Jaithoon Bibi Mohammed Saleem*
| Karpagavalli Shanmugam
© 2023 IIETA. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
In agricultural applications, the most essential task is to classify leaf diseases and their associated pests from various aspects. To achieve this, a Deep Convolutional Neural Network (DCNN) model was developed to classify the leaf diseases based on the soil and climatic features. But it needs a recommendation system to control the pesticide use for controlling the leaf diseases caused by specific pests. Hence, this paper hybridizes the Multi-dimensional Feature Learning-based DCNN (MFL-DCNN) with the Rough Set (RS) on an intuitionistic Fuzzy approximation space (RSF)-based decision support system to suggest the proper pesticides for a certain crop to be planted in a particular region. First, the leaf images are augmented by the Positional-aware Dual-Attention and Topology-Fusion with Evolutionary Generative Adversarial Network (PDATFEGAN) model. Then, the multi-dimensional data such as the created leaf images, pest, soil, weather, and pesticide data are fed to the DCNN with a softmax classifier for classifying leaf diseases and related pests. Then, the RSF-based decision model is applied, which determines the correlation between leaf disease and pests to recommend suitable pesticides. Finally, the experimental results reveal that the MFL-DCNN-RSF accomplishes a maximum efficiency than all other models for recommending pesticides to control leaf diseases and pests.
leaf diseases, PDATFEGAN, MFL-DCNN, pesticide, fuzzy rule, rough set, intuitionistic fuzzy approximation space, recommendation system
Crop productivity is endangered by many conditions, like environmental issues, crop diseases, and land erosion. The pathogenic illnesses of plants are worsened due to the growth of a wide range of natural commodities, and environmental degradation characteristics [1, 2]. Those illnesses are not appropriately recognized and diagnosed by human eyesight, which impacts yield productivity. To tackle this issue, Artificial Intelligence (AI) models including machine learning and deep learning algorithms have been adopted in crop/plant disease detection [3, 4]. The crop diseases are mostly identified by the leaves using a variety of methods. Many researchers have experienced the different machine learning algorithms for the detection and classification of various plant leaf diseases, including Support Vector Machine (SVM), Artificial Neural Network (ANN), random forest, and so on [5, 6]. But these algorithms need separate mechanisms for each process like pre-processing, feature extraction, feature selection, and classification. This leads to high computational time complexity.
So, deep learning algorithms have been developed for the detection and classification of crop leaf diseases from a huge number of images. Mostly used deep learning algorithms are pre-trained DCNNs [7-9], e.g., VGG, AlexNet, GoogleNet, etc. These algorithms achieved better feasibility and efficiency in identifying and classifying leaf diseases. Alternatively, images captured from farms were blurred. Poor image quality may degrade the accuracy of pre-trained classifiers, which were trained on clear high-resolution images. To increase the accuracy of leaf disease classification, low-resolution images should be regenerated into high-resolution images. For this purpose, a variety of Generative Adversarial Network (GAN) models has been employed [10], which generate more high-resolution images from the limited number of low-resolution images. Amongst, the GAN with the Dual-Attention and Topology-Fusion strategies called the DATFGAN model [11] outperformed classical GAN models in terms of sharpness and image details. It can generate sharper leaf disease images precisely by eliminating artifacts or noisy textures for increasing classification accuracy. The generated high-resolution leaf disease images were classified by the different pre-trained DCNN models to identify the types of diseases. In our previous works, the problems in the DATFGAN were solved: (a) the spatial correlation among the training images was learned with the position of disease region from the partial or whole leaf by the Positional-aware DATFGAN (PDATFGAN) model [12], and (b) the non-convergent iteration and adversarial learning ability were further improved by the Positional-aware Dual-Attention and Topology-Fusion with Evolutionary Generative Adversarial Network (PDATFEGAN) model which adopts an Evolutionary GAN (EGAN) [13]. The EGAN considers many adversarial objective values to reduce the different errors observed between the distribution of created and actual images. But the classification of leaf diseases was not only effective to enhance crop productivity.
Identification of causes for leaf diseases was also essential to control both pests and their related diseases efficiently. So, a few researchers focused on identifying pests from the leaf pest images [14, 15] of different plants using deep learning models. But, additional factors like soil and weather attributes were necessary to identify the pests that have more responsibility for a particular leaf disease. Also, there were no proper pesticide recommendation systems to advise farmers to use proper pesticides for controlling pests and leaf diseases.
Therefore, the MFL-DCNN-RSF model is proposed to control pests and leaf diseases by suggesting the proper pesticides. It enables a proper decision to predict the pesticide for a specific crop to be cultivated in a certain region having different soil and climate features. The RSF strategy is adopted to create the rule for recommending pesticides regarding multi-dimensional data such as crop, climate, soil, leaf diseases, and pests. Based on the created rules, the correlation between these characteristics is identified to suggest the proper pesticide to control leaf diseases and pests. Thus, this MFL-DCNN-RSF model can be helpful for cultivators to use proper pesticides for specific leaf diseases to improve crop productivity.
