A Comparative Study on Deep Learning Models for Plant Disease Detection and Organic Solutions

Abstract

Farming is essential to the economy of each and every country. Consistently, ranchers grow many harvests. Crop and forest products for both food and non-food products are included in agriculture. [ First, agriculture was an important part of the development of sedentary human civilization. One of the primary causes of crop devastation and inappropriate crop development is infection and disease. Several factors associated with plant disease diagnosis using deep learning techniques must be taken into consideration when developing a robust system for accurate disease management. This model utilizes profound learning and picture handling to distinguish plant sicknesses and prescribe natural answers for treating the plants. This model makes use of convolutional neural networks, which are crucial for visual imagery. Images are processed and libraries are extracted with Tensorflow. Keras is likewise utilized for highlighting extraction expectations and adjusting hyperparameters. In light of the kind of plant and sickness, this model additionally suggests natural arrangements like fertilizer, vermicompost, bonemeal, etc. Moreover, the model distinguishes bothers and proposes natural cures, for example, rejuvenating balms, cow manure, and neem glue to safeguard the harvest from harm. The project\'s objective is to reduce the economic and aesthetic damage caused by plant diseases and to provide farmers with an intuitive interface for organic plant cultivation. In order to provide the most precise Deep Learning results, this model looks for the ideal approach. Organic Solutions and a Model for Predicting Plant Diseases

Introduction

I. INTRODUCTION

The protection of plant health is essential in every industry. Plants not only provide non-food crops for energy, fiber, feed, and horticulture, but they also contribute to the world's food supply. The physiological functions of plants that are disrupted by persistent phytopathogenic organisms (biotic or infectious disease agents) are referred to as plant diseases. At various cultivation stages, it suffers from a variety of diseases. Because of the vast amount of land under individual farmers' control, the variety of diseases, and the occurrence of multiple diseases in the same plant, early detection and treatment of such illnesses is beneficial for ensuring high quantities and quality. In remote areas, expert knowledge in agriculture is unavailable, and the process takes time.

It developed an automated method for farmers to send images of diseased leaves to our server, where a neural network will identify the disease and provide the farmer with the disease classification as well as the organic solution that corresponds to it because everyone nowadays has a cell phone. Using our plant disease dataset, which was gathered over several months, the proposed architecture developed a deep learning strategy inspired by work on convolution neural networks. Sequential pre-trained VGG-19 models that have been trained on a lot of ImageNet data are used to identify plant diseases.

We must choose the most effective management strategies whenever we need to take control measures. Plant selection and cultural practices are the first lines of defence. As an essential component of agricultural sustainability, organic farming significantly enhances soil fertility.

Organic methods are cost-effective, improve the structure, texture, and aeration of the soil, increase its capacity to retain water, and encourage the development of healthy roots. Minerals, animal waste, sewage sludge, and even plants are all sources of organic fertilizer. Soil organic matter content increased as a result of vegetables, animals, and residue materials. As a result, it is recommended that you use integrated nutrient management, which involves the scientific management of animal manure, plant residue, and sewage sludge in order to improve soil productivity over time and maximize crop growth, yield, and quality.

The second section explains about the survey done to know about the plant disease detection using image processing and deep learning models,third section ensure to know about the various models used,its advantages,its architecture,the fourth section gives the description of datset and preprocessing the data,fifth section gives the details of results.

II. LITERATURE REVIEW

Preventing yield and quantity losses in agricultural products necessitates the detection of plant diseases. The study of patterns in plants that can be seen is an important part of research into plant diseases. For sustainable agriculture, disease detection and monitoring of plant health are essential. Plant diseases are extremely difficult to manually monitor. It requires extensive work, knowledge of plant diseases, and a prolonged processing time. Consequently, plant diseases are detected through image processing. Disease detection involves several steps, including image pre-processing, image acquisition, feature extraction, image segmentation, and classification. Images of plant leaves can be used to identify diseases, as discussed in this paper [1]. The introduced technique exhibits a mechanized strategy for crop infection ID utilizing Nearby Twofold Examples (LBPs) for highlight extraction and One Class Order for arrangement on different leaf test pictures relating to various yield species. The proposed method makes use of a separate Class Classifier for each plant health condition, such as healthy, powdery mildew, downy mildew, and black rot. When tested in other crops, the vine leaf-trained algorithms displayed exceptionally high generalization behaviour. A unique algorithm that proposes conflict resolution between one-class classifiers provides correct identification when ambiguous data instances may belong to more than one condition. A total success rate of 95 percent was achieved for the 46 plant-condition combinations tested[2].

