How to improve measurement of major wheat diseases using artificial intelligence?

Fungal diseases of wheat cause significant yield losses and threaten food security. Especially damaging are the five diseases that affect wheat leaves: septoria tritici blotch (STB), yellow rust (YR), brown rust (BR), powdery mildew (PM), and tan spot (TS). In the UK, fungicides and cultivar resistance are the primary ways to control them. Yet, both measures often fail because of the evolutionary adaptation of pathogen populations. Pesticides may negatively affect the environment and human health. Consequently, growers need to reduce fungicide use in favour of integrated pest management (IPM) that emphasizes cultural/genetic control. To achieve this, cropping systems need local adjustments, which requires comprehensive disease monitoring. However, current national monitoring [e.g. UK Department for Environment, Food and Rural Affairs (DEFRA) surveys conducted by ADAS uses visual scoring, which limits the number of fields sampled at a given budget, and has restricted accuracy/reproducibility. These factors limit the utility of the data.

The student will develop a novel method to measure the five diseases using digital imaging (DI) and artificial intelligence (AI). The method will exceed existing methods in accuracy and speed. When used in disease monitoring, the method will enable increased sample sizes and improved data quality. This will lead to a more informative disease monitoring, allowing policy-makers to learn more about IPM uptake, its effects, and boost its adoption.

Objective A. Develop a new method to detect the presence of each disease on wheat leaves using DI and AI techniques.

Objective B. Develop a new method to quantify the amount of each disease within wheat leaves using DI and AI techniques.

Objective C. Develop a portable platform to acquire high-quality images of wheat leaves in the field.

Objective D. Test the new method in subsets of DEFRA/ADAS's surveys.

Detection of disease presence is an image classification problem, while quantification of the amount of disease is an image segmentation problem. For both problems, AI techniques based on convolution neural networks (CNN) offer powerful solutions. The student will build CNN models based on the Python TensorFlow library. The models will be trained on reference datasets comprising leaf images labelled with their disease status (for image classification) and having pixels annotated as those corresponding to healthy or diseased leaf areas (for image segmentation). The student will acquire reference datasets that are sufficiently large, and capture the diversity of leaf phenotypes and disease symptoms.

Extensive datasets of STB- and YR-diseased leaf images are available in lead supervisor's lab (>6,000 images for each). The student will use these datasets to create high-quality reference datasets. To acquire reference datasets for BR, PM and TS, the student will conduct replicated field experiments that capture contrasting wheat phenotypes and disease resistance levels. Using these reference datasets, the student will train CNN models to measure the five diseases (Objectives A, B).

To improve measurement speed/logistics, we will develop a portable field platform (PhenoBox) that captures leaf images rapidly and non-destructively. PhenoBox consists of a plastic box to block ambient light, a photocamera with a light source will be mounted at the top and a slit at the bottom to insert the leaves. The student will test PhenoBox's prototypes to achieve a working configuration (Objective C).

We will bring together the PhenoBox with the trained CNN models, and test the system as part of DEFRA/ADAS's surveys. In parallel with ADAS's field scorers, the student will capture large numbers of leaf images in a subset of farmers' fields (10 out of 300). The student will use the trained CNN models to measure the five diseases, determine the optimal sample size and evaluate the accuracy gain versus existing methods (Objective D).