The remaining sections are prepared as follows: Section 2 discusses various recommendation systems used in agricultural activities. Section 3 describes the MFL-DCNN-RSF model and Section 4 demonstrates its competence. Section 5 summarizes the complete research and gives further enhancement.
Pinki et al. [16] designed an automated model using K-means clustering and the SVM to diagnose the different paddy leaf infections. Also, insect repellents were recommended based on the infection severity. But this model was not effective to classify multiple diseases simultaneously. Kosamkar et al. [17] developed a model, which performs preprocessing and feature extraction of leaf images followed by CNN to classify the diseases and recommend pesticides. But it needs other factors, which also affect the leaves.
Tetila et al. [18] analyzed various CNN frameworks for classifying the soybean pest images. However, it has a high learning time and less robustness. Rahman et al. [19] designed an enhanced 2-level training-based compact CNN model to recognize leaf infections and pests from the rice crop images. But it requires climate and soil features to improve the efficiency of automatically identifying pests. Ayan et al. [20] developed a genetic algorithm-based weighted ensemble of different pre-trained DCNN models using the sum of maximum probabilities mechanism to classify the crop pests properly. But its efficiency was less for large-scale datasets.
Wang et al. [21] developed a new DeepPest using an attention strategy for the primary categorization of insect photos into plant types. But, it has less efficiency when a small number of pest images were trained. Escola et al. [22] designed the SVM for the identification of cicadids in coffee trees. But it has a high maintenance cost and needs a recommendation system to suggest pesticides. Rao et al. [23] developed the AlexNet for automatically identifying grapes and mango leaf infections. But it needs to classify more classes of leaf infections and a recommendation system to suggest the proper solution to diagnose that infection.
Table 1. Comparison of different existing models
|
Ref. No. |
Algorithm used |
Problem solved |
Merits |
Demerits |
|
[16] |
K-means clustering and SVM |
Paddy leaf disease classification |
It can suggest pesticides based on the disease severity. |
It was not effective to classify multiple diseases simultaneously. |
|
[17] |
CNN |
Leaf disease classification and pesticide recommendation |
Better accuracy to suggest pesticides |
It needs climate and soil factors to increase accuracy. |
|
[18] |
CNN |
Soybean pest classification |
It can control pests in soybean fields. |
It has a high learning time and less robustness. |
|
[19] |
2-level training-based compact CNN |
Rice leaf disease and pest recognition |
Highest accuracy |
It requires climate and soil features to improve efficiency. |
|
[20] |
Genetic algorithm-based weighted ensemble DCNN |
Crop pest classification |
It can recognize pests earlier and suggest suitable pesticides. |
Its efficiency was less for large-scale datasets. |
|
[21] |
DeepPest using attention strategy |
Pest classification |
High robustness |
It has less efficiency when for a small-scale dataset. |
|
[22] |
SVM |
Cicadids recognition |
Better accuracy |
It has a high maintenance cost and needs a recommendation system to suggest pesticides |
|
[23] |
AlexNet |
Grapes and Mango leaf disease classification |
Low cost and high accuracy |
It needs to classify more classes of leaf infections and suggest the proper solution to diagnose that infection. |
|
[24] |
Tiny-YOLOv3 |
Pest classification |
It lowers pesticide costs and decreased environmental damage. |
More climatic features were needed to enhance efficiency. |
|
[25] |
OSSL |
Pest recognition |
It can prevent the contamination of a training dataset. |
It needs to control the pests by recommending pesticides. |
|
[26] |
Pre-trained CNNs and SVM |
Crop disease and pest classification |
Highest accuracy |
It takes more time due to more features. |
|
[27] |
VGG |
Multi-crop leaf disease classification |
Greater accuracy |
It needs to apply advanced CNN structures. |
Chen et al. [24] developed a Tiny-YOLOv3 model to recognize fruit tree pests. In contrast, several climatic features were needed to enhance efficiency. Rustia et al. [25] designed an Online Semi-Supervised Learning (OSSL) for an automated insect pest forecasting model. But it needs to control the pests by recommending pesticides. Turkoglu et al. [26] presented CNN and SVM classifiers to identify crop disease and pests. But it takes more time due to a large number of features. Paymode and Malode [27] employed the VGG model to identify the multi-crops leaf infection. But it needs to apply advanced CNN structures to achieve a deep analysis of leaf images.
Table 1 summarizes various models according to the algorithms used, problem solved, merits, and demerits. According to this summary, it is addressed that most of the models focused on pest and leaf disease classification, whereas a pesticide recommendation is essential to prevent pest and leaf diseases efficiently. So, this research develops a new recommendation system, which suggests pesticides to control environmental damages due to the use of excessive or improper pesticides.
