A novel approach for Plant disease Classification through Neural Network-Based Color Feature Analysis

doi:10.21203/rs.3.rs-4210118/v1

plant disease identification using machine vision, which is a challenge in terms of maximizing both the quality and quantity of plant growth. The infection makes plants susceptible to disease. This needs continuous monitoring by experts, which is prohibitively expensive in large farms, but in some instances, erroneous observations by farmers culminate in poor diagnoses. Consequently, we need a fast and accurate plant disease diagnosis predicted to increase the area under cultivation, eliminate heavy losses, and ensure high accuracy. The focus must be on identifying early symptoms of plant disease using computer vision. In order to solve this problem, deep learning can combine machine learning and pattern recognition, two hottest topics in this field. We propose a novel method to identify different plant diseases using deep convolutional neural networks (CNNs). In this study, we propose an image-based classification approach for rice plant diseases, focusing solely on color features. We investigated 12 distinct color spaces and derived 4 features from each color channel, resulting in a total of 48 features. The accuracy of this model is much higher than that of traditional machine learning models. Using the best-performing model, we achieved a classification accuracy of 96.03%. The simulation results show that the proposed method for identifying plant diseases is effective and feasible.

Plant disease

color analysis

Machine learning

convolutional neural networks and classification

Agricultural production is, of necessity, to be maximized because the economy greatly depends on it. Agriculture is a primary occupation, with rice and wheat taking pride in place as staple foods (N. Zhang, Wang, and Wang 2002) (Jones et al. 2017). At the same time, staples such as wheat and rice are vital to humans, ensuring a healthy diet and robust immune system to fight diseases and infections. So then, there is a need to pay attention to productivity in agriculture (AGRICULTURAL RESEARCH COUNCIL Annual Report for 2015/16 October 2016 2016). But Plants show stress as a result of environmental changes, pests, and disease, developing visible symptoms in terms of color, shape, and texture changes. So the plants require proper diagnosis in time so that the crops can be safeguarded from loss (Ramegowda and Senthil-Kumar 2015) (Darley and Middleton 1966). Profits became a viable factor after some time, generating larger harvests (Zhu et al. 2016).

(Singh and Kumari 2023) investigate the influence of agricultural technologies on both crop yield and farmer income within the Indian context. The research delves into various technological interventions implemented in agriculture and evaluates their effectiveness in enhancing both productivity and economic returns for farmers. By examining the adoption and outcomes of these technologies, the study provides valuable insights into the role of innovation in driving agricultural development and livelihood improvement in India, thereby contributing to the discourse on sustainable farming practices and rural prosperity.

Consequently, the use of deadly poisons in the form of herbicides has increased drastically. There is no doubt that productivity has increased, but environmental damage has gone unreported, ultimately resulting in long-term and severe health consequences. (Kissoudis et al. 2014) It was observed that there were several diseases that resulted in a variety of visible infections. Three types of diseases commonly affect crops: fungi, bacteria, and viruses. (Thakur and Sohal 2013) Identifying symptoms and signs and knowing how to control the disease are ongoing challenges, as illustrated in Table 1. In this way, technology strongly supports agriculture through the development of technologies (J. C. Zhang et al. 2012) (Rastogi, Arora, and Sharma 2015).

In addition to impacting food security, plant diseases affect small farmers (Harvey et al. 2014). According to earlier research, plant infection is purely determined by the common characteristics of the disease. A fungus or virus may have attacked, bad weather conditions may have caused it, or an incorrect observation may have caused it. In addition to affecting smallholder farmers, plant diseases impact food security.

(Dash, Kumar, and Kumari 2023) This paper proposes a novel approach for maize disease identification using an optimized support vector machine (SVM) model that leverages deep features extracted from DenseNet201, a convolutional neural network architecture. Authored by Arabinda Dash, Prabira Kumar Sethy, and Santi Kumari Behera, the study aims to improve the accuracy and efficiency of maize disease identification, crucial for effective agricultural management. By integrating the powerful feature representation capabilities of DenseNet201 with the robust classification capabilities of SVM, the proposed method offers a promising solution for accurate and timely detection of maize diseases, potentially aiding farmers in implementing targeted intervention strategies to mitigate crop losses. (Nan et al. 2023) It was observed that a novel method for maize leaf disease classification by utilizing RGB-D post-segmentation image data. Authored by Fei Nan, the study introduces an innovative approach that combines RGB color information with depth data obtained from post-segmentation images, offering a comprehensive representation of maize leaf diseases. By leveraging this multimodal data fusion technique, the proposed method aims to enhance the accuracy and robustness of disease classification, crucial for effective disease management in maize crops. This approach could potentially provide farmers with more reliable tools for early detection and intervention, ultimately contributing to improved agricultural productivity and sustainability.

