This article showcases Matogen Applied Insights’(MAI) analytics capabilities within the agricultural sector, examining how employing deep learning for plant disease image recognition can be used to identify crop disease.
Even in the modern era, plant disease still poses significant risks to crops, despite the widespread use of crop protection products. In fact, misdiagnosing plant disease can result in the misuse of chemicals, leading to the emergence of resistant pathogen strains. This, in turn, results in increased input costs and more outbreaks, impacting profitability, as well as the environment. Currently, disease diagnosis is conducted by humans, which is both a labour- and cost-intensive. Thanks to advances in computer vision technology, there is an opportunity to employ deep-learning based modelling to enable automated plant disease image recognition in order to increase efficiency in this process.
The Department of Plant Pathology at the University of Stellenbosch utilizes the latest technology as part of its holistic and interdisciplinary approach towards limiting the impact of plant diseases. In its research and training, conventional and molecular techniques are combined to control plant pathogens and increase plant resistance in a sustainable and economic manner for the benefit of both local and export markets. The department features research programmes in grape, deciduous fruit, citrus, vegetable and cereal crop disease.
The department also contains the Plant Disease Clinic, a service laboratory where specialists in the field of plant pathology diagnose problems on received samples. Diagnosis is conducted on bacterial, fungal and viral disease. The clinic also offers an insect identification service in cooperation with the university’s etymology department. Digital images, either via a conventional camera or smartphone, or through a microscope lens, have become a useful diagnostic aid, enabling verification by experts worldwide.
Plant disease image recognition
Building on this methodology, Matogen Applied Insights (MAI) has been providing ongoing technical support to the Plant Disease Clinic by engineering a comprehensive platform and workflow system. The platform includes a database incorporating migrated historic institutional data, reports and images. Preventing loss of historic information is crucial, especially in the case of rare plant disease.
The plant disease data warehouse contained within this platform enables detailed trend analysis and cutting-edge modelling. The machine learning technique, Convoluted Neural Network, was applied to the vast repository of plant disease images, in order to construct a model that would identify a specific plant disease on a previously unseen image (in theory, sent from the field on a smartphone, for example) with high levels of accuracy.
The model is able to classify a given image into “diseased” or “healthy” categories with great accuracy. Subsequently, distinctions are made between many different types of diseases. Rare cases and novel symptoms are also accommodated and variations in depth perception (angle, light, shade, leaf age) are addressed. In addition, the platform allows for the incorporation of expert knowledge for identifying, annotating, quantifying and guiding the computer vision search for relevant features.
Computer vision use case
Efficiency increases gained by the type of modelling outlined above, pose unparalleled opportunities to save time and money in combating plant disease, ultimately improving food security.