The MFL-DCNN-RSF model for recommending suitable pesticides based on the classification of leaf diseases and pests is explained briefly. Figure 1 depicts the overall representation of the presented model for classifying the leaf diseases with their related pests and predicting the proper pesticide.
Figure 1. Schematic overview of MFL-DCNN-RSF for pesticide recommendation
3.1 Leaf images and pest data collection
To begin, the PlantVillage Dataset (PVD) [28] is used to collect leaf infection images for 3 different plants: tomato, bell pepper, and potato plants. There are 20636 images in this dataset, which are divided into 15 categories: pepper-bell bacterial spot, pepper-bell healthy, potato early blight, potato late blight, potato healthy, tomato bacterial spot, tomato early blight, tomato late blight, tomato leaf mold, tomato septoria leaf spot, tomato 2-spotted spider mites, tomato target spot, tomato yellow leaf curl virus, tomato mosaic virus and tomato healthy. This PVD includes images of several types of leaf infections that can impact tomato, pepper, and potato plants. Those images are in the RGB color space and saved in the uncompressed JPG format.
Once those images are obtained, the PDATFEGAN model is used to generate super-resolution micro patches for all leaf images, which are fused to provide complete super-resolution leaf images. In addition to the leaf images, a dataset is formed, which contains the agricultural pests related to the given 12 classes of leaf diseases. In this dataset, 10 insect pests are included with the related leaf diseases and weather situations in the region of the Coimbatore district in Tamilnadu. Maximum Temperature (Tmax), Minimum Temperature (Tmin), Relative Humidity (RH) in the morning and evening, Rainfall (RF), Wind Speed (WS), and Sun-Shine Hours (SSH) are among the time series of weather features evaluated in the occurrence of pests. Tomato mosaic virus, Tomato leaf curl virus, Xanthomonas campestris, Alternaria solani, Phytophthora infestans, Corynespora cassiicola, Xanthomonas gardneri, Alternaria tomatophila, Passalora fulva, Septoria lycopersici, and Tetranychidae are among the pests associated with the given 12 classes of leaf diseases and those pest data are addressed for the different climatic features.
3.2 Training of the MFL-DCNN classifier
Once all these data are obtained, the MFL-DCNN classifier is trained to classify the leaf diseases and their related pests. The learning of the MFL-DCNN classifier is shown in Figure 2, wherein the DCNN consists of 3 different pre-trained structures such as ShuffleNetV2, DenseNet121, and MobileNetV2.
Figure 2. Learning of MFL-DCNN classifier
After classification, it is needed to recommend the proper usage of pesticides to control leaf diseases efficiently. So, a pesticide dataset is created, which contains 12 classes of pesticides like copper hydroxide, resistant cultivators, etc. Similarly, soil features such as pH, moisture, and availability of nutrients (e.g., nitrogen (N), phosphorus (P), and potassium (K)) are collected as a soil property dataset. Both pesticide and soil datasets are collected around the Coimbatore district in Tamilnadu, from November 2021 to May 2022. Table 2 presents a few examples of pesticides used for leaf diseases with their related pests, soil, and weather factors.
Table 2. Examples of pesticides used for different leaf diseases, pests, soil and weather factors
|
Leaf Disease Name |
Pest Name |
Soil & weather factors |
Pesticide Name |
|
Pepper bell bacterial spot |
Xanthomonas campestris |
High temperature, high RH, low pH, low nutrients |
Cuprofix |
|
Potato Early blight |
Alternaria solani |
High WS, high RH, high temperature, low pH, low nutrients |
Maneb |
|
Potato late blight |
Phytophthora infestans |
High moisture, low pH, low nutrients |
Mancozeb |
|
Tomato target spot |
Corynespora cassiicola |
High moisture, high RH, low nutrients |
Azoxystrobin |
|
Tomato mosaic virus |
Tomato mosaic virus |
High temperature, low pH, low nutrients |
Sulfoxaflor |
|
Tomato yellow leaf curl virus |
Tomato leaf curl virus |
High temperature, low pH, low nutrients |
Pyrafluquinazon |
|
Tomato bacterial spot |
Xanthomonas gardneri |
High temperature, high RH, low pH, low nutrients |
BASF Cabriotop |
|
Tomato early blight |
Alternaria tomatophila |
High WS, high RH, high temperature, low pH, low nutrients |
Bonide Liquid Copper |
|
Tomato late blight |
Phytophthora infestans |
High mositure, low pH, low nutrients |
Clutch |
|
Tomato leaf mold |
Passalora fulva |
High temperature, high RH, low pH, low nitrogen |
Spray chlorothalonil |
|
Tomato septoria leaf spot |
Septoria lycopersici |
Medium temperature, high RH, high RF, low pH, low nutrients |
Copper soap |
|
Tomato two spotted spider mite |
Tetranychidae |
High temperature, low RF, low RH, low pH, low nutrients |
Bifenthri |
3.3 Pesticide decision support and recommendation system
To recommend suitable pesticides, the RSF-based decision support model is proposed. The primary tasks in this RSF system as illustrated in Figure 3 are fuzzification, RSF inference engine training and testing.
