Münstererc, Klaus Kammhuberd, Oliver Hensela, Stuart O.J

Computers and Electronics in Agriculture 175 (2020) 105547

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers In-process investigation of the dynamics in drying behavior and quality development of hops using visual and environmental sensors combined with T chemometrics ⁎ Barbara Sturma,b, , Sharvari Rauta, Boris Kuliga, Jakob Münstererc, Klaus Kammhuberd, Oliver Hensela, Stuart O.J. Crichtona a Department of Agricultural and Biosystems Engineering, University of Kassel, Witzenhausen, Germany b School of Environmental and Life Sciences, Newcastle University, Newcastle upon Tyne, UK c Bavarian State Research Center for Agriculture, Hop Competence Centre Production Technologies, Wolnzach, Germany d Bavarian State Research Center for Agriculture, Hop Research Centre Hüll, Wolnzach, Germany

ARTICLE INFO ABSTRACT

Keywords: Hops are a key ingredient for beer brewing due to their role in preservation, the creation of foam characteristics, Hyperspectral imaging the bitterness and aroma of the beers. Drying significantly impacts on the composition of hops which directly Thermography affects the brewing quality of beers. Therefore, it is pivotal to understand the changes during the drying process Environmental measurement to optimize the process with the central aim of improving product quality and process performance. Hops of the Chemometrics variety Mandarina Bavaria were dried at 65 °C and 70 °C with an air velocity of 0.35 m/s. Bulk weights in- Sensor data fusion vestigated were 12, 20 and 40 kg/m2 respectively. Drying times were 105, 135, and 195 and 215 min, respectively. Drying characteristics showed a unique development, very likely due to the distinct physiology of hop cones (spindle, bracteole, bract, lupilin glands). Color changes depended strongly on the bulk weight and resulting bulk thickness (ΔE 9.5 (12 kg), 13 (20 kg), 18 (40 kg)) whilst α and ß acid contents were not affected by the drying conditions (full retention in all cases). The research demonstrated that specific air mass flow is critical for the quality of the final product, as well as the processing time required. Three types of visual sensors were integrated into the system, namely Vis-VNIR hyperspectral and RGB camera, as well as a pyrometer, to facilitate continuous in-process measurement. This enabled the dynamic characterization of the drying behavior of hops. Chemometric investigations into the prediction of moisture and chromatic information, as well as selected chemical components with full and a reduced wavelength set, were conducted. Moisture content prediction was shown to be feasible (r2 = 0.94, RMSE = 0.2) for the test set using 8 wavelengths. CIELAB a* prediction was also seen to be feasible (r2 = 0.75, RMSE = 3.75), alongside CIELAB b* prediction (r2 = 0.52 and RMSE = 2.66). Future work will involve possible ways to improve the current predictive models.

1. Introduction Historically, the hop markets were dominated by bitter and high α acid hops, accounting for 75–80% in the US market, for example. According Hop cultivation has a long tradition in Europe, with first mentions to the Barth report (Barth, 2018) in 2017, however, the market shares dating from around 1000 years ago (Hirschfelder and Trummer, 2016). between bitter and aroma hops in the US and Germany were 23.1 and Initially, hops were primarily used to increase the microbiological 76.9% and 56.7 and 43.3% respectively. This shows a clear shift in stability of the beers which helped increase the shelf-life and allowed market demands towards craft beer production. In this context, it is for long distance transport. Over time, hops have increasingly been worth noting that the specific hop demand in craft beer brewing is used for other purposes, such as creation of specific foam character- significantly increased, which is partially due to the addition of the cold istics, bitterness and aroma of beers (Rybáček, 1991; Hirschfelder and hopping process during the fermentation stage. Trummer, 2016). Whilst bitterness is still the most important motiva- Besides the variety, the optimum harvesting date depends on a tion for selecting specific types of hops, aroma and flavor hops have multitude of environmental factors, such as environmental tempera- received new attention due to recent trends in craft beer manufacturing. ture, precipitation, light, location, soil quality, and exposition.

