Development of Efficient CNN model for Tomato crop disease identification

Abstract Tomato is an important vegetable crop cultivated worldwide coming next only to potato. However, the crop can be damaged due to various diseases. It is important for the farmer to know the type of disease for timely treatment of the crop. It has been observed that leaves are clear indicator of specific diseases. A number of Machine Learning (ML) algorithms and Convolution Neural Network (CNN) models have been proposed in literature for identification of tomato crop diseases. CNN models are based on Deep Learning Neural Networks and differ inherently from traditional Machine Learning algorithms like k-NN, Decision-Trees etc. While pretrained CNN models perform fairly well, they tend to be computationally heavy due to large number of parameters involved. In this paper a simplified CNN model is proposed comprising of 8 hidden layers. Using the publicly available dataset PlantVillage, proposed light weight model performs better than the traditional machine learning approaches as well as pretrained models and achieves an accuracy of 98.4%. PlantVillage dataset comprises of 39 classes of different crops like apple, potato, corn, grapes etc. of which 10 classes are of tomato diseases. While traditional ML methods gives best accuracy of 94.9% with k-NN, best accuracy of 93.5% is obtained with VGG16 in pretrained models. To increase performance of proposed CNN, image pre-processing has been used by changing image brightness by a random value of a random width of image after image augmentation. The proposed model also performs extremely well on dataset other than PlantVillage with accuracy of 98.7%.