Table 1. Common diseases of Crops
Reference	Culture	Types of diseases and their pathogens	Favorable conditions for pathogens	Disease cycles	Key diagnostic signs	The extent of yield loss
(Phadikar, Sil, and Das 2013)	Rice Plant	Bacterial Leaf Blight (bacteria Xanthomonas oryzae)	Strong winds, heavy rains, and tropical and temperate environments	During maturity, affects all plant parts (leaf, neck, and node).	Stripes of yellowish water soaked into wavy margins. Leaves with severe infestations dry quickly,	Serious diseases yield a loss of up to 70%.
(Zhou et al. 2014)	Sugar Beet	Cercospora Leaf Spot (Cercosporabeticola)	Rainy and humid weather	It starts from a younger plant to an old plant.	Circular spots about 1/8 inch with ash gray center.	The infection starts from root tonnage
(S. Shrivastava, Singh, and Hooda 2015)	Soybeans	Downy mildew (Peronospora manshurica}	Humid, wet, warm weather. Infested soybean residue on the soil surface.	During the flowering stage.	Visible on leaves, Lesions are mostly-round spots, brown to gray, surrounded by a thin dark reddish ring.	During heavy loss, the leaves were prematurely dropped from plants.

According to researchers, image-based assessment approaches are more accurate and reproducible than human visual assessments. According to (Xu et al. 2011), tomato plants can be detected as having potassium and nitrogen deficiencies. The author defines a method for finding features in a colour image using an algorithm. In this study, genetic algorithms were used to identify unique features to diagnose the disease. In addition, fuzzy k-nearest neighbour classifier was used for the final classification. With the help of L*a*b* color space, the features are extracted (Bai et al. 2013). In 2011, (Macedo-Cruz et al. 2011) proposed a segmentation method to identify region-based color transformation in oat crops. A cropped image was captured in the direct field so that color transformation could be analyzed to predict diseased plants (Sladojevic et al. 2016). In addition, the resulting partition can be used to predict the infected area caused by the different crops. Lastly, frost damage to oat plants is predicted by automatic threshold techniques. The author defines a method [45] to the realm of plant disease detection, employing computational intelligence and image processing techniques. It likely presents a comprehensive study that explores the application of advanced algorithms and methodologies to automatically identify and classify plant diseases based on visual symptoms captured through image analysis. The authors likely investigate various computational intelligence approaches, such as machine learning and neural networks, to analyze plant images and extract relevant features indicative of disease presence. Through rigorous experimentation and validation, the paper likely provides insights into the efficacy and reliability of these techniques in accurately diagnosing plant diseases, offering valuable contributions to the field of precision agriculture and crop management. A variety of authors extract features using different methods from feature extraction, which contain information about shape, color, and texture, as summarized in Table 2.

Table 2

A summary of different features of a Plant disease
Reference	Different Plants species	Features
(Camargo and Smith 2009a)a	Soya, banana leaf, cotton, maize, alfalfa	Shape and color
(Camargo and Smith 2009b)	Cotton plant	Shape and color
(Phadikar and Sil 2008)	Paddy leaves	Shape and color
(V. K. Shrivastava and Pradhan 2021)	Soybean	Shape and color
(Bai et al. 2013)	Paddy leaves	Shape and color

Agriculture continues to pose challenges that need to be addressed. Certain specific challenges to be dealt with include the following:

Even though the researchers have used several databases, there is no standard database that can be utilized by researchers, which is becoming a serious issue as no Universal database is available. This can be resolved using a holistic approach or rules to build and maintain a standard dataset.
The diseases are identified by various methods such as naked eye observation, microscopic, and camera-based. Each method has its strengths and limitations. A system fusing all the best features from different methods would give better accuracy.
In addition, image processing becomes complicated when the lighting and angle of the objects need to be more appropriate (Thornton et al. 2009). Expert photographers may be unable to avoid image disturbances despite their best efforts. Therefore, this poses a challenge for the researchers. This can be resolved by adapting automatic techniques to capture images regularly when the environmental conditions are favorable.

The proposed model is more accurate according to the challenges discussed above and techniques using deep learning and machine learning (Dhaka et al. 2021). With all these things in mind, we propose a deep learning algorithm for detecting leaf diseases automatically in this paper. Three stages are involved in contributing to the framework mentioned above. To maximize the quality of the leaf samples, image pre-processing (Archana and Sahayadhas 2018b) and K-means clustering are employed first. By analyzing the K-means clustering response (Archana and Sahayadhas 2018a), it is possible to predict whether the leaf is diseased or not at an early stage of operation. Secondly, novel intensity-based color features (NIBCF) were used to extract the informative regions and features of the samples. Finally, CNNs are used to classify features in deep learning approaches (K. Zhang et al. 2018).

Image Processing has widespread applications in medicine, agriculture, remote sensing, pattern recognition, video processing, microscopic imaging, etc. Machine vision is used to analyze data in plant pathology to assist farmers technologically. This diagram reports automatic plant disease identification through various methods, as shown in Fig. 1: (1) Acquisition: input image taken through a digital camera or plant village, (2) Image pre-processing: noise removal, (3) Image segmentation: separate the pigment from the background image, (4) Feature extraction: analyze the unique features (5) Classification: separate diseased leaf and healthy leaf.

