AI-Driven Plant Leaf Disease Detection for Modern Agriculture

1. Introduction

For Indian farmers, tomatoes have been identified as their most important source of earnings. Over half of the world’s population depends upon tomatoes as it is most important part of cooking. It provides 20% of the world’s energy supply unchanged and is considered to be an integral part of many traditions and is also able to grow anywhere. In addition, it is also a good source of thiamine, niacin and riboflavin. Especially in India the dishes that are made using tomato also includes animal products such as fish to ensure a perfect nutrient balanced diet. The production of these tomatoes is declining day by day because the crop that is being affected by various diseases. Some of the major tomato crop diseases that are affecting the crop are Bacterial blight, tomato Blast, Brown spot, False smut, Tungro, Leaf Scald, Bakanae etc. It is the responsibility of the farmers to ensure the good yield of the crops. In order to ensure a better yield, the farmer must have thorough knowledge on the diseases that occur and should also have the knowledge on the correct use of pesticides and fertilizers. There are several solutions that have been proposed using some machine learning techniques. In the past few decades there is a rise in the Deep learning technology as it is being used for classifying the images. Deep learning methods are also promising because they are able to achieve high accuracy and existing research aims to increase the efficiency as well as affordability. These created new opportunities in the agriculture domain. By applying these machine learning and deep learning techniques, the farming agriculture is developing as it minimizes the overall losses in the crop production. The solutions that have been proposed until now only focused on the output i.e., they only describe the name of the particular disease that the crop is being affected. Our model which is implemented using CNN and LIME not only talks about the disease but also describes about the features of the diseased crop. It works by approximating the behavior of the model locally, in the vicinity of a particular instance, by training an interpretable model such as a linear model or a decision tree on perturbed versions of that instance. The overall goal of LIME is to identify an interpretable model. Firstly, CNN is used to detect and classify the images which belong to their classes respectively. Then LIME creates a local interpretable model based on these images which can then be useful to explain the features (Liang et al.,2022). This model is now useful to predict the image with the custom trained model. The performance of the model is evaluated on the tomato Image Dataset which contains healthy and four disease classes.