Figure 3. Major processes in RSF decision support system
First, a fuzzy set is defined, which gives various degrees of the membership function for its elements in the range of (0, 1). A fuzzy set is an extension of a crisp set, which allows only full membership or no membership, whereas fuzzy set allows partial membership. The list of multi-dimensional data created the fuzzy set as:
$X=\{$ temperature $, \ldots, N, P, K$, } (1)
Then, a membership function defines how each element in the input space is mapped to a membership value (or membership degree) between 0 and 1. The membership function maps all elements of X to a membership value between o and 1 as:
$\mu(x)=\{$ High, Moderate, Low $\}$ (2)
3.3.1 RSF System
Consider $U \neq \varphi$ is a quasi finite group of topics known as space and $x$ is a specific component of $U$. An Intuitionistic Fuzzy Set (ISF) $X$ of $U$ is described as $\left\{x, \mu_X(x), v_X(x)\right\}$, where $\mu_X: U \rightarrow[0,1]$ and $v_X: U \rightarrow[0,1]$ is the membership and non-membership degree, correspondingly, for all components $x \in U$ such that $0 \leq \mu_X(x)+v_X(x) \leq 1$. The range $\pi_X(x)=1-\left(\mu_X(x)+v_X(x)\right)$ is known as the indecision component that can provide either membership or non-membership or both ranges. Specifically, $\left(\mu_X(x), v_X(x)\right)$ is utilized to define the ISF $X$.
An intuitionistic fuzzy association (IA) on U is an ISF represented on (U×U) represented using the membership μIA and the non-membership vIA in Eq. (3).
$I A=\left\{\left(\mu_{I A}\left(x_i, x_j\right), v_{I A}\left(x_i, x_j\right)\right) \mid x_i, x_j \in U\right\}$ (3)
An $I A$ on $U$ is called an intuitionistic fuzzy neighborhood association when it ensures the below criteria, where $\mu_{I A}\left(x_i, x_j\right)$ is the membership degree and $v_{I A}\left(x_i, x_j\right)$ is the non-membership degree among 2 features $x_i$ and $x_j$.
1. $\mu_{I A}\left(x_i, x_j\right)=1$ and $v_{I A}\left(x_i, x_j\right)=0$ for each $x_i \in U$.
2. $\mu_{I A}\left(x_i, x_j\right)=\mu_{I A}\left(x_i, x_j\right)$ and $v_{I A}\left(x_i, x_j\right)=v_{I A}\left(x_i, x_j\right)$ for each $x_i, x_j \in U$.
Consider $J=\{(\alpha, \beta) \mid \alpha, \beta \in[0,1]\}$ and $0 \leq \alpha+\beta \leq 1$. After, for any $(\alpha, \beta) \in J,(\alpha, \beta)$-cut is provided as $I A_{\alpha, \beta}=$ $\left\{\left(x_i, x_j\right) \mid \mu_{I A}\left(x_i, x_j\right) \geq \alpha\right.$ and $\left.v_{I A}\left(x_i, x_j\right) \leq \beta\right\}$. It is observed that the 2 features $x_i$ and $x_j$ are $(\alpha, \beta)$-similar to $I A$, when $\left(x_i, x_j\right) \in I A_{(\alpha, \beta)}$ and $x_i I A_{(\alpha, \beta)} x_j$ is defined. Two features $x_i$ and $x_j$ are called $(\alpha, \beta)$-matching to $I A$, when a series of components $u_1, u_2, \ldots, u_n$ presents in $U$ such that $x_i I A_{(\alpha, \beta)} u_1, u_1 I A_{(\alpha, \beta)} u_2, \ldots, u_n I A_{(\alpha, \beta)} x_j$. In this scenario, it is observed that $x_i$ is a uniformity association $I A_{(\alpha, \beta)}$. The $I A_{(\alpha, \beta)}$-uniformity class of a component $x$ in $U$ is defined as $[x]_{(\alpha, \beta)}$. The pair $K=(U, I A(\alpha, \beta))$ is known as IF approximation space.