4.1 Dataset

A great deal of information is required to train intelligent visualization and classification systems. As a rule, machine learning and deep learning systems perform better when trained with a large amount of data. For this research work, we utilized images taken from Thanjavur district, India, as well as data from the Plant Village database. Here, plant datasets are divided into three groups based on their types, such as corn, tomato, and potato. The goal of this study was to generate a unique data set that contained images of different sizes (S. Shrivastava, Singh, and Hooda 2015). We used whole data set to process, extract, select, and classify plant leaf images. Deep learning systems should be trained and evaluated using research and assessment data. According to the data set, Fig. 2,3,4, and Table 3 represents the leaf diseases related to corn, tomato, and rice plants, respectively.

Table 3 Summary of an image dataset

Dataset	Trained image	Tested image	Total image	Classes
Corn	1282	700	1982	2
Tomato	1100	600	1700	2
Potato	700	450	1150	2

4.2 Image Segmentation of Plant disease

Segregation is the process of splitting images from unsupervised algorithms in computer vision. The goal is to distinguish the diseased from the targeted part (Wang, Sun, and Wang 2017). A distinction was made between normal and diseased regions to obtain the target region (Barbedo 2016). Image segmentation can be done in many ways, including region-based edge-based, and cluster-based. There are several types of clustering, but clustering is the most important.The infected portions of the leaf are segmented using different techniques to extract the pigment(Wijekoon, Goodwin, and Hsiang 2008). A segmentation method (Camargo and Smith 2009b), however, proposed a method to separate the pigment from the earliest symptoms from the color images and to identify the pigment from the earliest symptoms in color images by implementing a segmentation method and utilizing the features as inputs for classifying the images in the segmentation method (Munisami et al. 2015). Researchers (Garcia and Barbedo 2016) developed an to segment plant leaf disease symptoms.

Based on the extracted images, the segmentation method classifies the disease type based on the pigments extracted from infected regions (Espinoza et al. 2016).(Camargo and Smith 2009a) considered the pigment in color images and differentiated the features based on segmentation. The author introduced a novel algorithm for segmenting disease symptoms on leaves. As far as disease identification is concerned, automatic segmentation was more accurate than manual segmentation. According to the author, automatic segmentation produces much higher levels of accuracy than manual segmentation. Using the K-means cluster algorithm, pigment in the leaves was separated from the background in this research (Fig. 5 and Algorithm I) (Archana and Sahayadhas 2018a).

Algorithm I: Improved K- means segmentation

Input:Median Filter Output Image (M_Filt)

Output: Segmented Image.

STEP 1: Read the input image from the preprocessed image.

STEP 2: In the next step, use rgb2lab to convert the input image into RGB format.

STEP 3: Again, convert the RGB image using the model of L*a*b* conversion he redundant function of using apply form

STEP 4: Next, resized the previous image with row-wise and column specified values were identified.

STEP 5: Calculate the k points on the centroids of the initial group, which represent the objects to be clustered.

STEP 6: The closest point to the centroid is determined by assigning the pixel value to each point and calculating the distance.

STEP 7: After finding the region of interest using different clusters and again convert the image using the RGB conversion model as Final Segmented Image.

4.3 Feature extraction: RGB Color Model

A color model can be used to differentiate the symptoms of a damaged pathogen. It is necessary to represent several models to analyze the color image. A model may be represented using RGB, which is converted to HSV. An RGB image has been created by combining three channels, i.e., red, green, and blue, using an additive color model in various ways to create a vast array of colors(Shih and Cheng 2005).An HSV space model is applied in this study to distinguish pathogens from healthy regions on the rice plant.As shown in Fig. 6, the RGB color model was transferred to the HSI color model using the equations below.

$$Hue\left(H\right)=2-ACOS \left\{\frac{\left[\left(R-G\right)+\left(R+G\right)\right]}{\sqrt[2]{\left(R-G\right)2+\left(R-G\right)\left(G-B\right)}}\right\}, B>G \left(1\right)$$

Where,

In this case, H refers to hue, which describes the pure color in the image. After S, we have Saturation, which refers to the dilution of pure color over white color. Finally, V stands for Value, which describes the color's brightness.

$$\text{I}\text{n}\text{t}\text{e}\text{n}\text{s}\text{i}\text{t}\text{y} \left( \text{I}\right)=\frac{R+G+B}{3} \left(2\right)$$

Where,

Red, green and blue represent an R, G, and B color model respectively. The greener pixel represents healthier and the arbitrarily minute value of ε which helps to eliminate division by zero.

Those pixels with the greenest color represent the healthiest portions. Here, the HSV color space model is used to analyze the RGB components, as shown in Fig. 7. This color space model is used to extract the specific infected portion of the leaf. Additionally, the color calibration reduces the effects of illumination variations caused by the unpredictable sun after collecting the color image data of the plant leaves. As illumination changes, R, G, and B color components' intensity values also change, resulting in modified intensity values, which will affect results (Algorithm II).

