing from the 60th to the 90th LP. The accuracy of HC (CNN + LSTM) is 88.175 at the 60th LP, 92.46 at the 70th LP, 94.69 at the 80th LP, and 95.548 at the 90th LP. This shows that, better performance is obtained when LP is 90. The FNR of HC (CNN + LSTM) in the potato dataset decreases from the 60th to the 90th LP. The specificity of DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13] is 90.3, 85.86, 86.65, 93.65, 93.73, 90.95, 93.34 and 93.30 respectively for dataset 3. When compared to DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13], the specificity of HC (CNN + LSTM) is high around 95.41. Better results are obtained because a modified Unet model with increased multi-texon characteristics is used. Furthermore, improved SLF with hybrid CNN and LSTM models also contributed to better plant leaf classification. As a result, the suggested model’s superiority over the other models in use is demonstrated.

Fig. 4.

Examination on HC (CNN + LSTM) method for DHCLDC of underground crops for a) specificity b) sensitivity c) accuracy d) precision for dataset 1.

Fig. 5.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) FNR b) FPR for dataset 1.

Fig. 6.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) NPV b) MCC and (c) F measure for dataset 1.

Fig. 7.

Analysis of HC (CNN + LSTM)method for DHCLDC of underground crops for a) specificity b) sensitivity c) accuracy d) precision for dataset 2.

Fig. 8.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) FNR b) FPR for dataset 2.

Fig. 9.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) NPV b) MCC and (c) F measure for dataset 2.

Fig. 10.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) specificity b) sensitivity c) accuracy d) precision for dataset 3.

Fig. 11.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) FNR b) FPR for dataset 3.

Fig. 12.

Analysis of HC (CNN + LSTM) model for DHCLDC of underground crops for a) NPV b) MCC and (c) F measure for dataset 3.

6.4. Ablation study

The ablation analysis for HC (CNN + LSTM) over suggested with existing U-Net, proposed with existing multitex on, and suggested with existing SLF is shown in Table 5–7. The DHCLDC scheme’s net present value (NPV) while utilizing the HC (CNN + LSTM) technique is 0.912773; in contrast, the NPV when utilizing the proposed U Net, extant multitex, and extant SLF approaches are 0.781332, 0.785736, and 0.903983. Furthermore, the precision obtained with the HC (CNN + LSTM) technique is 0.932977, whereas the precision obtained with the HC (CNN + LSTM) technique with the existing U Net, the proposed multitex on, and the proposed SLF are 0.784864, 0.788046, and 0.908209.Better results are obtained by using an updated U net model with enhanced multi-texon characteristics. Additionally, updated SLF hybrid CNN and LSTM models helped to improve plant leaf classification.

6.5. Statistical analysis

Table 8 shows accuracy on HC (CNN + LSTM)over DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13]. The HC (CNN + LSTM) achieved the greater accuracy of 0.955484 for maximum case, while DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13] have achieved greater accuracy of 0.92758, 0.826896, 0.893675, 0.917395, 0.900547, 0.934243, 0.914368 and 0.926304 for dataset 1. The median case using HC (CNN + LSTM) is0.93208, while, DBN, NN, RF, KNN, CNN, LSTM, DCNN [33] and SVM [13]gained less accuracy of 0.880981, 0.776491, 0.859265, 0.880863, 0.885786, 0.870687, 0.875057 and 0.874353. Improved U net model with expanded multi-texon characteristics yields better outcomes. Additionally, improved SLF hybrid CNN and LSTM models produced superior plant leaf categorization.

