Deep learning applied to plant pathology: the problem of data representativeness

The rise of deep learning techniques has profoundly impacted both research and applications of pattern and object recognition in digital images. In plant pathology, the number of scientific articles on the use of deep learning for disease classification using images has grown steadily for at least a decade and targeted most important agricultural crops. Results have been encouraging, with accuracies of many prediction models usually approaching 100%. It is now widely accepted that, enough data being available, deep learning models can solve most of the image classification problems. However, determining what “enough” means in each context is far from trivial because this involves not only the number of samples used for training, but also the quality, in particular the representativeness of the dataset. More important than having a large sample size is to guarantee that all the variability associated to a given classification problem is represented in the dataset. Achieving this goal is particularly challenging for plant disease images because the agricultural environment is non-structured and highly dynamic, containing numerous variables that introduce variability to the problem. To make matters even more difficult, image annotation is time consuming and prone to inconsistencies due to its subjectivity. As a result, all studies in the literature employ datasets that represent only a fraction of the whole range of the variability, and many of these do not even acknowledge the limitations of the experimental conditions. Experiments with limited scope are valuable in the early stages of emerging research topics, but the application of deep learning to plant pathology has matured to the point where new studies need to contribute something more substantial. Unfortunately, many of the recent publications have been redundant, differing from previous research only by the adoption of slightly different experimental setups and improved model architectures. To move forward, new studies in this field need to address the data gap problem more effectively. This article delves deep into some technical and practical issues to achieve this goal and to increase the usefulness of the future studies. Although this article is dedicated primarily to proximal images, many of the remarks also hold for images captured using unmanned aerial vehicles.