Agricultural productivity is an essential component of the Indian economy. Consequently, the contribution of food and cash crops is essential to humans and the environment. Consistently, crops are obliterated by various illnesses. Inadequate disease diagnosis and ignorance of disease symptoms and treatments result in the death of numerous plants. This study gives an outline of plant illness location utilizing different calculations. This paper proposes a CNN-based approach for identifying plant diseases. Simulation studies and analyses of time complexity and infected region area are carried out on sample images. It is achieved using a picture-handling strategy. The model was fed 15 cases, 12 of which were successful [3].

It is difficult to localize and classify various insect species due to their high degree of similarity in features, making it even more challenging when dealing with those that have already been caught in traps. Inspired by the Deep Convolutional Neural Network (CNN) achievement, this paper proposes a method for identifying various species of trapped insects based on available images. This study developed a novel method for detecting plant diseases by combining four CNN models with a database of 200 images [4].

In this experiment, we used an open-source database of 36,258 images classified into 10 plant species and 61 classes of healthy and diseased plant leaves.

Two datasets were created from 36258 images. 4540 for the validation set and 31718 for the training set. Four CNN models, Commencement, Resnet, Origin Resnet, and Dense Net, were broadcast and the sequences of CNN models were processed using a stacking technique. Compared to the results obtained using a single CNN model, the accuracy rate obtained by the stacking method is 87%, a significant improvement. A relatively high accuracy rate suggests that a combination of CNN models was used [5].

Identifying plant diseases is essential to prevent loss of crop yield and quantity. It requires considerable labour, knowledge of plant diseases, and longer processing times. This makes it possible to detect plant diseases through image processing. Image acquisition, image pre- processing, image segmentation, feature extraction, and classification are all components of disease detection. This article described a strategy used to detect plant diseases using leaf images. In this article, we have discussed various methods of segmenting diseased parts of plants. In addition, this article described trait extraction and classification methods used to extract traits from infected leaves and classify plant diseases. Image processing can be used to accurately identify and classify plant diseases that are critical to successful crop cultivation. In this article, we have discussed various methods of segmenting diseased parts of plants. In addition, this article described trait extraction and classification methods used to extract traits from infected leaves and classify plant diseases. Classifying plant diseases using ANN methods such as self-organizing feature maps, backpropagation algorithms, and SVM. can be used effectively [6].

The color transform structure RGB is converted to HSV space since HSV is a good color descriptor. The application of biostatistics to detect plant leaf diseases is described. Cover and remove green pixels with a pre-processed cut-off level. The 32x32 patch size is used for segmentation in subsequent steps, thus yielding useful segments. These segments are used in texture analysis of the color co-occurrence matrix. Especially when you compare texture parameters to regular leaf parameters. [7]. Getting a high-quality image with as much color detail as possible was the hardest part of the job. Getting an image with all the details stored in a memory that can be processed is a common task. Such images are composed with a lofty goal and thus can be 6-10 MB in size. The Nikon D5200 camera used for this was very effective. The second hurdle was to eliminate the lighting conditions, as the lighting conditions are very different from the beginning to the end of the paddy season, even though the shooting time is fixed. However, the answer to this is variable client-defined threshold management and a significant adjustment to LCC tint.

When plants get sick, the quantity and quality of produce is greatly reduced. Identifying disease side effects with the naked eye is difficult for ranchers. Computer-aided image processing is used to protect crops, especially on large farms, and diseased leaves can be identified by color. Depending on the application, many image processing methods have been learned and solved through design validation and several programmed applicators. This white paper provides useful overviews of these proposed systems in the following sections. Although there are many automated or computer vision methods for disease detection and classification, research on this topic is still lacking. No single method can identify all diseases [8]. This study used image processing and machine learning techniques to identify and classify cotton leaf diseases. Investigations into background subtraction and segmentation methods were also discussed. According to the results of this research, color space conversion from RGB to HSV is useful for background removal. Moreover, thresholding was found to outperform other background subtraction methods in terms of results. To perform color segmentation, we first removed the background from the image by erasing the green pixels and then thresholded the resulting binary image. This helps in pinpointing disease features. We found that SVM yields excellent results in terms of infection ordering accuracy. The proposed work consisted of five main steps, of which he carried out three. Image segmentation [9].

