Title: Spectral signatures in the UV-range can be combined with secondary plant metabolites by deep learning to characterise barley – powdery mildew interaction

In recent studies, the potential of hyperspectral sensors for the analysis of plant-pathogen interactions was expanded to the ultraviolet range (UV; 200-380 nm) to monitor stress processes in plants. A hyperspectral imaging set-up was established to highlight the influence of early plant-pathogen interactions on secondary plant metabolites. In this study, the plant-pathogen interactions of three different barley lines inoculated with Blumeria graminis f.sp. hordei ( Bgh, powdery mildew) were investigated. One susceptible genotype (cv. Ingrid, wild type) and two resistant genotypes (Pallas 01, Mla1 and Mla12 based resistance and Pallas 22, mlo5 based resistance) were used. During the first five days after inoculation (dai) the plant reflectance patterns were recorded and in parallel plant metabolites relevant in host-pathogen interaction were studied. Hyperspectral measurements in the UV-range revealed that a differentiation between barley genotypes inoculated with Bgh is possible and distinct reflectance patterns were recorded for each genotype. The extracted and analyzedanalysed pigments and flavonoids correlated with the spectral data recorded. A classification of non-inoculated and inoculated samples with deep learning revealed that a high performance can be achieved with self-attention networks. The subsequent feature importance identified wavelengths, which were most important for the classification, and these wavelengths were linked to pigments and flavonoids. Hyperspectral imaging in the UV-range allows for a characterisation of different resistance reactions, can be linked to changes of secondary plant metabolites with the advantage of being a non-invasive method and therefore enables a greater understanding of the plants' reaction to biotic stress as well as resistance reactions.


Introduction
In recent plant phenotyping studies, the ultraviolet range (UV,  have various tasks such as visual signals, auxin transport, and resistance against plant pathogens (Petrussa et al., 2013). High contents of flavonoids such as kaempferol and pelargonidin were found in plants exposed to high levels of UV-radiation (Monici et al., 1993). Anthocyanins protect chloroplasts from

Accepted Article
This article is protected by copyright. All rights reserved UV-radiation but can also scavenge reactive oxygen species (Neill & Gould, 2003).
High levels were found in primary leaves of barley, where they are stored in the epidermis, because of their high absorptive properties of UV-light they can also be found in the mesophyll (Liu et al., 1995).
The flavone chrysin, present in different cereals (Liu et al., 2010) is often used to quantify flavonoids with spectrophotometric detection since their absorption maxima are at 240-290 nm as well as 310-370 nm (Mierziak et al., 2014). Flavonoids cannot only be synthesized by plants as a response to stress but also produced before the occurrence of stress and stored at important sites to play a direct role in the defensedefence mechanisms (Treutter, 2006). Studies proposed that they can be stored in epidermal cells and are released into the infected tissue where they might be involved in hypersensitive responses (HR) (Beckman, 2000). In addition, degradation of plant pigments like carotenoids and chlorophyll a and b with an absorption maximaabsorption maximum at 400-500 nm (Taniguchi & Lindsey, 2018) can be linked to a changing photosynthetic activity due to compatible and incompatible interactions (Brugger et al., 2018). Not only flavonoids are affected by an infection of barley with Bgh, but also other plant compounds. For example, genotypes susceptible to Bgh showed a reduced electron transport capacity due to a degraded photosystem II which results in a loss of chlorophyll during the infection development (Scholes et al., 1994).
Resistance breeding is a major protection strategy against Bgh infections in barley.
The cultivar Ingrid (wild type (WT)) is susceptible to infections while the near-isogenic lines cv. Pallas has two isogenic lines 01 and 22, which are resistant and show no typical symptoms of a powdery mildew infection. Pallas 01 has a race specific resistance and possesses the resistant genes Mildew loci Mla1 and Mla12, which