Table 5
Ablation study on developedDHCLDC approach for dataset 1

Metrics Proposed with extant U-Net Proposed with extant multitexon Proposed with extant SLF HC (CNN + LSTM)

MCC 70.05% 70.81% 81.01% 87.64%

Sensitivity 80.16% 80.49% 92.75% 92.98%

FPR 10.29% 9.86% 11.96% 8.27%

Precision 78.49% 78.80% 90.82% 93.30%

F-measure 79.30% 79.63% 91.76% 93.14%

Accuracy 78.35% 78.72% 90.65% 92.46%

NPV 78.13% 78.57% 90.40% 91.28%

Specificity 76.10% 76.52% 88.04% 91.73%

FNR 6.23% 5.90% 7.25% 7.02%

Metrics	Proposed with extant U-Net	Proposed with extant multitexon	Proposed with extant SLF	HC (CNN + LSTM)
MCC	70.05%	70.81%	81.01%	87.64%
Sensitivity	80.16%	80.49%	92.75%	92.98%
FPR	10.29%	9.86%	11.96%	8.27%
Precision	78.49%	78.80%	90.82%	93.30%
F-measure	79.30%	79.63%	91.76%	93.14%
Accuracy	78.35%	78.72%	90.65%	92.46%
NPV	78.13%	78.57%	90.40%	91.28%
Specificity	76.10%	76.52%	88.04%	91.73%
FNR	6.23%	5.90%	7.25%	7.02%

Table 6

Ablation study on developed DHCLDC approach for dataset 2

Metrics	Proposed with extant U-Net	Proposed with extant multitexon	Proposed with extant SLF	HC (CNN + LSTM)
NPV	87.93%	85.81%	92.39%	92.69%
Sensitivity	88.11%	86.05%	94.20%	93.30%
FNR	5.24%	4.84%	5.80%	6.70%
Precision	86.57%	84.65%	92.06%	93.59%
F-measure	87.33%	85.33%	93.11%	93.44%
MCC	81.17%	79.58%	84.19%	89.30%
Accuracy	87.25%	85.23%	92.21%	93.17%
FPR	6.91%	6.46%	10.27%	6.97%
Specificity	86.43%	84.43%	89.73%	93.03%

Table 7

Ablation study on developed DHCLDC approach for dataset 3

Measures	Proposed with extant U-Net	Proposed with extant multitexon	Proposed with extant SLF	HC (CNN + LSTM)
Specificity	86.78%	77.31%	90.47%	92.86%
MCC	82.09%	60.26%	84.90%	89.02%
Accuracy	87.33%	80.35%	91.44%	93.50%
Precision	86.58%	81.60%	92.54%	92.74%
FPR	5.79%	22.69%	9.53%	7.14%
Sensitivity	87.94%	82.85%	92.12%	94.18%
NPV	88.03%	78.77%	89.89%	94.24%
F-measure	87.25%	82.22%	92.33%	93.45%
FNR	4.64%	17.15%	7.88%	5.82%

Table 8

Statistical analysis onaccuracy for DHCLDC approach

Methods	Median	Mean	SD	Max	Min
Dataset 1
DBN	0.834158	0.835198	0.077911	0.92758	0.744897
NN	0.747832	0.74321	0.073827	0.826896	0.650279
RF	0.839079	0.835058	0.053668	0.893675	0.768398
KNN	0.884944	0.87238	0.05283	0.917395	0.802237
CNN	0.82299	0.8209	0.06757	0.900547	0.737075
LSTM	0.842123	0.84106	0.080413	0.934243	0.745753
DCNN [33]	0.824113	0.821758	0.083688	0.914368	0.724437
SVM [13]	0.847858	0.843748	0.074809	0.926304	0.752974
HC(CNN + LSTM)	0.935781	0.927199	0.032975	0.955484	0.881751
Dataset 2
DBN	0.880981	0.875833	0.060672	0.939296	0.802073
NN	0.776491	0.770677	0.060509	0.831967	0.697759
RF	0.859265	0.849619	0.040771	0.885049	0.794897
KNN	0.880863	0.876227	0.059064	0.939043	0.80414
CNN	0.885786	0.876114	0.064829	0.939631	0.793252
LSTM	0.870687	0.862733	0.061226	0.923482	0.786075
DCNN [33]	0.875057	0.867852	0.068865	0.938063	0.783229
SVM [13]	0.874353	0.869903	0.062499	0.937112	0.793795
HC(CNN + LSTM)	0.93208	0.930663	0.019405	0.952925	0.905568
Dataset 3
DBN	0.891485	0.884903	0.036094	0.919735	0.836908
NN	0.820934	0.811951	0.05283	0.859043	0.746894
RF	0.862509	0.858728	0.030413	0.891523	0.818371
KNN	0.906251	0.901948	0.037053	0.939291	0.856
CNN	0.898421	0.893997	0.045115	0.940495	0.838651
LSTM	0.885456	0.868831	0.052979	0.909838	0.794571
DCNN [33]	0.891118	0.885138	0.048511	0.93389	0.824424
SVM [13]	0.895464	0.890467	0.045825	0.936171	0.83477
HC(CNN + LSTM)	0.941327	0.938866	0.018098	0.957369	0.915441

