Accurate identification of disease and correct treatment policy can save and increase yield. Different deep learning methods have emerged as an effective solution to this problem. Still, the challenges posed by limited datasets and the similarities in disease symptoms make traditional methods, such as transfer learning from models pre-trained on large-scale datasets like ImageNet, less effective. In this study, a self-collected dataset from the DoctorP project, consisting of 46 distinct classes and 2615 images, was utilized. DoctorP is a multifunctional platform for plant disease detection oriented on agricultural and ornamental crop. The platform has different interfaces like mobile applications for iOS and Android, a Telegram bot, and an API for external services. Users and services send photos of the diseased plants in to the platform and can get prediction and treatment recommendation for their case. The platform supports a wide range of disease classification models. MobileNet_v2 and a Triplet loss function were previously used to create models. Extensive increase in the number of disease classes forces new experiment with architectures and training approaches. In the current research, an effective solution based on ConvNeXt architecture and Large Margin Cosine Loss is proposed to classify 46 different plant diseases. The training is executed in limited training dataset conditions. The number of images per class ranges from a minimum of 30 to a maximum of 130. The accuracy and F1-score of the suggested architecture equal to 88.35% and 0.9 that is much better than pure transfer learning or old approach based on Triplet loss. New improved pipeline has been successfully implemented in the DoctorP platform, enhancing its ability to diagnose plant diseases with greater accuracy and reliability.