Accepted Article
This article is protected by copyright. All rights reserved cause a HR after recognizing specific Bgh aviruelence genes (Schulze-Lefert & Vogel, 2000). The near-isogenic line Pallas 22 contains a dysfunction in the mlo5 gene and has a broad spectrumbroad-spectrum papilla based resistance against Bgh (Kølster et al., 1986). A papilla or cell wall apposition (CWA) is quickly developed below the penetration point of the pathogen and prevents further infection development. CWAs contain phenols, belonging to the secondary plant metabolite group of flavonoids (Jørgensen 1992).
The susceptible and resistant barley-Bgh interaction has previously been studied with hyperspectral imaging in the visible (400-700 nm) and near infrared range (700-100 nm) with emphasis on reflectance and transmission data (Thomas et al., 2017;Kuska et al., 2015). Reflection measurements enabled an early detection of infection two days before colonies became visible by the naked eye (Thomas et al, 2017). In addition, the genotypes were differentiated according to their susceptibility to Bgh (Kuska et al., 2015) and these this data were was combined with microscopic observations (Kuska et al., 2017) and results from invertase analysis (Kuska et al., 2018). At present, there is no research available that links the secondary metabolism of the plant with spectral changes of wavelengths. The genotypes used in this study serve as a model to prove the usability of the UV-range to describe phytopathogens and their effects on host plants.
Three hypotheses were investigated in this study: (i) an infection with Bgh affects the secondary plant metabolites, (ii) the changes in secondary metabolism can be detected non-invasive with hyperspectral imaging and (iii) the relevant wavelengths can be narrowed down by combining the recorded hyperspectral data with deep learning algorithms (Fig. 1). The last point is particularly important in order toto be able to limit future investigations to the relevant wavelengths of the UV-range in a

Accepted Article
This article is protected by copyright. All rights reserved targeted and cost-saving manner.

Plant cultivation and pathogen inoculation
The barley lines cv. Ingrid wild type (WT), Pallas 01 (Mla12), and Pallas 22 (mlo5) were grown in a greenhouse environment in plastic pots (5x5 cm) on commercial substrate (Topfsubstrat 1.5, Balster Erdenwerk GmbH, Sinntal-Altengronau, Germany) and wateredwatered, as necessary. After 10 days when reaching growth stage 11 (Witzenberger et al., 1989) the primary leaves were cut at 10 cm and place on 10 g/l phyotagar (Duchefa Biochemie B.V, Haarlem, Netherlands) containing 0.34 mM benzimidazole (Kuska et al., 2015). For each genotype 10 leaves were kept untreated as a control group while eight technical replications with 5 leaves each were inoculated with fresh spores of Blumeria graminis f.sp. hordei isolate A6. The agar plates were sealed and incubated at 19°C in a controlled environment with 1100 cd x m -2 illuminance and a photo-period of 16 h per day.

Extraction of secondary plant metabolites
For the extraction of secondary plant metabolites samples of inoculated and non-inoculated barley leaves of all three genotypes were collected 1 to 5 dai and kept in the freezer at -80°C until analysis. The extraction was carried out for 6 biological replications.

Chlorophyll and carotenoid extraction
Chlorophyll and carotenoid extraction was performed according to (Scholes et al., (1994). Frozen leaf samples with 0.5 M HClO4 were grounded in liquid nitrogen to a

Accepted Article
This article is protected by copyright. All rights reserved fine powder. Subsequently, 0.5 g per sample was transferred into a tube stored on ice and 1.5 ml of 80% acetone was added. The samples were kept for 3 h on ice in the dark and mixed every 20 min. Following the samples were centrifuged by 4°C at 13,000 rpm for 20 min and the absorption of the extract was measured at 470, 645, and 663 nm. The concentration of chlorophyll as well as carotenoid was calculated according to (Hiscox & Israelstam (1979).

Total flavonoid extraction
Total flavonoid extraction was performed according to Mihai et al., (2010). Frozen leaf samples were grounded in liquid nitrogen to a fine powder and were extracted with 96% ethanol. For this 1 g of each sample was mixed with 30 ml ethanol and kept overnight with constant stirring. Following the samples were filtered on qualitative filter paper for three times before the volume was added up to 100 ml for an initial extract concentration of 1%. Of this extract 3 ml were mixed with 1 ml 2.5 % ZrOCl2 and 21 ml methanol were added. After 30 min the absorption was measured at 288 nm against a blanc solution consisting of methanol. A calibration curve was established using chrysin. For this, a stock solution with 0.1 mg/ml was prepared and aliquots of 0.25, 0.5, 1, 1.5, and 2 ml were used. The measured absorbance was plotted against the concentration to establish the calibration curve.