⁎ Corresponding author at: Department of Agricultural and Biosystems Engineering, University of Kassel, Witzenhausen, Germany. E-mail address: [email protected] (B. Sturm). https://doi.org/10.1016/j.compag.2020.105547 Received 22 October 2019; Received in revised form 15 May 2020; Accepted 30 May 2020 0168-1699/ © 2020 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/). B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

According to Münsterer (2017a, b), the time frame between cone for- optimization potentials for the process (Aral and Bese, 2016; Ashtiani mation and harvest maturity is 4 weeks (minimum 25 days) for α hops, et al., 2018; Fatouh et al., 2006; Kaya and Orhan, 2009; Münsterer, 5 weeks (minimum 30 days) for high α hops and 5 weeks (minimum 2015; Orphanides et al., 2017; Sturm et al., 2012; Hermanek et al., 30 days) for flavor hops. At optimum harvest maturity, cone color 2016; Münsterer, 2016). In practice, air humidity regularly reaches changes from dark green to light green and yellowish. Münsterer 100% at the start of the drying process, which has direct negative ef- (2017a, b) further found that the dry matter content is a good indicator fects on the quality of the hop cones exposed to these conditions. for the determination of maturity. The timing of harvest has a sig- Firstly, the water which is transported to the surface of the cones cannot nificant impact on both the overall oil content and its composition. be removed from the surface, which leads to color degradation. Sec- Generally, the later the harvesting date the more pronounced the aroma ondly, these cones are highly likely to be colder than the surrounding expression. However, towards the end of the ripening stage, hops in- air, which in turn leads to a re-condensation of a proportion of the creasingly develop an oniony and garlicy aroma due to an increase in water in the air which further increases discoloration (Münsterer, 2015; sulfur compounds (Cocuzza et al., 2013; Lutz et al., 2009; Münsterer, Sturm et al., 2016). The knowledge of product temperature throughout 2017a). the process has also proven to provide valuable information for the Fresh hop cones are highly susceptible to rapid degradation and optimization of the process and has been linked to the development of have to be processed within 4–6 h after harvesting to retain optimum discoloration of the product and transition from one drying phase to quality (LfL, 2019). The most commonly found preservation method for another for products such as apples (Fiorentini et al., 2015; Sturm et al., hops is convection drying during which the fresh hop cones are dried 2014). from between 78 and 84% (LfL, 2019) to 10% desired final moisture Another aspect heavily impacting the final product quality, the content, usually in multi-stage kiln driers. overall energy demand and efficiency is the point in time when the Bitterness is primarily created by non-volatile components, such as drying process is terminated. This is, to date, largely a result of the α and ß acids, which are secondary metabolites of the hops and are experience of the processor. However, a uniform result cannot ne- responsible for the bitterness of beers (De Keukeleire, 2000; Nagel et al., cessarily be reached (Hermanek et al., 2018), which also has direct 2008). They mainly accumulate in the lupilin glands which lie between consequences for the subsequent conditioning phase (Münsterer, 2006, the braces of the cones (Priest and Stewart, 2006) and on the underside 2012; Rybka et al., 2019). Due to their structure, the leaves of the hop of young leaves (Kavalier et al., 2011). During storage, α acids in cones dry significantly quicker than the spindle (Hermanek et al., particular are susceptible to degradation. These changes can be ex- 2018), leading to a product which is not in an equilibrium state. pressed using the Hop Storage Index which was established as a para- Commonly, the leaves have a final moisture content of 6–8%, while the meter for the evaluation of the freshness of hops and hop pellets. spindle remains at 12% (Münsterer, 2006, 2012). Thus, conditioning is Cocuzza et al. (2013) showed in their research that the harvesting date an inevitable further processing step. also has a significant influence on the Hop Storage Index of fresh hops. Currently, temperature and humidity sensors are usually placed Forster and Beck (1985) found that α acid losses are between 3.5 and within the hop stack during drying. This method works well as a local 10% during drying, depending on the overall quality of the hops. Zeisig (in spatial terms) non-destructive method to determine hop moisture (1970) showed that an increase of loss by 5% occurred when the drying content. However, large variations in moisture content may exist within temperature was increased from 55 °C to 65 °C at constant air velocity the overall hop stack (Münsterer, 2006). This method also ignores any and bulk height. This is potentially related to a lack of sufficient air potential to continuously monitor the dynamics of changes within the throughput in the transportation of the evaporated water and the bulk product which go beyond water content. In recent years, development height within the dryer. of dryers has been strongly focused on the inclusion of product relevant Like bitter components, aromatic components are located in the information, with the aim of the realization of smart drying systems lupilin glands. Essential oils make up 0.5–3% of the dried hop weight which not only measure these changes non-invasively but also use these (Sharpe and Laws, 1986; Steinhaus and Schieberle, 2000). In contrast to as input variables for process control (Martynenko, 2017; Martynenko the bitter components, which are mainly affected by degradation due to and Misra, 2020; Martynenko and Sturm, 2019; Sturm, 2018; Sturm high water activity levels over an extended period of time, aromatic et al., 2018). To gain a deeper understanding of the dynamic processes compounds and essential oils are highly volatile and susceptible to within the drying of plant based products, such as hops, information degradation before, during and after the drying process (Aberl, 2016). from non-invasive measurement devices such as color cameras (CCD), During drying, these losses are mainly due to water vapor volatility spectroscopy and hyperspectral imaging (HSI), as well as thermo- (Münsterer, 2017b). More than 400 volatile components are known in graphy, can be combined with product related information from che- hops. The highest percentage of essential oils are carbon hydrates mical and textural analysis. With machine learning methods becoming which are responsible for 69% of odor activity (Steinhaus and more sophisticated, computing power increasing rapidly and prices Schieberle, 2000). Reactions of aromatic components with oxygen are falling, new ways of dealing with data and developing drying systems responsible for around 34% of the odor activity (Steinhaus and based on the principles of Quality by Design are emerging, i.e. using the Schieberle, 2000). Sulphuric compounds such as 4-MMP are highly raw material characteristics and change dynamics as basis for control aroma active despite them being only present in very low concentra- development and execution (Martynenko, 2017; Sturm et al., 2018). tions. Non-destructively monitoring chromaticity and moisture content, as Myrcene and linalool are seen as key substances for the overall well as the changes in chemical composition, over the entire hop stack aroma (Steinhaus and Schieberle, 2000). Myrcene’s smell is associated could bring great benefits to the hop drying industry. HSI offers a po- with geranium while linalool is described as sweet, flowery and citrusy. tential single solution to investigate both chromatic and chemometric The taste threshold for aromatic compounds is in the region of few µg/l changes in the hops continuously during drying (Crichton et al., 2016; beer (Kurzweiler et al., 2013). According to Aberl (2016), drying of Crichton et al., 2018; Pu et al., 2015; Pu and Sun, 2015; Su et al., 2018; hops results in a loss of hop oils of 30–40%. Myrcene is highly volatile Wu et al., 2012, 2013), while color imaging and thermography can at temperatures usual for hop drying operations. Kaltner (2000) and deliver further valuable information on the changes a product under- Kammhuber (2018) reported losses between 25 and 30 % during goes during drying (Martynenko, 2006; Martynenko and Kudra, 2015; drying, however it was also pointed out that hop aroma quality tended Martynenko, 2017; Sturm, 2018; Sturm et al., 2014). Several previous to increase with a decrease in myrcene concentration. studies have also shown, that HSI in the visible and near infrared areas Numerous previous studies on drying of sensitive agricultural pro- is suitable for the determination of phenolic and other bioactive com- ducts, including hops, have shown that the impact of air velocity, and pounds in plant-based products (Gaston et al., 2010; Huck, 2015; Yang thus specific air mass flow, needs to be considered when evaluating et al., 2015). Combined with other process data and their fusion, a