Target regions (prevalence patches) are then identified by R, G, and B color feature image segmentation. Image features such as borders, shapes, colors, and textures are later extracted for disease patches to identify the disease. Cotton leaf spots, cotton leaf color segmentation, edge detection-based image segmentation, and disease analysis and classification are the three components of this work [10]. In this section, we propose and experimentally test a software solution for the automatic detection and classification of plant diseases by image processing. In rural India, farmers have limited access to agricultural experts who can examine images of their crops and offer advice. Farmers often get the right answers to their questions too late. This paper develops an image processing algorithm that can automatically identify crop problems based on image color, texture, and shape, and automatically detect diseases and other conditions that can affect crops. to address this issue and provide farmers with fast and accurate solutions. SMS [11].

III. MATERIALS AND METHODS

We used photos of potato leaf diseases from the public Plant dataset to train, validate, and evaluate the proposed model. The photographs in the collection showed two distinct conditions for potato leaves: typical

i.e sound and an unhealthy one. Right now, the dataset is being loaded: The plant leaves dataset is utilized without augmentation. It contains 56,725 images representing 39 plant species, of which 42,312 have been trained and 14,413 have been tested. For both the training and testing datasets, the class model is categorical, and the image and batch sizes are set to 48 and 64, respectively.

Manipulation of images: This produces groups of picture information as tensors from continuous datasets. In this step, the batch processing model is defined. When the input exceeds a certain threshold, the transform function activates a node; The output is zero when the input is less than or equal to zero.

The proposed system's detection of these diseases assists in selecting the appropriate treatment to prevent plant infection. A specialist in the field who is able to treat these diseases with the naked eye is required for detection; however, this is a prohibitively expensive procedure; consequently, there is a demand for detection systems for crop diseases.

1. The candid and computationally resourceful method for identifying leaf disease and recommending organic solutions is provided by the proposed system.
2. Concepts of Convolutional Neural Networks like VGG-19, Sequential, and Mobile-net are included.
3. This system enables the identification and recommendation of the name of the disease and the organic remedy that is specific to that disease.

Advantages of the Proposed System

a. To identify key elements without the need for human intervention; to detect leaf diseases using CNN algorithms in less time.

b. Following proper identification, organic solutions are recommended.

A. Methodology

1. Dataset

New Plant Disease Dataset: This dataset was reconstructed using an offline supplement from the original dataset. You can find the original data set in the GitHub repository. This dataset contains approximately 87,000 RGB photos of both healthy and diseased plant leaves and is divided into 38 classes. The directory structure is preserved as the full dataset is split into a specific ratio of training and validation sets. Later, 33 test photos are taken in a new directory for prediction purposes.

2. CNN Model

The architecture of Convolution Neural Networks (CNNs) determines how well the network performs. It consists of the fully linked layer, the pooling layer, and the convolution layer. A deep CNN called VGG is used to categorize photos.

3. Layers Flow

The input to this network was a fixed-size RGB image (256*256), so the matrix was in the shape of a square (256,256,3).

The only pre-processing performed was to calculate the mean RGB value for each pixel in the entire training set, which was then subtracted from each pixel.

They covered the entire image using kernels with a size of (3 * 3) and a step size of 1 pixel. Spatial padding was used to preserve the spatial resolution of the images.

Max pooling was performed with side 2 across a 2 by the 2-pixel window. A rectified linear unit (ReLu) was then used to add non- linearity to the model to improve classification accuracy and computation time.

Unlike previous models that used tanh or sigmoid functions, this one performed much better.

B. Architecture of Plant Disease Detection

Therefore, due to the enormous amount of land under the control of individual farmers, the wide variety of diseases and the prevalence of many diseases in the same crop, early detection and treatment of such diseases is useful to ensure high quantity and quality. However, it is really difficult to do this.

Acquiring agricultural expertise is time-consuming and not possible in rural areas. Eliminating the need for image pre-processing and opening up built-in feature selection, convolutional neural networks have recently made great strides in image-based detection. Another problem that is quite difficult is finding huge datasets for such challenges. If the dataset is fairly small, it is preferable to use a model that has been previously pre-trained on a large dataset.

A neural network type called a Convolutional Neural Network or CNN or ConvNet is particularly good at processing inputs with a grid-like architecture like an image. A binary representation of visual data is a digital image. The brightness and color of each pixel is specified by its pixel value and they are arranged in a grid-like pattern. The human brain analyzes a vast amount of data and concepts while looking at an image. The entire visual field is covered by the connections between each neuron, each of which works in its own receptive field and has its own network of connections in the receptive field.

