Assessing the Performance of Domain-Specific Models for Plant Leaf Disease Classification: A Comprehensive Benchmark of Transfer-Learning on Open Datasets

Agriculture and its yields are indispensable to human life all over the planet. It is an essential part of many countries’ economies and without it the world’s population can not be fed. As such, guaranteeing harvest with minimal loss is a primary objective. One factor that heavily contributes to loss in crop harvesting are plant diseases, which often affect crops and their leaves. A plant’s leaf often carries symptoms that indicate whether or not a plant is infected, but traditional manual approaches to identifying these symptoms are tedious and laborious. Additionally the process of manually spotting diseases can be rather slow in a field where urgency and fast identification are very important. The sooner a disease gets identified, the sooner countermeasures can be carried out. To improve both the accuracy with which diseases can be recognized, as well as increasing the speed at which this can be carried out, deep learning methods have proven useful. Recently the field of plant disease recognition has seen a big uptick in the application of various convolutional neural network (CNN) models for the automatic classification of diseases. There exist many different highly-capable models at this time. There also exists a range of plant leaf disease classification image datasets containing different plants and diseases. However, there seems to be no consensus on which model is best suited to handle this task and the same can be said for the datasets. To the best of our knowledge, prior work has used a wide range of different models with different datasets in the way of feasibility studies, but without comprehensively identifying which models are best used in this field. In this work we test a large number of state-of-the-art CNN models on a wide range of openly available datasets to asses their performance and to identify models that are best suited for this field, in order to be able to built better models, and even new foundation models, based on these findings. 23 models have been tested on 18 datasets, both using transfer-learning and transfer-learning with additional fine-tuning added, for 5 iterations each. Transfer-learning allows models to utilize knowledge obtained from other previous tasks to be used for new tasks, reducing training time and lowering the need for training data. The experiments result in a total of 4,140 having been trained for this work. All results will be compared and contextualized in order to find the best models architecture for plant leaf disease classification as well as assessing which datasets are well suited for this task.