6.6. Jaccard and Dice scores

Table 9 reveals the evaluation on segmentation accuracy, Jaccard and Dice scores. The evaluation is done on K means, FCM and conventional U-net schemes over improved U-net segmentation method. For 3 datasets, the segmentation accuracy, Jaccard and Dice scores attained high values for improved U-net approach.

Table 9
Analysis on segmentation accuracy and Jaccard, Dice scores

Images K means FCM Extant U-Net Improved U-Net

Dataset 1

Dice 78.23% 64.20% 84.90% 86.30%

Jaccard 72.07% 65.49% 71.17% 77.60%

Segmentation accuracy 76.81% 76.28% 85.18% 91.73%

Dataset 2

Dice 71.04% 78.67% 88.39% 90.84%

Jaccard 64.62% 65.78% 72.53% 81.99%

Segmentation accuracy 81.08% 79.58% 86.87% 93.08%

Dataset 3

Dice 72.15% 71.73% 89.90% 91.90%

Jaccard 69.65% 65.14% 78.58% 80.80%

Segmentation accuracy 72.13% 74.27% 89.26% 92.43%

Images	K means	FCM	Extant U-Net	Improved U-Net
Dataset 1
Dice	78.23%	64.20%	84.90%	86.30%
Jaccard	72.07%	65.49%	71.17%	77.60%
Segmentation accuracy	76.81%	76.28%	85.18%	91.73%
Dataset 2
Dice	71.04%	78.67%	88.39%	90.84%
Jaccard	64.62%	65.78%	72.53%	81.99%
Segmentation accuracy	81.08%	79.58%	86.87%	93.08%
Dataset 3
Dice	72.15%	71.73%	89.90%	91.90%
Jaccard	69.65%	65.14%	78.58%	80.80%
Segmentation accuracy	72.13%	74.27%	89.26%	92.43%

6.7. Practical implications

Plant diseases are detectable from the start. It is possible to keep disease outbreaks at a strategic distance.

Through GSM interface, information about the plants and framework will be recommended to the client.

Depending on our needs, this can be set to take any form of leave.

This can be changed to accommodate any kind of leaf depending on our requirement.

When a plant disease detection system detects common irregularities on the leaves, like mold or burning, it can function as a universal detector.

7. Conclusion

This paper proposes a novel DHCLDC technique for underground crops. In this, preprocessing was done via deploying median filter. Then, improved U-net model was carried out for segmentation process. Moreover, the features extracted comprised of color features, shape features and improved multi texton features. Finally, HC model was suggested for DHCLDC that combines CNN and LSTM models. The outputs from HC (CNN + LSTM) were then subjected to improved SLF from which final detected outcomes were attained. Regarding the Cassava dataset, the HC (CNN + LSTM) accuracy was increasing from the 60th to the 90th LP. In the 60th LP, the HC (CNN + LSTM) accuracy was 88.175; in the 70th LP, it was 92.46; in the 80th LP, it was 94.69; and in the 90th LP, it was 95.548. This demonstrates that when LP is 90, higher performance was achieved. Additionally, the entire leaf area affected by the infection is identified by this research. Early plant disease identification is an extremely difficult endeavor in existing methods. But, this system will detect the plant disease earlier. The limitation of the suggested approach is processing time is not taken into account. We intended to apply the same technique to smartphones, making it simple for farmers to receive results within the allotted time. Future research on processing times must be conducted.