Hyperspectral image acquisition and data preprocessing
Spectral reflectance was recorded with a hyperspectral imagine line scanner in the UV-range according to (Brugger et al., (2019). with an exposure time of 800 ms, a frame rate of 0.4 frames per second and a linear axis speed of 0.3 mm/s. The reflectance was daily measured 1 to 5 days after inoculation (dai). The relative

Accepted Article
This article is protected by copyright. All rights reserved reflectance was calculated using the software Headwall Hyperspec III (Headwall Photonics, Bolton, MA, USA) and for this purpose, a white reference image of 95 % barium sulfatesulphate and a dark current image were recorded. Data was then analysed with the software ENVI 5.5 (Exelis Visual Information Solutions, Boulder, CO, USA). An amount of manually selected 7500 pixels of each sample was used to calculate the average reflectance. The selected pixels covered the entire barley leaf.
A Savitzky-Golay filter with a window size of 7 and a 3rd degree polynomial was used to pre-process the data. The Savitzky-Golay filter and the corresponding parameters were selected to reduce the signal noise while preserving the properties of the signal distribution (Madden, 1978). Figures were generated with SigmaPlot 14 (Systat Software GmbH, Erkrath, Germany). RGB visualization was performed 1 to 5 dai.

Experimental Setup of classification of non-inoculated and inoculated samples
The superpixels of the data were used as input for the (deep) learning algorithms to distinguish the measurements into non-inoculated and inoculated samples. A superpixel is defined as the average of P x P neighboringneighbouring pixels. A spatial area with P=7 was selected so it is likely to contain symptoms. The computed areas are non-overlapping. The data were split into two different sets for training and testing. The test set contained 20% of the data and the results were cross-validated so that five different models were trained and evaluated for each classification task.

Determination of relevant features with self-attention classification networks
To determine the meaning of the characteristics of the hyperspectral data, the neural network architecture Self-Attention Networks (SAN, Skrlj et al., 2020) is used. SANs are motivated by recent advances in natural language processing, e.g., through the

Accepted Article
This article is protected by copyright. All rights reserved language model BERT (Devlin et al., 2018) with the Transformer Network Architecture (Vaswani et al., 2017). A key feature of these architectures is the so-called self-attention mechanism. Skrlj et al., (2020) have shown that self-attention can also be used to identify the relevant features per data point and that doing so can result in better classification accuracy than previous methods. The network architecture can be described as follows (1) ( 2) where K is the number of parallel self-attention blocks Ω, X is the input, and W as well as b are learnable parameters of the network. The functions a and σ are activation functions. In this case a is a SELU (Klambauer et al., 2017) and σ softmax function. The symbol  corresponds to the Hadamard product while the symbol  refers to the Hadamard sum formation over individual blocks. The first neural network layer is used especially for maintaining the connection with the input features F. The input vectors are first used as input for the softmax-activated layer, whose neurons match with the number of features F. The softmax function is defined as following: (3) The self-attention mechanism effectively creates a sparsely populated input area and only emphasizes relevant features for solving the task at hand (e.g. language comprehension). In this way, the self-attention mechanism is often used to, e.g., better learn the relationships between words (Skrlj et al., 2020). In the present paper, SAN is used for the classification of individual signatures (pixels) or combined signatures (superpixels) of hyperspectral data in the UV-range, classified between healthy and inoculated samples of different genotypes. The networks' self-attention mechanism weights the input features and uses the weighted output for the neural

Accepted Article
This article is protected by copyright. All rights reserved classification network. Two self-awareness heads and a neural classification network with 64 latent neurons were selected for the Self-Attention Networks (SAN). The final feature importance is calculated by averaging the feature importance of the two heads (4) where X is the evaluated set of inputs and in this case k=2. The training setup of by Skrlj et al., 2020 was followed and the network was trained with the commonly used Adam optimization algorithm (Kingma & Ba, 2014). However, the described batch size seemed to be very low. To achieve a more stable training it was increased to 128 samples. A learning rate of 0.001 was used and it was gradually reduced during training. In a pre-processing step, the first wavelengths were removed, since they were characterized by increased noise, resulting in an input wavelength range from 260.232 nm to 501.219 nm.

