Classification of plant species and diagnosis of plant diseases using convolutional neural networks

Image classification provides services in many fields, especially agriculture, health, transportation, and it performes in a very short time and with much less cost, with the artificial intelligence architectures used in man-powered problems. In the field of agriculture, processes such as determination of plant species and estimates of plant diseases are carried out in human lab in a very long period of time, and the treatment to be taken can be delayed. Convolutional neural network, which is one of the deep learning algorithms, offers very successful results in very short periods in the classification of plants and in the diagnosis of their diseases. Within the scope of the study, a total of 8 data sets, including large-scale and small-scale cases, as well as multi-class and low-class, were created using the Plantvillage dataset, which is open to the public. Modelling studies were carried out in which it was used as a classifier in 3 phases: classification of plant species, prediction of plant diseases and diagnosis of plant diseases. In the first phase, Alexnet's comparisons were made on 6 different balanced datasets prepared for the classification of plant species and a better model was tried to be obtained with the changes in padding values. In the second phase, the balanced data set consisting of 2 classes prepared for the plant disease prediction problem was applied as a classifier in the most popular 9 different deep learning models of CNN observed in the literature, and the results were compared and 99 for the 3 cross-validations (CV-3) part. With 91 test accuracy, the GoogleNet/Inception model was chosen as the most optimal model. In the last phase, it has been observed that there is a DenseNet model with the dataset for the plant diseases diagnosis problem and a DenseNet model with a test accuracy of 99.14 for the 10 cross-validations (CV-10) parts. In this study, the literature of which was also carried out, comparisons of different CNN models and the performance of the models rearranged with transfer learning were evaluated.