References

https://www.kaggle.com/competitions/cassava-leaf-disease-classification/data.

Ahmad,

Abdullah,

Moon and

Han, Plant disease detection in imbalanced datasets using efficient convolutional neural networks with stepwise transfer learning, IEEE Access 9 (2021), 140565–140580. doi:10.1109/ACCESS.2021.3119655.

AlSuwaidi,

Grieve and

Yin, Feature-ensemble-based novelty detection for analyzing plant hyperspectral datasets, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11(4) (2018), 1041–1055. doi:10.1109/JSTARS.2017.2788426.

A.O.

Anim-Ayeko,

Schillaci and

Lipani, Automatic blight disease detection in potato (Solanum tuberosum L.) and tomato (Solanum lycopersicum, L. 1753) plants using deep learning, Smart Agricultural Technology 10 (2023), 100178.

Bizabani,

S.J.

Rogans and

M.E.

Rey, Differential miRNA profiles in South African cassava mosaic virus-infected cassava landraces reveal clues to susceptibility and tolerance to cassava mosaic disease, Virus Research 303 (2021), 198400.

Charoenvilaisiri,

Seepiban,

Kumpoosiri,

Rukpratanporn,

Warin,

Phuangrat,

Chitchuea,

Siripaitoon,

Chatchawankanphanich and

Gajanandana, Development of a triple antibody sandwich enzyme-linked immunosorbent assay for cassava mosaic disease detection using a monoclonal antibody to Sri Lankan cassava mosaic virus, Virology Journal 18(1) (2021), 100. doi:10.1186/s12985-021-01572-6.

C.C.

Chen,

J.Y.

Ba,

T.J.

Li,

C.C.

Chan,

K.C.

Wang and

Liu, EfficientNet: A low-bandwidth iot image sensor framework for cassava leaf disease classification, Sens. Mater. 33 2021, 4031–4044.

W.L.

Chen,

Y.B.

Lin,

F.L.

Ng,

C.Y.

Liu and

Y.W.

Lin, RiceTalk: Rice blast detection using Internet of things and artificial intelligence technologies, IEEE Internet of Things Journal 7(2) (2019), 1001–1010. doi:10.1109/JIOT.2019.2947624.

G.A.

da Costa,

R.L.

Freitas-Lopes,

J.B.

Lino,

E.C.

Julião,

S.J.

Michereff and

U.P.

Lopes, An efficient method for inducing sporulation of cercosporoid fungal species causing leaf spots in cassava, Journal of Plant Pathology 102 (2020), 201–204.

10.

K.S.

Devi,

Srinivasan and

Bandhopadhyay, H2K–a robust and optimum approach for detection and classification of groundnut leaf diseases, Computers and Electronics, in Agriculture 178 (2020), 105749.

11.

Dwivedi,

Dey,

Chakraborty and

Tiwari, Grape disease detection network based on multi-task learning and attention features, IEEE Sensors Journal 21(16) (2021), 17573–17580. doi:10.1109/JSEN.2021.3064060.

12.

Erick and

Mayengo, Modelling the dynamics of cassava mosaic disease with non-cassava host plants, Informatics in Medicine Unlocked 33 (2022), 101086.

13.

H.I.

Ghazalli and

N.A.

Roslan, Bacterial blight and mosaic disease detection on cassava leaf using support vector machine, Journal of Islamic 7(46) (2022).

14.

