Plants can be attacked by several pathogens, often at the same time, and the leading crops that provide most of our food are no exception. Farmers need to know what diseases are present in a field so they can apply the correct treatment, while crop scientists need to identify diseases accurately in field trials.
Textbooks present beautiful pictures of typical symptoms to teach students how to identify a disease in ideal conditions. Correct identification in the field can be much harder. Many plant diseases display a range of symptoms that overlap with other diseases, so accurate diagnosis may need expert knowledge of the pathogen and the plant. This can easily cause confusion, not least when one needs to score hundreds or even thousands of field plots in a day.
A well-known case where symptoms can be confused is that of two important foliar diseases of wheat; yellow rust and Septoria leaf blotch. Both first appear on a leaf as patches of yellowish chlorosis, and both (despite the common name of yellow rust) go on to form small black dots within chlorotic or necrotic areas. The early chlorotic lesions are usually longer and narrower in yellow rust, while yellow rust telia are slightly larger than Septoria pycnidia and, when studied closely, have a slightly fuzzy appearance. Easy enough to describe the symptoms but sometimes hard to distinguish accurately, especially with less than a minute to score each plot.
In our paper, we explored the potential for the latest advances in machine learning and artificial intelligence to automate recognition of fungal diseases, using wheat as an example. We aimed to assess the scope for this technology to produce a practical tool for farmers and crop scientists. We took pictures of four important foliar diseases: Septoria, yellow rust, brown rust and mildew; in realistic field and glasshouse situations; with diverse weather conditions and on different wheat varieties. We included the kinds of background object that naturally appear in photos of field trials: sky, trees, vehicles, soil, boots and so forth. We trained a deep learning network to distinguish the four diseases (as well as healthy leaves) and, by optimising the network parameters, we achieved a classification accuracy that was competitive with expert pathologists.
Scoring crop diseases is often time-consuming and can require the expertise of trained pathologists. Deep learning has the potential to increase reliability, by taking out human error and bias, and speed up the process of monitoring crops. In this project, we worked closely with plant breeders and we believe an important improvement will be to train a network, not only to recognise different diseases of many crop species, but, to quantify the severity of symptoms. This may eventually support efficient, reliable selection of varieties with improved disease resistance.
Computers will never replace expert pathologists, but they may eventually allow us to spend less of our time scoring field trials!
Megan Long, Matthew Hartley, Richard Morris and James Brown published this study in Plant Pathology:
Classification of wheat diseases using deep learning networks with field and glasshouse images. Classification of wheat diseases using deep learning networks with field and glasshouse images.