Algorithm II: NIBCF -Novel Intensity Based Color Feature

Input: Final_ Seg_ Image

Output: Color_ Feature

Step1: RGB to Gray conversation for the Future process.

Gray Image – rgb2gray (Final_ Seg_ Image);

Step2: After the Gray conversion we take the color features for the gray image by using the NIBCF.

Output Image = NIBCF(Gray_Image,6,2,1)

Where is Novel Intensity Based Color Feature.

Step3: To calculate the NIBCF Values based on the following steps.

Input_ Img = round (Input_ Img)

[r,c] = size(Input _ Img)

Mini_intelsity = min (min ( Input_Img)

Maxi_intelsity–max(Input _ Img)

Out_I = zeros (Maxi_intelsity – Mnini_intelsity + 1)

Then find Dir xandDir_yvalues

Dir_c = Dir_c * Dis

Dir_y = Dir _y * Dis;

Then find Intensity 1and Intensity 2 values

Based on the Intensity value we calculate the Output_ image.

Notes: Intensity1and Intensity2 values depend on the Minimum Intensity and Maximum_ Intensity.

Step4: After the, Output_ Image calculation we estimate the Color_Features for the input image.

M = Output _Image

for k1 – Output _Image

Idxr = 1 + fix(k1/4)

for k2 = 1:4:size(M,1)-3

MF = M(k1:k1 + 3,k2:k2 + 3)

A(idx,idxc) – mean(MF(:);

end end

Color_ Feature = abs(A(1:28))

The Y component of the color calibration shows the amount of energy (intensity) in the sensed color picture. Instead of using the original intensity value, a normalized value is used to overcome this problem. Assume that the real diseased image has M rows, N columns, and three channels, R, G, and B. A color image is collected and three channels are extracted, one luminance Y channel and two chrominance channels. Finally, the optimal threshold value is calculated by p₀ ….p_n-2 between the variance of the corresponding histogram. The following equation shows the whole process

$$\left\{{p}_{0, }{p}_{1}\dots ..,{p}_{n-2}\right\}=avg \&\text{max}\left\{{a}_{2}\left({t}_{0, }{t}_{1}\dots ..,{t}_{n-2}\right) \right(3)$$

Where

$${\sigma }_{n}^{2} ={\sum }_{i=1}^{n}({\mu }_{1}- {\mu }_{n}\left) \right(4)$$

Here µ is for calculating the image’s intensity and cumulating the final class value.

$${\omega }_{i }= {\sum }_{i=c}{p}_{i} \left(5\right)$$

The threshold quality can be determined by estimating the distance ratio between the inter-class and the total variance.

$$\eta =\frac{{\sigma }_{b}^{2}}{{\sigma }_{t}^{2}} \left(6\right)$$

A class fusion algorithm based on expert learning has been developed to automatically separate the discolorations of the sycamore bug from the leaves. The case of k1 > bk1 indicated no sycamore lace bug discoloration (natural leaf discoloration or leaf veins). Hence, the threshold is used for separating these two classes of the mask. Alternatively, the threshold was used unless the leaves were highly discoloured or showed trichomes or mildew spots. Since this is based on the intensity a subjective assessment of disease severity is based on the color distribution of lesions. Thus, the Intensity-based color index can be used objectively to quantify disease severity levels. Different pixels' red and green color values are considered to calculate the I value. According to this paper, the R-G-based Intensity value effectively indicates all diseases on green plant leaves. In diseased pixels, the lesion color ranges from yellow to dark brown.

4.4 Classification

Data models are evaluated based on their performance metrics during image classification. In machine learning, performance metrics are used to measure and evaluate the data.It is generally used for refining the parameters and selecting the appropriate model. (Gayathri Devi and Neelamegam 2018) Accuracy, Sensitivity, Specificity, Precision, and Recall are some of the most common performance metrics. Research and applications for identifying plant diseases can be extended with DL advancements. Early application of the right measures requires fast and accurate models. It depends on the goal of minimization or maximization when choosing the network architecture for a classifier system.

Our paper describes a multi-stage-CNN configuration inspired by the classical and successful Here, LeNet-5 and AlexNet CNN architectures and their improved performance by (Liu et al. 2018).Hence, Convolution, stochastic pooling, and softmax layers are presented in the CNN-based model(Pandian J. et al. 2022) A diagram and related parameters are shown in Fig. 8.