Groundnut-https://www.kaggle.com/datasets/muhammadazeemabbas/groundnut-leaves-dataset.

15.

Hormhuan,

Viboonjun,

Sojikul and

Narangajavana, Enhancing of anthracnose disease resistance indicates a potential role of antimicrobial peptide genes in cassava, Genetica 148 (2020), 135–148. doi:10.1007/s10709-020-00097-0.

16.

Huang,

Zhou,

He,

Chen,

Zhang and

Hu, Detection of peach disease image based on asymptotic non-local means and PCNN-IPELM, IEEE Access 8 (2020), 136421–136433. doi:10.1109/ACCESS.2020.3011685.

17.

D.O.

Igwe,

C.B.

Anyanwu,

C.A.

Afiukwa,

C.V.

Nnamani,

F.N.

Nweke,

G.N.

Ude and

B.E.

Ubi, Phenotypic and molecular screenings for determination of cassava mosaic disease (CMD) status in farmers’ fields in Ebonyi State, Nigeria, Molecular Biology Reports 48 (2021), 227–240. doi:10.1007/s11033-020-06039-5.

18.

E.C.

Julião,

M.D.

Santana,

R.D.

Freitas-Lopes,

A.D.

Vieira,

J.S.

de Carvalho and

U.P.

Lopes, Reduction of brown leaf spot and changes in the chlorophyll a content induced by fungicides in cassava plants, European Journal of Plant Pathology 157 (2020), 433–439. doi:10.1007/s10658-020-02001-0.

19.

Kanju,

V.N.

Uzokwe,

Ntawuruhunga,

Tumwegamire,

Yabeja,

Pariyo and

Kawuki, Varietal response of cassava root yield components and root necrosis from cassava brown streak disease to time of harvesting in Uganda, Crop Protection 120 (2019), 58–66.

20.

Kaur,

Harnal,

Tiwari,

Upadhyay,

Bhatia,

Mashat and

A.M.

Alabdali, Recognition of leaf disease using hybrid convolutional neural network by applying feature reduction, Sensors 22(2) (2022), 575. doi:10.3390/s22020575.

21.

T.Y.

Lee,

I.A.

Lin,

J.Y.

Yu,

J.M.

Yang and

Y.C.

Chang, High efficiency disease detection for potato leaf with convolutional neural network, SN Computer Science 2(4) (2021), 297.

22.

I.C.

Leite,

F.A.

Lima Filho,

R.L.

Freitas-Lopes,

S.J.

Michereff,

A.S.

Capucho and

U.P.

Lopes, Standard area diagrams to aid assessments of the severity of blight leaf spot symptoms on cassava leaves, Journal of Plant Pathology 101 (2019), 155–160. doi:10.1007/s42161-018-0132-9.

23.

U.K.

Lilhore,

A.L.

Imoize,

C.C.

Lee,

Simaiya,

S.K.

Pani,

Goyal,

Kumar and

C.T.

Li, Enhanced convolutional neural network model for cassava leaf disease identification and classification, Mathematics 10(4) (2022), 580.

24.

G.H.

Liu,

Zhang,

Y.K.

Hou,

Z.Y.

Li and

J.Y.

Yang, Image retrieval based on multi-texton histogram, Pattern Recognition 43(7) (2010), 2380–2389. doi:10.1016/j.patcog.2010.02.012.

25.

E.J.

McCallum,

R.B.

Anjanappa and

Gruissem, Tackling agriculturally relevant diseases in the staple crop cassava (Manihot esculenta), Current Opinion in Plant Biology 38 (2017), 50–58.

26.

G.M.

Oktavian and

Santoso, Leveraging MobileNet, InceptionV3, and CropNet to classify cassava plant disease, MIND (Multimedia Artificial Intelligent Networking Database) Journal 6(2) (2021), 183–193.

27.

