Abstract: These days crops are being affected by various types of bacteria and fungi which leads to huge menace to food patronage. It is very difficult to identify these diseases, because of the paucity of the basic infrastructure. As the smartphones are emerging and improvement in computer-mission enabled that neural network has suggested for wireless-based disease diagnosis. Using a dataset that includes different types of ailing and healthy plant leaflet images, we train a Deep Neural Network (C-NN) to diagnose different types of crop diseases. The model we created achieves a good accuracy as shown in the result page. This approach creates a path for wireless-based crop analysis on a gigantic scale.
Keywords - Convolution neural network, deep learning, plant diseases detection.
INTRODUCTION sustainable development are to scale back
the usage of pesticides, cost to save lots of India is eminent for Agriculture meaning the environment and to extend the standard. most of the people are engaged towards the Precise and early diagnosis may reduce the agriculture industry. The agriculture usage of insecticides pesticides. industry acts as a big role within the economic sectors. Most of the plants are Plant disease can be defined as anything that infected by various fungal and bacterial causes unhealthy to the plant. There are diseases. Thanks to the exponential different reasons for that, which includes inclination of the population, the climate poor nitrogen, air pollution, bacteria, fungus, also causes disease. The main challenges of virus, etc. This damage leads to poor growth of plants and loss of productivity. This will be a solution to every problem, the same also cause major loss in the economy in way we have a solution to this problem, terms of agriculture worldwide. Early deep learning helps us to come out of this detection can help us to take necessary problem, which is used to detect the type of actions or remedies which can help to raise disease. the productivity as well as economic in Nowadays, technology is widely used for agriculture. disease prediction. Plant diseases aren't only Due to this damage, crucial agricultural a risk to food haven worldwide, it can also plants are being threatened by different have catastrophic aftermath for small-scale types of diseases and pests. As this damage farmers whose source of revenue rests on can reduce the food yield and its quality we crop yield. Within the emerging world, can say that it is a major threat to food almost eighty per cent of the agricultural security which in turn leads to illness of a product is generated by small-scale farmers, man. This disease has the capability to lower and reports of produce loss of fifty per cent. the vegetable and fruit quality. About Furthermore, the most important fraction of Forty-Two percent of the crops in the world hungry people sleep in small-scale farming are destroyed yearly due to the damage of communities, making them a gaggle which the crops. Therefore to reduce this damage, is particularly liable to pathogen-derived a good and accurate diagnosis is needed. interruptions in the food supply.
The main support of human existence on Whenever the disease is identified correctly earth is with the help of food and the plants. on its earliest appearance could be an Plants which produce good quality food are important step for effective disease control. prone to a variety of diseases because of This process is supported by many external conditions of the environment. Due organisations. Furthermore, due to to this, there is a huge loss in agriculture. advancement in technology this plant Therefore early detection and necessary diagnosis can be done online by providing remedies are to be taken to save our valuable the required information. time and money. In the way that there will The current process for identifying diseases completed on these images and group them is observation through the naked eye by into an array. Then it's divided into testing experts. For this process, they should and training. Then it's fed to our C-NN continuously monitor the plants which can model and accuracy is noted. be expensive. At an equivalent time in a few In this project, we propose a unique deep countries, they don't have suitable services, learning architecture that uses a maybe they will contact experts, but Convolutional Neural Network (C-NN). consulting them can be laborious and These algorithms take images as input and expensive in most cases. In these cases, the convolve it with predefined filters as a part most favourable technique for farming of the process. A BXB sized image is taken community is to opt for automatic detection and it is convolved with a kernel of a certain for plant diseases which operates just by size. This process helps the model to learn a observing the leaf of a diseased plant and it's feature of the image and these features after a non-expensive process each operation are learnt by a respective Plant disease detection by observing is a feature map. These maps can help us in strenuous task as well as it is not accurate understanding the image and work with enough. If the above-recommended weights. technique is employed, it will take time to CNN can be contained in different detect the disease in a visual way. Most convolution layers in which sets of different common disease symptoms that can be seen filters are used to identify the required in plants mostly occur on a leaf which could pattern counting with other layers like resemble coloured spots which can be a sign pooling. In classification of image, patterns of fungal or bacterial infection are squeezed through convolution and the In this report, we present an efficient and different layers which are present in the effective method for disease detection. The model repetitively to get the outputs workflow of disease detection is described respectively. So therefore to train a model, a as follows. First, images are located in the dataset is given for training as input to the directory and therefore the pre-processing is model along with that, the weights and kernel parameters are given to extract the applied not only in the agriculture domain, patterns in the hidden layers to solve the but also in various domains which is problem. If an error is occurred at the output successful. So therefore we apply this side, then by using a technique called back technique for leaf disease detection propagation is used to reduce the classified automatically. This detection in the early error. In this model we have visualized the stage could be beneficial in terms of layers present, to identify the behaviour of economic and production. the method design for plant disease diagnosis.