From the input image, edges, lines, corners, and other low-level features are extracted using the first convolutional layer. In each output map feature, multiple input maps are combined with convolutions. Here, the input maps Mj represents a set of input layers, Kij represents the convolutional kernel, bj represents bias, and l represents the lth layer. In addition to sigmoid and tanh functions, CNNs can also be implemented with additive biases.

$${x}_{j}^{l}=f ( \sum _{i \in M j}{x}_{i}* {k}_{ij}+ {b}_{j }^{i} (7)$$

The different kernel dimensions in CNNs were: Pi and Qj are its height and width, and wijpq is its weight at the position (p, q) connected to the layer (i, j), which we call sigmoid(). Unsupervised training is usually performed on CNN parameters such as bias bij and kernel weight wijpq.

$${v}_{ij}^{xy}=sigmoid \left({b}_{ij}+ \sum _{q=0}^{p\left(i-1\right)}\sum _{q=0}^{q\left(j-1\right)}{w}_{ij}^{pq}{v}_{\left(i=1\right)}^{\left( x+p\right)\left(y+q\right)}\right) \left(8\right)$$

From these kernel dimensions the InceptionV3 perception network had the lowest accuracy percentage compared to the other perception networks using the same learning transfer technique. As a result, the AlexNet did not achieve similar results because it had the best results but scored much lower than the authors do. Lastly, some classes in PlantVillage have more images than others, which could result in overfitting if trained incorrectly(Kaur et al. 2022). Finally, if done incorrectly, the Plant Village data set is not balanced due to a lack of balance regarding images per class, which could lead to overfitting.

To measure the performance of these classification methods, experiments are sorted into the following categories to assess the performance of the parameters and select the best analysis methods. A variety of metrics were collected to evaluate the models which is to identify and classify different diseased plant, including sensitivity, precision, recall, and accuracy. The following four statuses are used to determine these measures: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). According to Table 4, the predictive analytics of error rate represents the confusion matrix such as true positives, false negatives, false positives, and true negatives. To measure the correct speculation, the confusion matrix generally consists of two columns and two rows.

Table 4

Confusion Matrix for prediction
Performance observation	Performance evaluation
	Positive	Negative
	TP	FN
	FP	TN

The following measures were used to evaluate the classification model:

Accuracy

A dataset's accuracy is calculated by dividing the number of accurate true positives and true negatives by the number of datasets, as follows:

$$Accuracy= \frac{TP}{TP+TN+FP+FN}*100 \left(9\right)$$

Sensitivity

In terms of mathematical equations, the sensitivity is calculated by the total number of correct positive cases divided by the total number of negative cases.

$$Sensitivity=\frac{TP}{TP+TN}*100 \left(10\right)$$

Precision

In terms of mathematical equations, the precision is calculated by the total positive of total positive and false positive cases by the equation:

$$Sensitivity=\frac{TP}{TP+fp}*100 \left(11\right)$$

The great ability of deep learning is to form a characteristic in an automated way without a human intervention it has led to great results in many fields of research. It gives you precise information on the architecture that suits your needs for designing and implementing classification automation of plant leaf diseases. The computational cost of ResNet50, however, is greater. As a result of many depth layers, ResNet50, ResNet101, and InceptionV3 are not as accurate as the shallower networks, despite being the deepest. Lastly, different CNNs were analyzed in terms of performance and time. The CNN consists of input (IL), and hidden layers (HL), output (OL), which are multiple. Then, the HL includes convolutional layers, RELU layers, which perform activation functions, pooling layers, fully connected layers, and normalization layers. In terms of mathematics, it is obvious that the matrix is cross-correlated instead of convolutional, and the indices in the matrix are significant (Li, Zhang, and Wang 2021).

The proposed model is tested with a plant disease dataset with the evaluation metrics such as precision, sensitivity, and accuracy were tabulated in Table 6. This proposed model (k-means, NIBCF, CNN) provides a higher accuracy than the other existing models, as shown in Fig. 10. It seems that (1) For classifying the corn plant, our method achieves: accuracy of Cercospora leaf spot (CLP) is 86.8%, accuracy of Common rust (CR) is 80.2% and accuracy of Healthy plant is 81.9%. (2) For classifying the Tomato plant, our method achieves: accuracy of Bacterial spot (BS) is 92.8%, accuracy of Early blight (EB) is 89.6% and accuracy of Healthy plant is 95.9%. (3) For classifying the Potato plant, our method achieves: accuracy of Early blight (EB) is 95.3%, accuracy of Late Blight (LB) is 95.5% and accuracy of Healthy plant is 96.3%. Hence, some of the models under the deep learning techniques were also compared here in Table 7(S. Shrivastava, Singh, and Hooda 2015)(Phadikar, Sil, and Das 2013) (W. Zhang, Teng, and Wang 2013) (Asfarian et al. 2013) (Shi et al. 2015). As a result, it can be seen our proposed method such as k-means, NIBCF, CNN produced better accuracy compared to existing model, shown in the Fig. 10.

