Comparison of Convolutional Neural Networks Transfer Learning Models for Disease Classification of Food Crops

Abstract

Indonesia is an agricultural country with 29% of the workforce working in the agricultural sector, however, farmers' knowledge and practices depend on informal local wisdom based on inherited past practices. Moreover, identifying diseases in plants is difficult to do with human vision so that intelligent technology is needed. In this paper, an architecture of CNN models such as MobileNetV2, ResNetV50, InceptionV3 and DenseNet121 will be built to detect diseases based on leaf images of several crops obtained from the agroai dataset containing multiple crops namely bean, chili, corn, potato, tomato and tea. The model is used through transfer learning for feature extraction of the trained model with imagenet weights, with 4 fully connected layers. Each model for each crop will be compared to get the best model based on the accuracy of training, evaluation and testing. ResNet50 has the best performance for four types of plants, including bean plants with training accuracy of 99.49%, validation of 99.52%, testing of 98.96%, chili plants with training accuracy of 98.03%, evaluation of 98.75%, testing of 100%, tea plants with training accuracy of 99.62%, evaluation of 99.6%, testing of 99.74% and tomato plants with training accuracy of 99.62%, validation of 99.7%, testing of 99.37%. Moreover, MobileNetV3 has the best performance for 2 types of crops that is corn with training accuracy of 99.22%, validation of 99.69%, testing of 99.55%, and potato with training accuracy of 99.62%, evaluation of 99.60%, testing of 99.74%.