SYSTEM ANALYSIS The images we have used in this model were taken from the internet sources. The dataset Existing System contains both types of ailing and healthy Outbreak of plant diseases can lead to death plant leaflet images which are of fifteen and famine. The result of famine in 1943 thousand images with fifteen different types which was caused by helminthosporium of crop diseases. The images were rifted into caused a heavy loss of crop yield and deaths training and testing data-sets that comprises in North India. These losses are caused by the ratio of 7:3. The image size we have not taking preventive steps to protect the taken in this model is 256 X 256 which are crop-yield by identifying the disease and used to feed into our model. This is highly treating them. recommended architecture as there are repeating convolution sections such that it Continuous monitoring of experts is needed makes it difficult to interpret by the other which is the main default method for layers in the model. identifying crop diseases. This is an expensive method for both small scale and In the recent study, it has been said that deep large scale farmers. Furthermore, in learning which is a front line for image developing third world countries, there is a based processing gives us accurate or shortage of experts and consulting them meticulous results and huge potential in agriculture areas. This modern technique is could be expensive and time consuming for training and remaining thirty percent for them. testing. These images are generally appearing in the form of a matrix holding the values of the cells in the form of Proposed System neurons. These neurons are fed as an input to the layers of the network. The There are a number of strategies currently construction of the CNN model is done for used to utilize computer-based systems of further analysis. Now, the model is tested plants derived from images to specific for implementation. If the model is accurate, classifier systems as input parameters. then the data is satisfied with the network or In this proposed system the images are else the model is modified for achieving collected. These images are given as input to better construction. our proposed classifier and it displays the disease it has affected with. To make this possible we have created a model using Convolutional Neural Network. This can be useful to both small scale and large scale farmers as this could be more accurate and less time consuming. Moreover it could be helpful to poor farmers as it could be less expensive.
SYSTEM DESIGN System Architecture
The diagram below shows how the
prediction mechanism occurs in the model. At first the dataset is pre-processed, after that it is converted to an array. From the dataset, seventy percent of data is used for Fig. Process flow of complete project. General explanation : 1.The input test image is acquired and preprocessed in the next stage and then it is converted into array form for comparison. 2.The selected dataset is properly segregated and preprocessed. 3.The model is trained using CNN and then classification takes place. 4.The comparison of the test image and therefore the trained model happen followed by the display of the result.
Fig. Use case diagram
UML DIAGRAMS One of the basic methods to show how a Use Case Diagrams system works is to use a UML diagram. It defines and develops a Single Modeling From the above diagram, the user interacts Language (UML) via the user case study. with the system by uploading an image, after The connection between actors, their pre-processing an image, the predicted objectives and their dependencies is output will be displayed. supplied by the functionality of the system. State Chart Diagram The functionality of the individual actors will be demonstrated. The roles of the actor State charts are used in “computer” science are performed throughout the process. and other related fields. State diagrams allow a finite number of states to make up the structure described often in this case, sometimes these are logical abstractions. There are several types of state diagrams that are slightly different from different ReLu, convolution_layer, normalization logical expressions. layer and dropout Layers. In Recognition, an image is taken as an input as well as label name conversely class name would be the output.
Input Layer
In the input layer of C-NN an image is given
as input. Deep learning models need enormous image data for training them. The number of channels is equal to the size of the image. For example if the number of channels are three then it is a colour image and for gray-scale image it is one. The size of the dataset can be increased by data augmentation techniques, these techniques are done on image before these are fed into the model. These augmentation techniques include transposing, zoom, cropping and rotate operations whereas conserving the Fig. State chart diagram label of the image
Convolutional Layer
ARCHITECTURE OF DEEP This is the first layer of C-NN and images
are taken as input. The major operation of CONVOLUTIONAL this layer is to convolute the images. With the help of filters the image is convoluted NEURAL NETWORK and produces the feature maps. The output of this layer can be written as A C-NN is identical to a neural network; it encompasses input as well as output layers in inclusion to this a few middling hidden layers are also present among them. These hidden layers encompass a pooling layer, ReLU Layer scaled by an element gamma. The equation for output of this layer is This Layer is also known as the activation layer, Relu is used to induce non-linearity to the system.It is also used for neutralizing negative activation values. ReLu is initialised after every convolution. In case of multiclass classification ReLu can be used but for binary classification sigmoid function should be used Fully Connected Layer
It implies, these are the layers that each and
every node is connected to the next node in the coefficients. These coefficients support each and every node in the pooling data. The multiple sets of output from the pooling layer is connected to F-CL. The job of FCL is to collectively draw the top best cases for the object or an image. At the end of this layer it follows different techniques which Max-Pooling Layer includes S-VM, soft-max for the probability distribution to pick the top best solution. Its In Max pooling, the image is down-sized for equation is output. No learning is done in this layer. The output from Conv layers is a feature map, after pooling this map is of reduced size and conserving its features in the output based on filter and strides defined. This layer can help us avoid overfitting.