Table 6. Comparison of different classification techniques
Techniques	Disorders	Precision	Sensitivity	Accuracy (%)
Corn	CLP	91.3	72.2	86.8
	CR	88.4	81.6	80.2
	Healthy	87.8	83.2	81.9
Tomato	BS	87.4	85.2	92.8
	EB	73.4	83.2	89.6
	Healthy	86.4	98.4	95.9
Potato	EB	92.4	94.5	95.3
	LB	93.4	93.4	95.5
	Healthy	97.4	98.2	96.3

Table 7

A summarized survey of plant disease detection using image processing
Reference	Features	No. of Samples	No. of disease	Plants	Proposed technique	Accuracy	Limitations
(S. Shrivastava, Singh, and Hooda 2015)	Color and Texture	1000 images	6	Soybean	SVM, KNN, PNN	90%	Compared to SVM and KNN, PNN gives better classification results. Accuracy can be improved.
(Phadikar, Sil, and Das 2013)	Color and shape	500 images	4	Rice plant	Fermi energy Otsu method, k-means method, Rule Generation algorithm	90%	In the rule-generation algorithm, a rank-based system is used, but classifying requires a long trainingperiod.
(W. Zhang, Teng, and Wang 2013)	Morphological, color, and texture	50 images	6	Jujube trees leaves	GLCM, STEPDISC, DISCRIM, Artificial neural network	91.00%, 89.00%, 94.00%, 88.00%, 73.00%, 81.00%	The number of samples used for detection was less.
(Asfarian et al. 2013)	Color and shape	50 images	4	Rice plant	HSV, probabilistic neural networks	83.00%	Two diseases relatively involved the same color that is one feature combined with another feature which leads to a low accuracy level.
(Shi et al. 2015)	Color, texture, and shape	100 images	3	Cucumber leaf	PNN	91.08%	Increase of computational time since it requires several steps

The ability of deep learning to automatically form characteristics without the intervention of humans has led to great results in many areas of research (Vishnoi, Kumar, and Kumar 2021). Hence, many works have been proposed using deep learning to detect and classify plant diseases, so we have proposed using this CNN to create a tool to protect plants against plant diseases. The results of AlexNet were better, while ShuffleNet achieved them in less time by being below 1.53%. A similar approach was used in our study to detect and locate the disease regions using the activation layers of AlexNet (Palareti et al. 2016).

Agricultural productivity is low primarily due to plant diseases. Most farmers have difficulty controlling and detecting plant diseases. Therefore, farmers should detect these diseases early to prevent further losses. As a result, artificial intelligence should be introduced to help rescue diseases at the earliest possible time and treat them. The proposed method consists of (i) applying advanced segmentation parameters to the data, (ii) a major factor is color analysis, hence, we developed NIBCF for color feature extraction and (iii) finally, a convolutional neural network was used to classify the type of disease and early symptom recognition and warning. According to the results of this study, the proposed method achieves 96.03% accuracy when tested on diseases and compared with other methods. The main challenge during the analysis was controlling the large number of datasets of diseased leaves and the large number of iterated AlexNet CNN. It can be used to provide instant information about plant diseases using this study. In the future, improved background separation methods can be used to separate leaf objects from complex backgrounds to improve performance.

Funding Statement: The authors received no specific funding for this study.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the

present study.

Author Contribution: Archana K.S have developed the model and the source code. Arun S have done

all the simulations.

Data Availability: The datasets generated during and/or analyzed during the current study are available from

the corresponding author on reasonable request.

Ethical Approval: This article doesn’t contain any study with human participant or animal performed by

any of the author

Acknowledgment: The authors would like to thanks Management and Dean of SRM Institute of Science and

Technology, Kattankulathur and SRM Agriculturefor providing us the state of the art facilities to carry out this research work