2 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547 better understanding of the dynamics of quality changes within hops measurements an additional 24 samples were taken at 0.15 m/s. Air during drying can be gained. Based on this information, optimized temperature, product temperature, relative humidity, color, and spec- process settings and dryer set-ups can be realized. tral characteristics were measured. All investigations were conducted in In this study, for the first time, a full array of visual sensors was triplicate. integrated into a pilot scale drying system to investigate the dynamic Samples were taken at the start of the drying process, after 15, 25, changes of selected quality criteria, namely color, Hop Storage Index, α 35, 45, 75 min and then every 30 min thereafter until the samples and ß acids, product and air temperature, as well as air humidity, reached storage moisture content. RGB and hyperspectral images were throughout the drying process, with different bulk weights and drying taken directly before the samples were removed from the dryer. Image temperatures. These criteria were then fused with the results of color capture at these set time intervals across the 4 different drying condi- and HSI to predict water content and distribution thereof. Furthermore, tions led to the capture of 134 hyperspectral images and 134 corre- selected chemical components, i.e. myrcen, ß-Ocimen a and 1-octen-3- sponding calibrated RGB images. Three hyperspectral images were ol, were chemometrically linked to HSI. Conclusions for optimized hop subsequently discarded due to capture error, so all subsequent analysis drying were drawn from the combined data. The study is a proof of was performed upon 131 images. The test train split was carried out on concept for integration of non-invasive measurement systems into fully a 2:1 basis, resulting in a 88:43 basic train test sample split. The cali- operational small-scale commercial driers and, thus, lies the foundation bration of the system is detailed in Sections 2.2.3 and 2.2.4 and within for the realization of intelligent drying systems for sensitive agricultural Crichton et al. (2016). produce. Sampling was conducted on the surface of the bulk underneath the HSI system and the amount removed was replenished from other areas 2. Materials and methods of the surface (1 m2). For sampling, a beaker of 2000 ml capacity was