4. CNN Model

The loaded picture data set needs to be used for both training and testing. Images and class labels are stored separately in arrays for training purposes. The train-test split strategy utilizes 70% of the information for preparing and 30% for testing. After 30 percent of the data is used for validation, another 70 percent is separated.

The class labels are converted to integers with the help of a dictionary. Keras is delinked of the fully linked layers. changes to the system that can't be taught. After flattening the output of a SoftMax feature extractor, it applies the SoftMax filter.

A model was constructed from the ground up using the Adam optimizer and category cross entropy as the loss function to classify plant disease identification. After fifty iterations, there was no change in the outcomes. The model was trained using VGG-19, Sequential, VGG-19 with Early Stopping, and Mobile Net models. The VGG-19 model performed well and produced better results than any of these approaches.

5. Databases and Connection to Model

For each disease of the plant, organic methods of control have been gathered. The authoritative Google pages, books on plant care, and manuals for organic farming were used to compile these organic remedies. For each plant disease, several organic remedies are gathered.

MYSQL databases were used to create an entry in the database for each of these organic solutions. Two tables were created with two columns—one for the name of the disease and another for an organic remedy—to identify plant diseases.

A connection to the databases and previously trained models was established by utilizing the Python MySQL Connector. Python is used to connect the model for identifying plant diseases.

V. RESULTS AND DISCUSSION

Python is used to combine the plant disease identification model, and the databases for plant disease organic remedies are linked to this integrated model using Python's MYSQL connector package.

To determine whether a plant is contaminated, a picture of the plant must be uploaded. Organic solutions are not offered if healthy. Otherwise, the disease's name and natural treatments are offered.
Five alternative algorithms, including VGG-19, Inception, MobileNet, ResNet, and VGG-16, were used to train a model of plant disease. The MobileNet model out of these five has the highest accuracy and outperforms the others in terms of performance.

Conclusion

This model makes use of a deep learning model to recognize plant ailments and then suggests natural remedies fit for the general care of the plants. The dataset has over 80K images of plant diseases. Several pre-trained models, including VGG-19, VGG-16, Mobile Net, and Inception, were used to build the CNN model, with MobileNet emerging as the top performer. For each plant and insect, organic remedies are gathered and entered into MySQL databases. After combining both datasets and CNN models, the user is shown the results. The aim is to give farmers detailed information on how to manage various plant diseases. The accuracy of the results must be improved by the use of real-time data. Giving farmers detailed information on how to manage various plant diseases is the aim. The accuracy of the results must be improved by the use of real-time data. This can be expanded to offer some DIY gardening solutions to home gardeners. This can also be applied to a mobile application, where the services could be offered via NLP in more languages.

References

[1] A.B. Patil, S.D. Khirade, Plant disease detection using image processing, in: Proceedings of the International Conference on Computing Communication Control and Automation, 2015, doi:10.1109/iccubea.2015.153. [2] D. Moshou, X.E. Pantazi, Automated Leaf Disease Detection in Different Crop Species through Image Features Analysis and One Class Classifiers, A.A. Tamouridou Aristotle University, School of Agriculture, Agricultural Engineering Laboratory, Thessaloniki 54124, 2021Greece [3] Deepsikha ; Garima Shrestha; Majolica Das; NaiwritaDey: Plant disease detection using CNN INSPEC Accession Number:20257553 doi:10.1109/ASPCON49795. 2020.9276722, 2021. [4] https://extension.sdstate.edu/organic-pest-control-methods [5] https://extension.psu.edu/organic-and-biological-control-of- plant-diseases [6] https://earthsally.com/disease-control/common-plant- diseases.html [7] Jun Liu, Xuewei Wang :Plant diseases and pests detection based on deep learning: a review,2021 [8] Tuannam Nguyen, PhanDuy Hung :Pest detection on Traps using Deep Convolutional Neural Networks: DOI:10.1145/3232651.3232661,2018 [9] Xulang guan, Brownell Talbot School, Omaha, NE, US, A Novel Method of Plant Leaf Disease Detection Based on Deep Learning and Convolutional Neural Network, doi:10.1109/ICSP51882.2021.9408806, 2021.

Copyright

Copyright © 2023 Dr K. Prasanna, B. Nayanika Reddy, Chaganti Sathvika, D Himaja, Barenkala Laxmi Prasanna. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.