Development of an Active Multispectral System for Agricultural Crop Diagnostics

Abstract

Protecting the global food supply is a historic tradition, with modern advances in sensing technology bringing about unprecedented production capabilities. The feasibility of high efficiency food systems has been demonstrated with reductions in chemical usage, transport emissions, wastage and even improvements in crop quality, yield and taste shown. Such advancements were created by evolving automation, driven by the industry’s increasing labour demand, for example, manual inspection of plants cost £3.5m per year in the indoor tomato industry alone. The reducing availability of skills, and the increasing number of non-ergonomic tasks requiring quantitative assessments of symptoms, has led to widespread implementation of machine systems throughout the supply chain. This reactive movement has catalysed the employment of remote imaging, due to its utility as a diagnostic tool providing an online alternative to manual screening. Spectral technology has been demonstrated in the laboratory by optically monitoring physical changes in plant tissue samples due to disease; a method often used in the process industry to detect impurities or adulterants. There is, however, a lack of systems targeted specifically at plant science duties, due to practical and economic motives, resulting in low precision for close range measurements. The aim of this study was to extend the potential of diagnostic crop imaging, as a vehicle to address current limitations within agricultural productivity. In short, it was to detect previously undetected disease phases in commercially grown crop varieties. In order to achieve this, a system was engineered with the spatial discrimination and measurement sensitivity to detect physiological changes on a leaf-by-leaf scale.v The work allowed the detection method to be taken out of the laboratory and applied on an online basis in an industrial production setting; bringing the benefits into the industry. The prototype was used in conjunction with highly-developed data processing and machine learning techniques, to effectively analyse a variety of crop samples. This was the first implementation of a plant tissue specific active multispectral system and was well received in the phenotyping and production industries. The developed technology was demonstrated over a number of trial experiments, including nitrogen quantification, drought detection, aphid classification and fungal disease detection. Most notably the proposed method was able to detect a hemibiotrophic fungal infection, when imaging oilseed rape plants at plant canopy scale, before the appearance of visible symptoms. Spectral vegetation indices were used to quantify disease severity and distribution within the plant, while classification algorithms were applied to automate the process. As a result of preliminary tests, the structure of the plant was also recorded, in later trials, which allowed for reconstruction of the leaf angle and surface texture. The importance of capturing this structural information was outlined, with its effect on reflectance and classification underlying the patent application for the final system. The results of this study expand the horizons of plant disease detection, and opened up an undeveloped area of agricultural technology. The findings of this study have an international impact, particularly for the plant breeding community, who can enhance their selection of resistant cultivars, with its early and quantitative capabilities. With the inclusion of this technology, and ones like it, automated agricultural systems will be enabled and their full potential realised.

Date of Award	31 Dec 2018
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Hujun Yin (Supervisor) & Bruce Grieve (Supervisor)