Consider $X \subseteq U$. Then, $(\alpha, \beta)$-minimum and $(\alpha, \beta)$ maximum approximation of $X$ in the generalized approximation space $K=(U, I A(\alpha, \beta))$ is defined as $\left(X_{M i n}^{\alpha, \beta}, X_{\text {Max }}^{\alpha, \beta}\right)$ in Eqns. (4) & (5):
$X_{M i n}^{\alpha, \beta}=\cup\left\{Y \mid Y \in I A_{\alpha, \beta}^*\right.$ and $\left.Y \subseteq X\right\}$ (4)
$X_{\operatorname{Max}}^{\alpha, \beta}=\cup\left\{Y \mid Y \in I A_{\alpha, \beta}^*\right.$ and $\left.Y \cap X \neq \varphi\right\}$ (5)
A given set $X$ is known as $(\alpha, \beta)$-rough when $X_{\text {Max }}^{\alpha, \beta} \neq X_{\text {Min }}^{\alpha, \beta}$. Similarly, a given set $X$ is known as $(\alpha, \beta)$-crisp when $X_{\text {Max }}^{\alpha, \beta}=$ $X_{\text {Min }}^{\alpha, \beta}$. Evenly, a set $X$ is known as $(\alpha, \beta)$-rough when the limit $L_{I I M}^{\alpha, \beta}=X_{\text {Max }}^{\alpha, \beta}-X_{\text {Min }}^{\alpha, \beta}$ such that $L I M_{I A}^{\alpha, \beta} \neq \varphi$.
3.3.2 Fuzzification
In this task, a crisp input is converted into a lingusitic variable using the membership function provided by the fuzzy knowledge base. The triangular membership function is used because the linguistic variables are modeled into 3 sets i.e., all the elements in X are converted into Low (L), Medium (M), and High (H). This process needs to get input provided by the user. The compositional ranges of multi-dimensional data are great determinants for predicting suitable pesticide. Such multi-dimensional data are grouped into 3 linguistic variables in the range of 0 to 1 (as illustrated in Table 3).
Table 3. RSF-based decision input variables (for weather and soil attributess)
|
Attributes |
Low (L) |
Moderate (M) |
High (H) |
|
Temperature |
0–35 |
36–65 |
66–100 |
|
RH |
0–29 |
30–59 |
60–80 |
|
RF |
0–40 |
41–70 |
70–100 |
|
WS |
15–30 |
30–69 |
70–100 |
|
SSH |
0.2–0.45 |
0.46–0.7 |
0.71–1.0 |
|
Soil moisture |
10–25 |
25–40 |
40–60 |
|
pH |
0.1–0.3 |
0.3–0.6 |
0.6–1.0 |
|
Nitrogen (N) |
1.0–1.99 |
2.0–3.99 |
4.0–6.0 |
|
Phosphorous (P) |
0.15–0.35 |
0.36–0.59 |
0.60–0.80 |
|
Pottasium (K) |
0.1–2.49 |
2.50-4.49 |
4.50–8.50 |
3.3.3 Membership and non-membership calculation
The intuitionistic fuzzy tolerance finds the maximum indiscernibility of all attributes. The RSF generates (α, β) uniformity classes, where α refers to the membership degree and β refers to the non-membership degree, correspondingly. The membership degree should be large and the non-membership degree should be small to obtain a better prediction of proper pesticides.
Because each prediction can include accurately a particular pesticide, the model fails to analyze the data if belongingness is set to 1 and non-belongingness is set to 0. It is due to the feature values being non-qualitative. The membership and non-membership associations are calculated such that the total ranges obtain from 0 to 1 and such factors should be symmetric.
Here, pesticide prediction is performed using the different features related to leaf diseases, pests, soil, and climatic conditions. By changing the values of α and β, these features may diverge from each other. The range of α is decremented and the range of β is incremented, gradually the number of features will become more essential. The membership degree (μ) and the non-membership degree (v) amid xi and xj are described in Eqns. (6) & (8), respectively:
$\mu_A\left(x_i, x_j\right)=1-\frac{\left|V_{a_i}^{x_i}-V_{a_i}^{x_j}\right|}{\max \left(V_{a_i}\right)}$ (6)
$v_A\left(x_i, x_j\right)=1-\frac{\left|V_{a_i}^{x_i}-V_{a_i}^{x_j}\right|}{2 \times \max \left(V_{a_i}\right)}$ (7)
In Eqns. (6) & (7), $V_{a_i}^{x_i}$ denotes the value of $x_i$ for a particular crop $a_i$.