IMPLEMENTATION Batch Normalization Layer
In this layer we can find activation of each Preprocessing
channel by subtracting and dividing mini-batch mean and standard deviation. Pre-processing the dataset will improve the This layer comes before the ReLu layer. The accuracy and performance. We perform input is shifted by a beta offset and then it is pre-processing techniques on our image After creating our CNN model we will be dataset by resizing or reshaping them. After optimizing it using Adam optimizer from that, we check for a .DS_Store file which Keras library and the loss function is cross holds the attributes /metadata information. If entropy. it is present is removed and then the images Then the training is commenced by calling a are converted into an array. This process is model.fit_generator to which image data done on the whole dataset which includes generator is supplied and trained. C-NN has test and train data. The dataset consists of different layers like Activation, Flatten, almost 15000 images of different diseased Convolution_2D and Max_Pooling_2D. and healthy plants. The training data is After the completion of training, the model employed to train the model so that it can can detect disease among plant species. classify the required test-set. After Training and pre-processing, comparison of the trained model and test image takes place to predict the disease. Training
After Pre-processing the images we will be
converting them into array using a function. Testing Using label_binarizer from scikit-learn library we will convert each image label to binary levels and this instance is saved using pickle. Then we further pre-process the data by scaling the data points from [0, 255]. After that we will be splitting our data into test/train sets using 80 percent of images for training purposes and rest for testing. An image data generator object is initialised After training the classifier we test its
which performs random shifts, rotations, knowledge on the test set which we created
crops and flips on our image dataset. earlier and accuracy is calculated After testing the model we save it with the ● After this the dataset is help of a library called pickle. Then the divided into test and training model is saved in GDrive. To test the model set based on requirement. we trained we will be uploading an image ● After pre-processing, the next belonging to a certain plant which is move is to encode the labels affected by the disease. After uploading the with LabelEncoder and file in drive we will be accessing it using its thereby dividing the dataset file path and then pre-process and predict into two groups, train and test the class it belongs using predict_class sets. Additionally we will be function and print the result. transforming input features to Tensors for both numeric and categorical variables and RESULT column features are described. In this project we have designed a program using CNN. The programming language being used is python. Its packages which are We have trained our model and saved it being here are numpy, sklearn, matplotlib, before exploring results. Let's have a glance keras, scikit-learn. The following are the at the summary of our model. The summary steps to be taken in of our model’s Architecture is order to achieve experimental results.
● The initial step is to import
the necessary libraries and read the data from the file to the dataset. ● Pre-processing the dataset and encoding it so that the machine can understand and easily parse it. exploring results. The testing accuracy resulted is 97.00118%.
For testing we use a separate image
which has to be pre-processed, hence for testing any image can be used as a test image. If any disease is identified in the image, the label will be shown and the aim of the project will be achieved. CONCLUSION Among image-classification methods, C-NNs beat outdated image-processing This project aims to implement an inventive methods. This is also observed in the idea to determine the damaged crops. By automatic detection of PlantDiseases. We using the above-mentioned technology, the images are given as input to it or technology have trained our model and saved it before for the recognition of diseases. This model is very helpful for the agriculture community, especially in remote villages. As India is a developing country, this technology can help and what possible solutions are automatic in many ways. This system helps in terms of once the disease is identified. reducing the time to detect diseases. This technology helps us for fast identification of the disease and take necessary actions or remedies for good productivity in agriculture. As further the project can be developed or implemented by using multimedia REFERENCES (Audio/Video) to tell what type of disease is [1] Detection and Classification of Plant [4] Plant Disease Detection and Leaf Diseases by using a Deep Learning Classification by Deep Learning PMCID: Algorithm. ISSN: 2278-0181. PMC6918394 PMID: 31683734