AGRICULTURAL RESEARCH COUNCIL Annual (2016) (October)
Archana KS (2018a) and Arun Sahayadhas. Automatic Rice Leaf Disease Segmentation Using Image Processing Techniques. International Journal of Engineering and Technology(UAE) 7(3.27 Special Issue 27): 182–85
Archana KS, and Arun Sahayadhas (2018b) Comparison of Various Filters for Noise Removal in Paddy Leaf Images. Int J Eng Technol 7(2):372–374.
Asfarian A, Herdiyeni Y, Rauf A (2013) and Kikin Hamzah Mutaqin. Paddy Diseases Identification with Texture Analysis Using Fractal Descriptors Based on Fourier Spectrum. Proceeding – 2013 International Conference on Computer, Control, Informatics and Its Applications: Recent Challenges in Computer, Control and Informatics, IC3INA 2013: 77–81
Bai XD et al (2013) Crop Segmentation from Images by Morphology Modeling in the CIE L*a*b* Color Space. Comput Electron Agric 99:21–34
Barbedo JGA (2016) A Novel Algorithm for Semi-Automatic Segmentation of Plant Leaf Disease Symptoms Using Digital Image Processing. Trop Plant Pathol 41(4):210–224. http://dx.doi.org/10.1007/s40858-016-0090-8
Camargo A, Smith JS (2009a) An Image-Processing Based Algorithm to Automatically Identify Plant Disease Visual Symptoms. Biosyst Eng 102(1):9–21. http://dx.doi.org/10.1016/j.biosystemseng.2008.09.030
Camargo. (2009b) Image Pattern Classification for the Identification of Disease Causing Agents in Plants. Comput Electron Agric 66(2): 121–125
Darley EF, Middleton JT (1966) Problems of Air Pollution in Plant Pathology. Annu Rev Phytopathol 4(1):103–118
Dash A, Kumar P, and Santi Kumari (2023) Maize Disease Identification Based on Optimized Support Vector Machine Using Deep Feature of DenseNet201. J Agric Food Res 14:100824. https://doi.org/10.1016/j.jafr.2023.100824
Dhaka V, Singh et al (2021) A Survey of Deep Convolutional Neural Networks Applied for Prediction of Plant Leaf Diseases. Sensors 21(14):4749. https://www.mdpi.com/1424-8220/21/14/4749
Espinoza K et al (2016) Combination of Image Processing and Artificial Neural Networks as a Novel Approach for the Identification of Bemisia Tabaci and Frankliniella Occidentalis on Sticky Traps in Greenhouse Agriculture. Comput Electron Agric 127:495–505. http://dx.doi.org/10.1016/j.compag.2016.07.008
Garcia J, and Arnal Barbedo (2016) ScienceDirect A Review on the Main Challenges in Automatic Plant Disease Identification Based on Visible Range Images. Biosyst Eng 144:52–60. http://dx.doi.org/10.1016/j.biosystemseng.2016.01.017
Gayathri Devi T, Neelamegam P (2018) Image Processing Based Rice Plant Leaves Diseases in Thanjavur, Tamilnadu. Cluster Comput
Harvey CA et al (2014) Extreme Vulnerability of Smallholder Farmers to Agricultural Risks and Climate Change in Madagascar Extreme Vulnerability of Smallholder Farmers to Agricultural Risks and Climate Change in Madagascar Author for Correspondence: Phil. Trans. R. Soc.
Jones JW et al (2017) Brief History of Agricultural Systems Modeling. Agric Syst 155:240–254. http://dx.doi.org/10.1016/j.agsy.2016.05.014
Kaur P et al (2022) Recognition of Leaf Disease Using Hybrid Convolutional Neural Network by Applying Feature Reduction. Sensors 22(2):575. https://www.mdpi.com/1424-8220/22/2/575
Kissoudis C, van de Wiel C, Visser RGF, Gerard van der Linden (2014) Enhancing Crop Resilience to Combined Abiotic and Biotic Stress through the Dissection of Physiological and Molecular Crosstalk. Front Plant Sci 5:207. http://journal.frontiersin.org/article/10.3389/fpls.2014.00207/abstract
Li L, Zhang S, Wang B (2021) Apple Leaf Disease Identification with a Small and Imbalanced Dataset Based on Lightweight Convolutional Networks. Sensors 22(1):173. https://www.mdpi.com/1424-8220/22/1/173
Liu B, Zhang Y, He DJ, Li Y (2018) Identification of Apple Leaf Diseases Based on Deep Convolutional Neural Networks. Symmetry 10(1)
Macedo-Cruz A, Pajares G, Santos M, Isidro Villegas-Romero (2011) Digital Image Sensor-Based Assessment of the Status of Oat (Avena Sativa L.) Crops after Frost Damage. Sensors
Munisami T, Ramsurn M, Kishnah S, and Sameerchand Pudaruth (2015) Plant Leaf Recognition Using Shape Features and Colour Histogram with K-Nearest Neighbour Classifiers. Procedia Comput Sci 58:740–747
Nan F et al (2023) A Novel Method for Maize Leaf Disease Classi Fi Cation Using the Image Data. (September): 1–14
Palareti G et al (2016) Comparison between Different < scp > D - <scp > D Imer Cutoff Values to Assess the Individual Risk of Recurrent Venous Thromboembolism: Analysis of Results Obtained in the < scp > DULCIS Study</scp >. Int J Lab Hematol 38(1):42–49. https://onlinelibrary.wiley.com/doi/10.1111/ijlh.12426
Pandian J, Arun KK, Rajalakshmi NR (2022) and G.Arulkumaran. An Improved Deep Residual Convolutional Neural Network for Plant Leaf Disease Detection ed. Muhammad Fazal Ijaz. Computational Intelligence and Neuroscience 2022: 1–9. https://www.hindawi.com/journals/cin/2022/5102290/
