Multispectral imaging for presymptomatic analysis of light leaf spot in oilseed rape

The development of precision agriculture has brought about huge advances in monitoring technologies that allow for quantifiable and early detection of plant stress factors [1, 2]. Despite these technological advances and continuous improvement in plant varieties, yield improvements in many crops have plateaued in recent years [3]. This has been blamed in part on ineffective crop management, due to a lack of reliable tools for in-situ monitoring and intervention in increasingly varying conditions [4].

Plant disease is a leading contributor to global crop losses [5]. The selection of disease resistant crop varieties plays a central role in negating this diminution; typically achieved by visual scoring of symptom severity. Imaging techniques have considerable potential to improve this process, by enabling quantification of disease severity and development rate. This would facilitate breeding for crops with quantitative resistance; with individual traits and associated genes each contributing a small improvement to plant performance but, in combination, providing effective and sustainable disease resistance [6].

Plant diseases result in physiological and morphological alterations, with plant pigmentations caused by pathogen interactions [7]. In foliar fungal disease, infection is quantified by estimating the infected area on the leaf surface; a major limitation is the low accuracy due to the subjective nature of visual assessment [8]. Molecular methods have been exploited in plant pathology to improve the accuracy of diagnosis with serological assays, such as enzyme-linked immunosorbent assay (ELISA), used extensively for detection of pathogens [9] and nucleic acid-based methods, such as quantative polymerase chain reaction, based on using nucleotide primers with high affinity to target sequences. Although molecular methods are relatively low cost, they have disadvantages, as the sample preparation is destructive and labour intensive, and producing specific antibodies can be inefficient with the presence of inhibitors reducing the sensitivity of nucleic acid-based methods [10, 11]. As pathogens often do not spread uniformly inside plants, destructive molecular methods can be non-diagnostic, especially at the asymptomatic stage [10]. Therefore, new methods are required for precise non-invasive, non-destructive and continuous diagnosis.

The light leaf spot (LLS) pathogen (Pyrenopeziza brassicae) is an important disease of winter oilseed rape (OSR) (Brassica napus) [12–14]. This versatile crop has multiple uses, and is responsible for £0.7 billion of the UK agricultural market [15], with LLS responsible for losses of £400 million, between 2012 and 2014, with 95% OSR crops affected [16]. The onset of infection can be subtle, due to the hemibiotrophic nature of the pathogen. This makes in-situ analysis almost impossible during the initial infection stages [14, 17], relying on the emergence of late stage visual symptoms. Thus, novel approaches would benefit plant breeders seeking to develop new resistant varieties as well as farmers where early detection in the field opens the possibility of early intervention by sparse selective fungicide application, improving the efficacy of chemical applications.

Multispectral imaging (MSI) collect light reflected from the leaf surface. The spectrum of the reflected light is governed by both physical and biochemical interactions with the leaf, processes that will change during disease infection [18, 19]. Images are collected at different wavelengths in the optical range (350–1000 nm), where each pixel in the image represented by a vector, known as the spectral signature. Analysis typically uses spectral information to study plant properties and conditions through spectral vegetation indices (SVIs), used commonly in remote sensing to describe vegetation health and density [20]. Several SVIs have been developed to identify and detect plant disease [21, 22].

Using photometric stereo (PS) for analysing plant tissue has attracted limited attention in the literature. Recent work such as [23] has highlighted the usefulness of the approach for extracting high resolution 3D imagery of leaves in order to aid vein extraction. It has been shown that different wavelengths used for light sources affect the validity of the assumption that the surface exhibits Lambertian reflectance. For example in [24], reconstructions of human faces were shown to be more accurate under NIR than visible. This is thought to be due to the sub-surface scattering of NIR, which penetrates further into the skin, leading to a more diffuse reflectance. However, there is a compromise in using NIR for improving overall 3D reconstruction in that surface detail is lost. Whether the same issue affects plant tissue is not clear and will be the focus of further work, but for the purposes of this study visible light is used in order to extract as much higher resolution information as possible.