Table 4. Sample rules for pesticide recommendation
|
Temperature |
RH |
RF |
WS |
SSH |
Soil moisture |
pH |
N |
P |
K |
Leaf diseases |
Pests |
Pesticide |
|
H |
H |
L |
M |
H |
L |
L |
L |
L |
L |
Pepper bell bacterial spot |
Xanthomonas campestris |
Cuprofix |
|
H |
H |
L |
H |
M |
H |
L |
L |
L |
L |
Potato Early blight |
Alternaria solani |
Maneb |
|
L |
L |
L |
L |
L |
H |
L |
L |
L |
L |
Potato late blight |
Phytophthora infestans |
Mancozeb |
|
M |
H |
L |
M |
M |
H |
L |
L |
L |
L |
Tomato target spot |
Corynespora cassiicola |
Azoxystrobin |
|
H |
M |
L |
L |
M |
M |
L |
L |
L |
L |
Tomato mosaic virus |
Tomato mosaic virus |
Sulfoxaflor |
|
H |
M |
L |
M |
H |
L |
L |
L |
L |
L |
Tomato yellow leaf curl virus |
Tomato leaf curl virus |
Pyrafluquinazon |
|
H |
H |
L |
L |
L |
L |
L |
L |
L |
L |
Tomato bacterial spot |
Xanthomonas gardneri |
BASF Cabriotop |
|
H |
H |
L |
H |
H |
L |
L |
L |
L |
L |
Tomato early blight |
Alternaria tomatophila |
Bonide Liquid Copper |
|
L |
L |
M |
L |
L |
H |
L |
L |
L |
L |
Tomato late blight |
Phytophthora infestans |
Clutch |
|
H |
H |
L |
H |
H |
L |
L |
L |
L |
L |
Tomato leaf mold |
Passalora fulva |
Spray chlorothalonil |
|
M |
H |
H |
L |
L |
M |
L |
L |
L |
L |
Tomato septoria leaf spot |
Septoria lycopersici |
Copper soap |
|
H |
L |
L |
L |
H |
L |
L |
L |
L |
L |
Tomato two spotted spider mite |
Tetranychidae |
Bifenthrin |
*Note: H – High; M – Medium; L – Low
3.3.4 Inference engine and knowledge base
The knowledge base is a component, where knowledge is developed, accumulated, arranged, analyzed, and distributed. It comprises a dataset and a rule base. The dataset gives the requied elements to define the linguistic variables and rules using IF-THEN control constructs. The dataset involves a group of facts utilized to match against the IF (condition) parts of rules accumulated in the knowledge base.
The rule knowledge base follows the Mamdani rule creation. Table 4 shows a few rules derived from the use of Mamdani rule. For an example of pepper bell bacterial spot disease, if the temperature is H && RH is H && RF is L && WS is M && SSH is H && soil moisture is L && pH is L && nitrogen is L && phosphorus is L && potassium is L, then this leaf disease is caused by the pest called Xanthomonas campestris. So, to prevent this pest, the Cuprofix pesticide is recommended.
The performance of the MFL-DCNN-RSF model is assessed by implementing it in Python 3.7.8 using the PVD. Figure 4 shows the application process of MFL-DCNN-RSF. To validate the efficiency of this model, 2250 leaf images with 15 distinct classes are acquired. For the training process, 1500 leaf images (100 images from each class) are collected randomly. For the testing, 750 leaf images (50 images from each class) are collected. Also, the efficiency is analyzed in terms of different metrics. Figure 5 portrays a few samples of the diseased leaf images from each given class.
Figure 4. Application process of MFL-DCNN-RSF
Precision is determined by Eq. (8).
Precision $=\frac{\text { No.of exactly classified diseased leaves and pests }}{\text { No. of exactly classified diseased leaves and pests }+ \text { No.of inexactly classified diseased leaves and pests }}$ (8)
Recall is determined by Eq. (9).
Recall $=\frac{\text { No.of exactly classified diseased leaves and pests }}{\text { No.of exactly classified diseased leaves and pests }+ \text { No.of inexactly classified healthy leaves and pests }}$ (9)
F-measure is calculated as Eq. (10).
$F-$ measure $=2 \times \frac{\text { Precision } \cdot \text { Recall }}{\text { Precision }+ \text { Recall }}$ (10)
Accuracy is determined by Eq. (11).
Accuracy $=\frac{T P+\text { True Negative }(T N)}{T P+T N+F P+F N}$ （11）
The findings of MFL-DCNN-RSF, Tiny-YOLOv3 [24], OSSL [25], and SVM [26] applied on the multi-dimensional dataset (super-resolved leaf disease images, pests associated with the weather, soil factors, and pesticides) are given in Table 5 and the graphical depiction of accuracy ranges is presented in Figure 6.