Phadikar S (2008) and Jaya Sil. Rice Disease Identification Using Pattern Recognition Techniques. In International Conference on Computer and Information Technology, IEEE, 25–27
Phadikar S, Sil J, and Asit Kumar Das (2013) Rice Diseases Classification Using Feature Selection and Rule Generation Techniques. Comput Electron Agric 90:76–85. http://dx.doi.org/10.1016/j.compag.2012.11.001
Ramegowda V, and Muthappa Senthil-Kumar (2015) The Interactive Effects of Simultaneous Biotic and Abiotic Stresses on Plants: Mechanistic Understanding from Drought and Pathogen Combination. J Plant Physiol 176:47–54. http://dx.doi.org/10.1016/j.jplph.2014.11.008
Rastogi A, Arora R (2015) and Shanu Sharma. Leaf Disease Detection and Grading Using Computer Vision Technology & Fuzzy Logic. In 2nd International Conference on Signal Processing and Integrated Networks, SPIN 2015
Shi Y, Wang XF, Zhang SW, Chuan Lei Zhang (2015) PNN Based Crop Disease Recognition with Leaf Image Features and Meteorological Data. Int J Agricultural Biol Eng 8(4):60–68
Shih FY, and Shouxian Cheng (2005) Automatic Seeded Region Growing for Color Image Segmentation. Image Vis Comput
Shrivastava S, Singh SK, Dhara Singh Hooda (2015) Color Sensing and Image Processing-Based Automatic Soybean Plant Foliar Disease Severity Detection and Estimation. Multimedia Tools Appl 74(24):11467–11484
Shrivastava VK, Pradhan MK (2021) Rice Plant Disease Classification Using Color Features: A Machine Learning Paradigm. J Plant Pathol 103(1):17–26
Singh J (2023) and Mamta Kumari. Socio – Economic Impacts of Agricultural Technologies on Farmers ’ Livelihood. National Academy Science Letters (0123456789). https://doi.org/10.1007/s40009-023-01351-7
Sladojevic S et al (2016) Deep Neural Networks Based Recognition of Plant Diseases by Leaf Image Classification. Comput Intell Neurosci 2016:1–11. 10.1155/2016/3289801
Thakur M (2013) and Baldev Singh Sohal. Role of Elicitors in Inducing Resistance in Plants against Pathogen Infection: A Review. ISRN Biochemistry 2013: 1–10. http://www.hindawi.com/journals/isrn.biochemistry/2013/762412/
Thornton PK, van de Steeg J, Notenbaert A, Herrero M (2009) The Impacts of Climate Change on Livestock and Livestock Systems in Developing Countries: A Review of What We Know and What We Need to Know. Agric Syst 101(3):113–127. http://dx.doi.org/10.1016/j.agsy.2009.05.002
Vishnoi V, Kumar K, Kumar, and Brajesh Kumar (2021) Plant Disease Detection Using Computational Intelligence and Image Processing. J Plant Dis Prot 128(1):19–53. https://link.springer.com/ 10.1007/s41348-020-00368-0
Wang G, Yu Sun, and Jianxin Wang (2017) Automatic Image-Based Plant Disease Severity Estimation Using Deep Learning. Comput Intell Neurosci
Wijekoon CP, Goodwin PH, Hsiang T (2008) Quantifying Fungal Infection of Plant Leaves by Digital Image Analysis Using Scion Image Software. J Microbiol Methods 74(2–3):94–101
Xu G et al (2011) Use of Leaf Color Images to Identify Nitrogen and Potassium Deficient Tomatoes. Pattern Recognit Lett 32(11):1584–1590. http://dx.doi.org/10.1016/j.patrec.2011.04.020
Zhang J, Cheng et al (2012) Detecting Powdery Mildew of Winter Wheat Using Leaf Level Hyperspectral Measurements. Computers and Electronics in Agriculture
Zhang K, Wu Q, Liu A, and Xiangyan Meng (2018) Can Deep Learning Identify Tomato Leaf Disease? Advances in Multimedia
Zhang N, Wang M, Wang N (2002) Precision Agriculture—a Worldwide Overview. Comput Electron Agric 36(2–3):113–132
Zhang W, Guifa Teng, and, Wang C (2013) Optik Identification of Jujube Trees Diseases Using Neural Network. Optik - Int J Light Electron Opt 124(11):1034–1037. http://dx.doi.org/10.1016/j.ijleo.2013.01.014
Zhou R et al (2014) Disease Detection of Cercospora Leaf Spot in Sugar Beet by Robust Template Matching. Comput Electron Agric 108:58–70. http://dx.doi.org/10.1016/j.compag.2014.07.004
Zhu H, Cen H, Zhang C, and Yong He (2016) Early Detection and Classification of Tobacco Leaves Inoculated with Tobacco Mosaic Virus Based on Hyperspectral Imaging Technique.: 1. &t=2&redir=&redirType=. http://elibrary.asabe.org/abstract.asp?aid=46925

A novel approach for Plant disease Classification through Neural Network-Based Color Feature Analysis

Status:

Version 1

Abstract

Figures

1. Introduction

2. Related work

3. Proposed Model for plant disease identification

4. Materials and methods

4.1 Dataset

4.2 Image Segmentation of Plant disease

4.3 Feature extraction: RGB Color Model

4.4 Classification

5. Performance measure of leaf disease prediction

6. Experiment results of image classification techniques

7. Conclusion

Declarations

References

Status:

Version 1