Chapter 5 - Artificial intelligence for plant disease recognition

Abstract

Plant disease is one of the most terrifying challenges threatening global food security every year. Accurately detecting plant health problems is the very first step to tackle the problem, and recently, we have new AI technologies promising to help. With the rise of deep learning, neural networks once again became prime options to tackle a variety of classification problems, especially when digital images are involved. The popularization of this type of technique gave rise to an active community that has made publicly available most of the deep learning architectures developed so far. Comprehensive documentation and detailed tutorials associated with those AI techniques ensure that anyone with basic programming knowledge can carry out experiments almost effortlessly. As a result, there has been an explosion of articles applying deep learning to a wide variety of problems. Despite the remarkable results that are being achieved by deep learning techniques, deep learning models are often applied without the necessary precautions to avoid unrealistic or biased results. There are many subtleties which are rarely mentioned in basic reference material associated with deep learning models, thus being frequently ignored when experiments are designed. The “black box” nature of deep learning models aggravates this issue, because potential problems with the model fitting cannot be easily detected. One way to avoid some of those problems is by training the models with comprehensive datasets that cover the entire variability associated to the problem. Building truly comprehensive datasets is arguably the biggest challenge faced by plant pathology researchers wishing to use AI technologies, because the variability associated to the problem tends to be very high, thus requiring a large number of samples to be properly covered. The objective of this work is to detail each step of the workflow normally adopted for training deep learning models in the context of plant disease classification. Special attention is dedicated to the numerous pitfalls that may render the generated models useless for practical use, as well as to practical ways to overcome some of the challenges involved in the development of technologies dedicated to plant pathology.