Figure 6 addresses that the accuracy of the MFL-DCNN-RSF is 10.81% higher than the SVM, 8.07% higher than the OSSL, and 6.19% higher than the tiny-YOLOv3. The precision of the MFL-DCNN-RSF is 10.33% higher than the SVM, 7.41% higher than the OSSL, and 6.01% higher than the tiny-YOLOv3. The recall of the MFL-DCNN-RSF is 12.04% higher than the SVM, 10.17% higher than the OSSL, and 7.41% higher than the tiny-YOLOv3. The comparative scrutiny clarifies that the MFL-DCNN-RSF model guarantees the highest accuracy in classifying the leaf diseases and pests, as well as predicting the suitable pesticides compared to the other models applied to the multi-dimensional dataset. The proposed MFL-DCNN-RSF uses a variety of attributes for pesticide recommendation by classifying leaf diseases and pests, whereas the existing models focus only on the low-resolution leaf and pest images, which may not sufficient to provide maximum accuracy. Also, existing models take more time for learning and have high complexity while increasing the number of data. So, according to the observed accuracies, the MFL-DCNN-RSF model is highly helpful to classify leaf diseases, and pests and recommend suitable pesticides with proper usage for controlling the leaf diseases.
Table 5. Analysis of MFL-DCNN-RSF and existing models for pesticide recommendation
|
Models |
Precision (%) |
Recall (%) |
F-measure (%) |
Accuracy (%) |
|
SVM |
89.47 |
87.19 |
88.33 |
89.28 |
|
OSSL |
91.90 |
88.67 |
90.29 |
91.54 |
|
Tiny-YOLOv3 |
93.11 |
90.95 |
92.03 |
93.16 |
|
MFL-DCNN-RSF |
98.71 |
97.69 |
98.20 |
98.93 |
Figure 5. Few sample images of different classes of diseased leaf images
Figure 6. Comparison of proposed and existing pesticide recommendation models
In this article, the MFL-DCNN-RSF model was developed for preventing leaf diseases and pests by recommending suitable pesticides. This model can be helpful for cultivators to use appropriate pesticides for corresponding pests and leaf diseases. As a result, it reduces environmental damages due to the excessive/improper usage of pesticides and enhances crop productivity significantly. The test outcomes proved that the MFL-DCNN-RSF model has a mean accuracy of 98.93%, respectively, contrasted with the other classification models.
[1] Rehman, H.U., Anwar, S., Tufail, M. (2020). Machine vision based plant disease classification through leaf imagining. Ingénierie des Systèmes d’Information, 25(4): 437-444. https://doi.org/10.18280/isi.250405
[2] Ristaino, J.B., Anderson, P.K., Bebber, D.P., Brauman, K.A., Cunniffe, N.J., Fedoroff, N.V., Wei, Q. (2021). The persistent threat of emerging plant disease pandemics to global food security. Proceedings of the National Academy of Sciences, 118(23): 1-9. https://doi.org/10.1073/pnas.2022239118
[3] Lamba, M., Gigras, Y., Dhull, A. (2021). Classification of plant diseases using machine and deep learning. Open Computer Science, 11(1): 491-508. https://doi.org/10.1515/comp-2020-0122
[4] Noola, D.A., Basavaraju, D.R. (2021). Corn leaf disease detection with pertinent feature selection model using machine learning technique with efficient spot tagging model. Revue d'Intelligence Artificielle, 35(6): 477-482. https://doi.org/10.18280/ria.350605
[5] Zamani, A.S., Anand, L., Rane, K.P., Prabhu, P., Buttar, A.M., Pallathadka, H., Dugbakie, B.N. (2022). Performance of machine learning and image processing in plant leaf disease detection. Journal of Food Quality, 2022: 1-7. https://doi.org/10.1155/2022/1598796
[6] Hussein, M.A., Abbas, A.H. (2019). Plant leaf disease detection using support vector machine. Al-Mustansiriyah Journal of Science, 30(1): 105-110. https://doi.org/10.23851/mjs.v30i1.487
[7] Shah, N., Jain, S. (2019). Detection of disease in cotton leaf using artificial neural network. In: IEEE Amity International Conference on Artificial Intelligence, pp. 473-476. https://doi.org/10.1109/AICAI.2019.8701311
[8] Ferentinos, K.P. (2018). Deep learning models for plant disease detection and diagnosis. Computers and Electronics in Agriculture, 145: 311-318. https://doi.org/10.1016/j.compag.2018.01.009
[9] Das, P., Yadav, J.K.P.S., Yadav, A.K. (2021). An automated tomato maturity grading system using transfer learning based AlexNet. Ingénierie des Systèmes d’Information, 26(2): 191-200. https://doi.org/10.18280/isi.260206
[10] Sun, R., Zhang, M., Yang, K., Liu, J. (2020). Data enhancement for plant disease classification using generated lesions. Applied Sciences, 10(2): 1-16. https://doi.org/10.3390/app10020466
[11] Dai, Q., Cheng, X., Qiao, Y., Zhang, Y. (2020). Crop leaf disease image super-resolution and identification with dual attention and topology fusion generative adversarial network. IEEE Access, 8: 55724-55735. https://doi.org/10.1109/ACCESS.2020.2982055
[12] Bibi, M.J., Karpagavalli, S. (2021). Positional-aware dual attention and topology fusion GAN for plant leaf disease image super-resolution and classification. Linguistica Antverpiensia, 12(3): 13-25.