Olasanmi,

Kyallo and

Yao, Marker-assisted selection complements phenotypic screening at seedling stage to identify cassava mosaic disease-resistant genotypes in African cassava populations, Scientific Reports 11(1) (2021), 1–8.

28.

Owomugisha,

Melchert,

Mwebaze,

J.A.

Quinn and

Biehl, Matrix relevance learning from spectral data for diagnosing cassava diseases, IEEE Access 9 (2021), 83355–83363.

29.

Potato-https://www.kaggle.com/datasets/muhammadardiputra/potato-leaf-disease-dataset.

30.

Ramcharan,

McCloskey,

Baranowski,

Mbilinyi,

Mrisho,

Ndalahwa,

Legg and

D.P.

Hughes, A mobile-based deep learning model for cassava disease diagnosis, Frontiers in plant science (2019), 272.

31.

P.K.

Rao, Cassava leaf disease classification using separable convolutions UNet, Turkish Journal of Computer and Mathematics Education (TURCOMAT) 12(7) (2021), 140–145.

32.

Saberi Anari, A hybrid model for leaf diseases classification based on the modified deep transfer learning and ensemble approach for agricultural aiot-based monitoring, Computational Intelligence and Neuroscience 2022 (2022).

33.

G.A.

Sambasivam and

G.D.

Opiyo, A predictive machine learning application in agriculture: Cassava disease detection and classification with imbalanced dataset using convolutional neural networks, Egyptian informatics journal 22(1) (2021), 27–34.

34.

Shi,

Duan,

Yang,

Feng,

Ding and

Jiang, An improved U-net image segmentation method and its application for metallic grain size statistics, Materials 15(13) (2022), 4417.

35.

U.P.

Singh,

S.S.

Chouhan,

Jain and

Jain, Multilayer convolution neural network for the classification of mango leaves infected by anthracnose disease, IEEE access 27(7) (2019), 43721–43729. doi:10.1109/ACCESS.2019.2907383.

36.

Sowmya and

Jose, Contemplate on ECG signals and classification of arrhythmia signals using CNN-LSTM deep learning model, Measurement: Sensors 24 (2022), 100558.

37.

C.K.

Sunil,

C.D.

Jaidhar and

Patil, Cardamom plant disease detection approach using EfficientNetV2, IEEE Access 10 (2021), 789–804.

38.

Udutalapally,

S.P.

Mohanty,

Pallagani and

Khandelwal, sCrop: A novel device for sustainable automatic disease prediction, crop selection, and irrigation in Internet-of-agro-things for smart agriculture, IEEE Sensors Journal 21(16) (2020), 17525–17538. doi:10.1109/JSEN.2020.3032438.

39.

Yang,

Chen,

He,

Yan,

Guo and

Ding, Self-supervised collaborative multi-network for fine-grained visual categorization of tomato diseases, IEEE Access 8 (2020), 211912–211923. doi:10.1109/ACCESS.2020.3039345.

40.

Yuan,

Xu and

Lu, SPEDCCNN: Spatial pyramid-oriented encoder-decoder cascade convolution neural network for crop disease leaf segmentation, IEEE Access 9 (2021), 14849–14866. doi:10.1109/ACCESS.2021.3052769.

41.

Zhong,

Huang and

Tang, Classification of cassava leaf disease based on a non-balanced dataset using transformer-embedded ResNet, Agriculture 12(9) (2022), 1360.

42.

Zhou,

Lin,

Zhang,

Shao,

Chen and

Liu, Improved itracker combined with bidirectional long short-term memory for 3D gaze estimation using appearance cues, Neurocomputing 21(390) (2020), 217–225. doi:10.1016/j.neucom.2019.04.099.

43.

Zhou,

M.M.

Rahman Siddiquee,

Tajbakhsh and

Liang, Unet++: A nested u-net architecture for medical image segmentation, in: InDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Proceedings, Vol. 4, Springer International Publishing, Granada, Spain, 2018, pp. 3–11.