Gradient Boosting as classification method
The framework XGBoost (Chen & Guestrin, 2016) was used to implement the baseline Gradient Boosting (GB) classifier. The machine learning technique GB can be used for classification as well as regression problems and generates a predictive model in the form of an ensemble of weak predictive models, usually decision trees.
As in other boosting methods the model is build up stage-wise. The hyperparameters to optimize the GB model were chosen empirically. The reported results were achieved using a maximum tree depth of 6 and a learning rate of 0.3. Further, a L2 regularization was used with the weight λ = 1.

Accepted Article
This article is protected by copyright. All rights reserved

Impact of compatible and incompatible barley Bgh interactions on pigment and flavonoid concentration
Chlorophyll, carotenoid, and total flavonoid content presented changes depending on the resistance of the host plant. The total chlorophyll content of all genotypes decreased from 1 to 5 dai. In WT leaves there was a strong decrease from 5.2 µg/ml 1 to 3.1 µg/ml 5 dai, whereas in mlo5 leaves only a slight decrease from 5.7 µg/ml 1 dai to 5.1 µg/ml 5 dai was observed (Fig. 2). The carotenoid extraction revealed a similar pattern for WT leaves. The highest value was measured 1 dai and decreased by 76% until 5 dai. Mla1 leaves demonstrated a strong decrease from 2.2 µg/ml 1 dai to 1.4 µg/ml 3 dai but exhibited no further decrease 5 dai. mlo5 leaves were represented by a constant carotenoid concentration. The flavonoid content of WT leaves decreased from 49% 1 dai to 40% 5 dai. Mla1 leaves displayed a decrease of 12% from 1 to 3 dai and stayed constant 5 dai. mlo5 leaves demonstrated an increase from 47% 1 dai to 59% 3 dai and remained constant 5 dai.

Phenotypic development of Bgh on barley leaves
The phenotypic development of non-inoculated as well as with Bgh inoculated barley leaves of Ingrid and the near-isogenic lines Pallas 01 and Pallas 22 were visualized with RGB images (Fig.3 and Fig.4). Non-inoculated leaves of all three genotypes were represented by constant vitality and showed no visible disease symptoms

Classification of non-inoculated and inoculated samples with deep learning
The superpixels of the data were used as input for the (deep) learning algorithms to distinguish the measurements into non-inoculated and inoculated samples.
A superpixel is defined as the average of P x P neighboring pixels. A spatial area with P=7 was selected so it is likely to contain symptoms. The computed areas are non-overlapping. As classification method, a neural network architecture with attention mechanism so called self-attention networks (SAN) (Skrlj et al., 2020) was chosen. Attention-based neural networks are a rather novel deep learning architecture which have achieved recent advances in various fields. However, they have not been considered as feature importance extractors for biological or hyperspectral data. The attention mechanism in the first layers of the network was used to classify particularly important elements of the feature space and filter them out of the rest. Afterward, the classification layer used the as important identified parts to classify the input. The data were split into two different sets for training and testing. The test set contained 20% of the data and the results were cross-validated

Accepted Article
This article is protected by copyright. All rights reserved so that five different models were trained and evaluated for each classification task.

Feature importance identifies relevant wavelengths
The self-attention block of the trained network was used to determine the feature importance. For this, the softmax-activated output of the self-attention block was extracted. Fig. 6 visualizes the average over the cross-validated models. Only features whose weighting is above 5% are highlighted for easy presentation. In all three genotypes, 262 or 264 nm were identified as important wavelengths, for WT leaves 263 nm was identified as well. In addition, 280 and 478 nm were also identified for the susceptible genotype. For both resistant genotypes, 501 nm was additionally identified.

Changes of secondary plant metabolites can be linked to relevant wavelengths
The wavelengths identified by feature importance are in the range from 262 to 291 nm and 442 to 500 nm. In both ranges, WT and Mla1 leaves were represented by an increase of reflectance from 1 to 5 dai and alsoand showed a decrease of chlorophyll, carotenoid, and flavonoids. mlo5 leaves displayed a decrease of reflectance in the range from 262 to 291 nm and an increase of flavonoids but similar values of reflectance were recorded in the range from 442 to 500 nm from 1 to 5 dai and extraction experiments presented a small decrease of chlorophyll and constant

Accepted Article
This article is protected by copyright. All rights reserved value of carotenoid content.