This study aims to develop a MSI system for disease diagnosis. This has been applied in a controlled environment for the analysis of disease progression and has the potential, with further development, to be used in the field [25]. This MSI system uses narrowband light sources and a broadband silicon detector, in contrast to typical MSI and HSI systems where a broadband illumination source is used with a narrow band (typically diffraction grating based) detector. This selected spectral resolution allows the technology to be applied at a fraction of the cost of traditional systems, whilst improving signal-to-noise ratio (SNR) due to targeted reflectance peaks. The use of machine vision algorithms allows for early symptom detection (before symptoms can be detected by eye) allowing better quantification of disease progression with the potential, in field production scenarios, to allow earlier and hence more effective intervention. The main contributions of this work lie in (1) introducing a refined MSI device, (2) detection of a hemibiotrophic infection before visible symptoms appear and (3) improvements in classification performance using tailored machine learning techniques.

The aim of this study was to demonstrate the detection capabilities of MSI whilst investigating the effect of monitoring at canopy scape versus leaf scale. Detached leaf and entire plant assays were undertaken to provide information with samples at different angles as found within canopies versus the orthogonal orientation achieved using detached leaves and compare the time to diagnosis between the two conditions.

Pathogen: light leaf spot

Pyrenopeziza brassicae penetrates the cuticle directly as germ tubes [26], see Fig. 1, however no appressoria are formed and entry via stomata has not been confirmed [27]. The expression of cutinases assists the initial penetration of the cuticle. [26, 28]. P. brassicae has a hemibiotrophic lifestyle. The fungus initially grows as a hypomecylium where hyphae develop slowly in the subcuticular space between the cuticle and epidermal cells -no cell perforation or systematic spread is observed [27]. Therefore, the biotrophic phase of P. brassicae is visually symptomless, however the fungal growth will affect the plant tissue reflection properties [25], particularly in the NIR. Asexual sporulation signifies the later phase of infection when visual symptoms become apparent. The pathogen interacts with the metabolism of the host plants resulting in morphological and physiological perturbations such as stunting and chlorotic lesions, see Fig. 1. In common with many foliar pathogens, cytokinins are proposed to have a critical role in P. brassicae pathogenicity, promoting the formation of localised carbohydrate sinks and a redirection of nutrients from host to the pathogen. Sporulation causes the breaks in the cuticle and lesions develop, expanding concentrically. Chlorotic regions break down or become sunken due to the separation of the epidermal membrane and cuticular layer, or because of the production of toxins by the pathogen. In cases of severe infection, lesions merge and leaves become necrotic [17]. An LLS isolate (LLS160803, Scotland’s Rural College) was grown on Malt Agar Media (LabM) at \(18\,{^{\circ }}\hbox {C}\) for two months. Conidia were collected by flooding the Petri dishes with 3–5 ml of sterile water plus 0.01% (v/v) Tween20 and agitating the mycelium with a plastic spreader. The spore suspension was collected in a 50ml Falcon tube and adjusted to \(10^{6}\) spores/ml. The spore suspension was applied by spraying to plants when they had formed two true leaves. Infection was ensured by spraying the whole leaf surface area of the host until run-off. Mock controls were inoculated with sterile water plus Tween20 (0.01% v/v). The plants were covered immediately by plastic lids for 48h to maintain the leaf humidity above 90% and the temperature set at \(16\,{^{\circ }}\hbox {C}\) [29].

Trials

In order to investigate the pathogen-host interaction, alongside sensing limitations, two trials were used at different perspectives with plants and detached leaves.

Canopy assay

A susceptible OSR rape (Brassica napus) genotype, Charger, was selected based on moderate resistance rating (4/10) A [30]. The irradiance in the growth chamber was \(200\;\upmu \hbox {mol m}^{-2}\hbox { s}^{-1}\). The study used 18 pathogen-inoculated replicates and 5 mock-inoculated with scans taken at 03, 06, 09, 12, 15, 18, 21, 24, 27 and 31 days after inoculation (DAI). Images were taken of the main canopy for all time points before 24 DAI and taken on a mounted leaf after, due to the size of the samples outgrowing the field of view, see Fig. 2.

Detached leaf assays

Four OSR genotypes, Bristol(2), Charger(4), Cracker(9), Temple(7), were selected on the basis of their resistance, as per their given ratings (/10) [30]. The irradiance in the growth chamber was \(160\;\upmu \hbox {mol m}^{-2}\hbox { s}^{-1}\). At 25 DAI the inoculated leaves were removed with a sterile scalpel and placed on petri dishes within plastic trays with a small amount of distilled water. A small piece of tissue paper was wrapped around the leaf petiole and dipped into the water to keep the leaves moist. The study used 6 inoculated replicates and 6 mock-controls, from each of the four different genotypes. Data was collected at 26, 28, 31 and 34 DAI. Images were taken of leaf samples held in place using a wire mesh, see Fig. 2.

