In the modern world with ever rising population, the demand and stress on the agricultural produce to meet the ever growing demand is on record-level high. With the innovations like Smart Agriculture, Natural Fertilizers and Genetically Modified Plants, there still are various areas to focus upon.One of the main areas to focus that is majorly gone unnoticed is the diseases in the plants. In many developing countries with less access to human healthcare there is little to no scope to consider for plant health and disease detection. Due to this a large amount of produce is lost due to diseases in plants. Furthermore most of these diseases can spread from one plant to another starting a domino effect destroying the entire fields.Though some work on this area, the accuracy achieved by the systems can still be increased and systems be made easier to incorporate, use. This paper aims to solve the problem of Detection of Plant Disease by analyzing the image of plant leaves. It aims to increase the existing accuracy of the various existing works with the proposed approach using Transfer Learning and identifying disease in a broader number of leaves than the existing works.
In this modern world, India still depends a lot on Agriculture. Agriculture provides 17% of the GDP and provides employment to more than 60% of the Population. Also this sector is responsible for feeding and providing with various Raw Products for Agro-Based Industries to satisfy the needs of the 1.36 Billion People in India.
With the ever growing demand and the land for the supply being limited and even depleting, it is very much required to make the best use of the resources like land, water. For improving the quantity and quality of the products, many innovative technologies like the use of Genetically Modified Plants produce more Produce per Plant. But still there are huge losses endured due to diseases in plants.
Roughly the losses in agricultural production due to pathogens, diseases, weeds 20% to 40% of the global production. Pathogens and pests are causing 10 percent to 28 percent losses in wheat, 25 percent to 41 percent losses in rice, 20 percent to 41 percent losses in maize, 8 percent to 21 percent losses in potato, and 11 percent to 32 percent losses in soybeans on a global scale, according to a study published in the journal Nature, Ecology & Evolution. In this paper we acknowledge and focus on the disease of the Plants. We believe that the diseases of the plants if detected earlier can help to contain the spread of the disease and also help to produce more.
In this paper, we aim to take up this issue of losses due to diseases in plants and propose a method of solving it. So, in this paper, we aim to study the existing models and propose a model which will help in identifying disease with good accuracy in the leaves of Tomato, Apple, Blueberry, Cherry, Grapes, Corn, Orange, Peach, Raspberry, Soya bean, Squash, Strawberry.
The proposed work focuses on 26 different categories of diseases that occur in the mentioned 12
types of plants.
2. Literature Review of Existing Methods
[
KNN
[
From the survey done on the existing techniques, we found many techniques, where the most popular being K-means , SVM and Bayesian classification, and ANN(Artificial Neural Networks). We were not able to find many techniques using Transfer Learning for the purpose.
Based on the review, it was observed that though some work has been done in the field, no widespread work has been done taking into account a generalized and large number of plants and many diseases. Most of the present works are concentrated on diseases in 1 category of plants.
Therefore, it is required to propose a generalized approach to predict numerous diseases of several plants.
As a result, the goal of this research is to provide an approach/method for classifying leaf detection with improved accuracy rates for different leaves and disease categories.
For any supervised learning project, the main component is the dataset. For this project, we used the
publicly available PlantVillage dataset [
The dataset was accessed directly from Kaggle Notebook which was used to make the model for the paper.
Over 50,000 pictures of healthy and diseased plant leaves are included in this collection. It contains the infected images in 38 categories where 26 categories of diseases of the 12 plants, namely, tomato, apple blueberry, cherry, grapes, corn, orange, peach, raspberry, soya bean, squash, and strawberry Input leaf image.
The disease into which the diseases were classified:1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.
For the dataset, the image contains background noise. In order to extract the relevant region from the input picture, the Tiramisu model must be used. It is based on DensNet,where all the layers are interconnected. Also the Tiramisu model adds skip connections to the up-sampling layer like Unet.
For making the model which would be trained on the images, we incorporated the transfer learning technique.
Transfer learning is a novel method to deep learning in which models that have been pre-trained for one task are repurposed for another.
One of the biggest advantages being that there is no need for extra feature extraction step. This is because they are very deep neural networks where the initial layers act as feature extractor.
In our case we went on with the use of Inception V3 pre-trained neural network. It is a family of Inception neural networks where all the previous features of inception v1 and v2 are incorporated along with label smoothing, factorized 7x7 optimizer, RMSProp optimizer and BatchNorm.
Also, in comparison with its counterparts like VGGNet, Inception networks work better and provide more computationally efficiency both in the terms of parameters generated by the network and the economical cost incurred in terms of memory and other resources.
In this model we added further layers at the end to get the prediction into the 5 categories as we desired. After the model being compiled with trained it with our large dataset with 25 epoch and batch size of 16.
The model is being trained on the training dataset, which is produced by taking 70% of the entire dataset for training and 30% for testing.
In the model, we were able to achieve an accuracy of 96.74% on Validation Accuracy.
In this section, we are briefly explaining the result of the proposed model.
With the Tiramisu model, we were able to extract the only leaf image from the overall image.
The approach aimed to capture a large base of leaves that can be checked for any diseases. With the work, we were able to achieve the aim of targeting 12 different plant leaves and detecting 26 different diseases.
With the work we were able to achieve pretty good performance as we can see from the visualizations.
The system when fed with the input of the image of a plant leaf, predicts the top 5 diseases with bar chart visualization. With this, we aim to address anomaly that can occur in predictions. Meaning it provides options and its opinion on the diseases that the plant leaf has with the amount of confidence in each diseases.
Also it provides the user,i.e., farmers to view alternative diseases that can be present.
We had started work on the paper focusing on detecting multiple diseases in multiple plants. We focused on 12 categories of plants namely
On our model, we achieved an accuracy of 96.74%. With our paper, we conclude that with the model developed we can classify the images into vast 26 categories of the diseases from 11 different plant types efficiently and effectively with an overall accuracy of 96.74% higher than any of the existing works and also into more categories than any existing works.
We think that the future work on this can be to convert this model into an application, which can easily be used by the real farmers to use the model for fast and efficient disease identification.