Discussion
In this study, hyperspectral imaging in the UV-range was used to investigate changes of secondary plant metabolites as well as pigments in susceptible and resistant barley powdery mildew interactions. It is well described that hyperspectral imaging can provide non-invasive information on host-pathogen interactions by assessing specific changes in plant reflectance patterns, which can be associated with

Accepted Article
This article is protected by copyright. All rights reserved stopping further development (Jørgensen et al., 1994). Inoculated plants containing the resistant Mla1 gene also showed no symptoms but from 4 dai on brown necrotic spots were visible on the leaves. Once the penetration peg ruptures the cell wall and enters the epidermal cell, a race-specific resistance gene recognizes the Bgh avirulence factors and H2O2 is produced (Caldo et al., 2004). H2O2 triggers a HR and causes cell death of penetrated epidermal cellcell, which is visible in brown, necrotic spots (Schulze-Lefert & Vogel, 2000).
The interaction of the sensor and illumination caused significant peaks from 250 to 321 nm as well as 410 to 440 nm which could not be removed by normalization.
Therefore, a correlation of these wavelengths to specific secondary plant metabolites cannot be verified. Untreated barley leaves showed no changes in spectral signatures and no senescence during the observation period, which corresponds to the phenotypic development in of barley leaves (Fig. 3). The increase in reflectance of inoculated WT leaves from 450 nm onward can be linked to a decrease in plant pigments like chlorophyll and carotenoids (Brugger et al., 2018,;Scholes & Rolfe, 1994) and is also reflected in the pigment analysis. Previous studies showed that already 1 dai, susceptible WT leaves show a reduced photosynthetic performance due to reduced pigment content in the leaves (Brugger et al., 2018). Similar toLike WT, Mla1 leaves showed an increase in reflectance from 450 to 500 nm starting 2 dai and a corresponding decrease in chlorophyll and carotenoid content. It is described that Mla resistance leads to a deterioration of photosynthesis and pigment metabolism, as HR causes necrosis (Matile et al., 1999). Previous studies with hyperspectral imaging showed that a reduction of plant pigments can be already

Accepted Article
This article is protected by copyright. All rights reserved released rapidly in the infested tissue (Beckman, 2000). As this occurs in a Mla1-based resistance and results in a decrease in the flavonoid amount, the decrease of flavonoids in inoculated Mla1 leaves can thus be explained. In addition, a previous study suggested that the release of flavonoids occurs particularly in early stages of infection, which explainings the strong decrease of flavonoids from 1 dai to 3 dai in the present study. In contrast to inoculated WT and Mla1 leaves, mlo5 leaves featured a decrease of reflectance from 250-370 nm during time-series measurements corresponding to the strong increase of flavonoids 3 dai. Flavonoids are involved in auxin metabolism, which causes a tightening of the plant tissue (Beckman, 2000).
This leads to the formation of callose and cell wall-phenolics (von Roepenack et al., 1998)

Accepted Article
This article is protected by copyright. All rights reserved disease in soybean stems (Nagasubramanian et al., 2019) and other applications (Tetila et al., 2019;, Polder et al., 2019, Schramovski et al. 2020. Data mining methods were also used to localize HR from hyperspectral imaging data before they were visible on RGB images (Kuska et al., 2017) were also consistent with the spectra recorded in the UV-range. This can be used as information to characterize different host-pathogen interactions as they lead to differences in secondary plant metabolites, which was also reflected in changes in reflectance. Thus, it could be shown that hyperspectral imaging in the UV-range leads to information about changes in secondary plant metabolites such as chlorophyll, carotenoid, and flavonoids depending on susceptibility or resistance reaction.
This study showed that spectral information in the UV-range of different host-pathogen interactions corresponds to changes of secondary plant metabolites.
Specific resistance responses in incompatible barley-Bgh interactions can also be differentiated by spectral reflectance. In addition, deep learning revealed that these secondary plant metabolites can be linked to wavelengths which are of importance

Accepted Article
This article is protected by copyright. All rights reserved

Accepted Article
This article is protected by copyright. All rights reserved networks (SAN) is shown in comparison to the established Gradient Tree Boosting (GB) method and standard deviation is mentioned for each classification result.  genotype for which data was tested with Kolmogorov-Smirnov-Test for normal distribution with p ≤ 0.05 and aTukey's range test was applied with α = 0.05.

Accepted Article
This article is protected by copyright. All rights reserved  For each genotype, wavelengths with a weighting of more than 5 % are highlighted and standard deviation is indicated. Wavelengths from 262 to 291 nm were