[13] Bibi, M.J., Karpagavalli, S. (2021). An evolutionary optimization of positional-aware dual-attention and topology fusion generative adversarial network for plant leaf disease detection. Turkish Online Journal of Qualitative Inquiry, 12(3): 2904-2922.
[14] Ngugi, L.C., Abelwahab, M., Abo-Zahhad, M. (2021). Recent advances in image processing techniques for automated leaf pest and disease recognition–a review. Information Processing in Agriculture, 8(1): 27-51. https://doi.org/10.1016/j.inpa.2020.04.004
[15] Liu, J., Wang, X. (2021). Plant diseases and pests detection based on deep learning: A review. Plant Methods, 17: 1-18. https://doi.org/10.1186/s13007-021-00722-9.
[16] Pinki, F.T., Khatun, N., Islam, S.M. (2017). Content based paddy leaf disease recognition and remedy prediction using support vector machine. In: IEEE 20th International Conference of Computer and Information Technology, pp. 1-5. https://doi.10.1109/iccitechn.2017.8281764
[17] Kosamkar, P.K., Kulkarni, V.Y., Mantri, K., Rudrawar, S., Salmpuria, S., Gadekar, N. (2018). Leaf disease detection and recommendation of pesticides using convolution neural network. In: IEEE Fourth International Conference on Computing Communication Control and Automation, pp. 1-4. https://doi:10.1109/iccubea.2018.8697504
[18] Tetila, E.C., Machado, B.B., Astolfi, G., de Souza Belete, N.A., Amorim, W.P., Roel, A.R., Pistori, H. (2020). Detection and classification of soybean pests using deep learning with UAV images. Computers and Electronics in Agriculture, 179: 1-11. https://doi.org/10.1016/j.compag.2020.105836
[19] Rahman, C.R., Arko, P.S., Ali, M.E., Khan, M.A.I., Apon, S.H., Nowrin, F., Wasif, A. (2020). Identification and recognition of rice diseases and pests using convolutional neural networks. Biosystems Engineering, 194: 112-120. https://doi.org/10.1016/j.biosystemseng.2020.03.020
[20] Ayan, E., Erbay, H., Varçın, F. (2020). Crop pest classification with a genetic algorithm-based weighted ensemble of deep convolutional neural networks. Computers and Electronics in Agriculture, 179: 1-10. https://doi.org/10.1016/j.compag.2020.105809
[21] Wang, F., Wang, R., Xie, C., Yang, P., Liu, L. (2020). Fusing multi-scale context-aware information representation for automatic in-field pest detection and recognition. Computers and Electronics in Agriculture, 169: 1-11. https://doi.org/10.1016/j.compag.2020.105222
[22] Escola, J.P.L., Guido, R.C., da Silva, I.N., Cardoso, A.M., Maccagnan, D.H.B., Dezotti, A.K. (2020). Automated acoustic detection of a cicadid pest in coffee plantations. Computers and Electronics in Agriculture. 169: 1-8. https://doi.org/10.1016/j.compag.2020.105215
[23] Rao, U.S., Swathi, R., Sanjana, V., Arpitha, L., Chandrasekhar, K., Naik, P.K. (2021). Deep learning precision farming: Grapes and mango leaf disease detection by transfer learning. Global Transitions proceedings, 2(2): 535-544. https://doi.org/10.1016/j.gltp.2021.08.002
[24] Chen, C.J., Huang, Y.Y., Li, Y.S., Chen, Y.C., Chang, C.Y., Huang, Y.M. (2021). Identification of fruit tree pests with deep learning on embedded drone to achieve accurate pesticide spraying. IEEE Access, 9: 21986-21997. https://doi.org/10.1109/ACCESS.2021.3056082
[25] Rustia, D.J.A., Lu, C.Y., Chao, J.J., Wu, Y.F., Chung, J.Y., Hsu, J.C., Lin, T.T. (2021). Online semi-supervised learning applied to an automated insect pest monitoring system. Biosystems Engineering, 208: 28-44. https://doi.org/10.1016/j.biosystemseng.2021.05.006
[26] Turkoglu, M., Yanikoğlu, B., Hanbay, D. (2021). PlantDiseaseNet: convolutional neural network ensemble for plant disease and pest detection. Signal, Image and Video Processing, 16(9): 1-9. https://doi:10.1007/s11760-021-01909-2
[27] Paymode, A.S., Malode, V.B. (2022). Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG. Artificial Intelligence in Agriculture, 6: 23-33. https://doi.org/10.1016/j.aiia.2021.12.002
[28] Emmanuel, T.O. (2019). PlantVillage Dataset. https://www.kaggle.com/emmarex/plantdisease.