The classification accuracy of the plant analysis (83-97%) was in-line with existing literature (61–99%) [20] however, the time from inoculation to 90+% diagnosis was a significant improvement (12 DAI), particularly due to the clandestine nature of the hemibiotrophic pathogen at early stages of its life-cycle. The active MSI technique allowed detection of disease regions allowing for localised classification of infected tissue areas. The method, with application of machine learning algorithms, allowed infection detection for entire plants at 92% accuracy 13 days earlier than visual inspection. The authors would urge caution with the early rates of classification for ND-SVM, caused by the difference in residue left from the control and pathogen inoculum. Although this is of interest in this study, it is a similar infection mechanism that could occur in-field or during high-throughput trials, which did not give any visible symptoms and thus outlines the requirement for spectral applications.

The dual-scale nature of the experiment allowed for a detailed investigation into the orientation effects that are so problematical in the HSI imaging of plant material [51–53]. The detached-leaf assay also had a number of drawbacks despite being the measurement of reference for flat samples. This difference in the time-scales between the two experimental set-ups, was chosen to reduce the development of chlorosis on detached leaf assay but also meant the infection took longer to manifest; resulting in visual symptoms seven days later for detatched leaves. The simulated view of the entire leaf on a flat plane gave an important insight into the pathogen mechanism on a leaf scale (see Fig. 3) and gave an example of what could be seen on the canopy, should orientation reflections be fully compensated. Since the leaf images were taken alongside the canopy, the significance of this effect on detection rates can be examined (see Fig. 5) with an average increase in classification rate of 15% for flat orthogonal samples.

The 3D capabilities are still in the testing phase as there are a number of limitations to the current design. The PS method manages to represent the surface features with high-fidelity but suffers on discontinuous surfaces; to negate this an OSAVI mask is applied to remove non-plant shapes from the scan, but in future, the inclusion of structured light will prevent this and allow for more macro surface orientation based morphology to be extracted. The current LED design means that a Gaussian distortion is introduced into the reconstruction, which can be seen in the model in Fig. 5; as the curvature of the leaf tip is exaggerated.

The study found that automated FS can play an important role in improving classification performance, when compared to SVIs or even full spectra with SVM. This is due to the limitation of specific wavelengths with mixed relevance to disease ingress and additional data-points obfuscating the classification respectively. A comparison demonstrating the improvement is shown in Table 1. Another important finding was that the improvement in prediction accuracy between the conventional (SVM) and one-class (ND-SVM) classifiers, illustrated in Table 2.

There are not sufficient replicates of the cultivars to be able to infer a clear relationship between the resistance and classification rates. However, since these rates have been defined on visual symptoms [30] then they would not be a good reference when compared to destructive molecular techniques. This system would need some modification before it could be utilised outside the lab environment as the performance would be affected by additional stresses (e.g. nutrient deficiency) and ambient lighting; thus the current design may prove of more value in controlled indoor environments.

There are a number of potential improvements to improve the data acquisition process, with a major focal point on fully incorporating the canopy orientation models into the spectral processing to remove the variation seen in Fig. 5. A faster global shutter system could replace the rolling shutter sensor; allowing a significant improvement in SNR. On-line visualisation of the multispectral cube and 3D information would inform the user of current limitations of imaging technique, which is often not found until the data is analysed. Finally, to improve the static nature of the current system, a common communication interface to a translational stage would allow integration in high-throughput applications. A revised system will be implemented in a field setting, incorporated into the spot-spraying equipment. This location has the added benefit of the shade created by the health and safety enclosures on all operational farm machinery, which in turn improves the SNR.

In this paper, a low-cost active MSI system has been developed. Different analysis techniques have been used with the primary goal of evaluating the application of this system to the study of plant pathogen interactions. The findings clearly indicate the ability to detect disease using spectral information, with the minimum detection level affected by leaf orientation. The paper also exploits machine learning methods to extend the diagnosis beyond user supervised techniques. The outcome of this will not only help to detect the onset of disease but will also help in breeding varieties in the future by extending current breeding capabilities, by allowing for better differentiation of resistant cultivars.