used and for each sample the equivalent of 1600 ml (mleq) was re- Trials were conducted at the Research Station Hüll, Hallertau, moved. A small proportion was then used for dry matter assessment

Germany on ﬁve consecutive days in September 2015 to minimize (24 h, 105 °C) (AOAC, 1995) whilst the rest (ca. 1200 mleq) was vacuum variations in raw material quality due to harvest dates (Cocuzza et al., packed and immediately frozen at −20 °C for chemical analysis. All 2013). tests were conducted in triplicate.

2.1. Materials 2.2.3. RGB imaging The RGB imaging system was directly integrated into the drying Hops of the variety Mandarina Bavaria were investigated. chamber (Fig. 1). To prevent overheating to the unit, a permanent Mandarina Bavaria is a hybrid between Cascade and a male Hüll pressurized air blast was provided, and the camera temperature con- breeding strain (Lutz et al., 2013), and is described as less hop-typical, tinuously measured using a thermocouple. The maximum device tem- but with a pronounced fruity and exotic aroma of mandarin and orange perature was 31 °C. A color DFK 72AUC02-F (The Imaging Source, (Lutz et al., 2015). Germany) of a resolution of 1944*2592 (H*W) was used. Illumination All raw materials originated from the same hop garden close to the was provided by 2 Philips fluorescent light sources (Philips, Nether- research station. Hops were harvested one hour before drying tests were lands) of CCT 6500 K and luminance of 2100 Lumens each. conducted. Hop cones were separated from the vines using a commer- Color calibration of the RGB camera using polynomial color cor- cial de-vining machine (Hopfenpflückmaschine Wolf WHE 220), rection (Finlayson et al., 2015) was carried out. An X-Rite Color weighed and then transferred into a pre-heated dryer and drying tests Checker Classic 24 patch chart (X-Rite Inc., Michigan, USA) was imaged were started immediately. using a set exposure under imaging conditions. Colorimetric data of each patch was then measured using a calibrated spectrometer (Q mini, 2.2. Methods RGB Photonics GmbH, Kelheim, Germany). Calculated XYZ values were then converted into CIELAB co-ordinates. 2.2.1. System set-up To carry out the calibrated CIELAB capture correctly the halogen A pilot plant scale dryer with a drying area of 1 m2 and an in- light source was turned off prior to RGB image capture (leaving only the tegrated heat recovery system was used as described by Hofmann et al. 6500 K fluorescent sources on). (2013). The system was modified by the introduction of an RGB camera, a hyperspectral imaging system and the respective lighting 2.2.4. HSI set-up, imaging, calibration and statistical analysis installations as described by Crichton et al (2016). To enable the imaging of hops during drying, a Specim V10E hy- The dryer was further equipped with an array of temperature, hu- perspectral camera (Specim Spectral Imaging Ltd., Finland) was placed midity and air flow sensors to assess the temporal and spatial conditions within the dryer inside a protective casing above the hop stack within the system. Combined temperature/humidity probes (Testo (Crichton et al., 2016). The camera was used in combination with a 174H, Testo, Lenzkirch, Germany) were installed and positioned at mirror translation unit and a 35 mm Schneider lens (Xenoplan 1.9/35, different filling heights within the bulks to measure the development of Schneider Optische Werke GmbH, Germany). Illumination was pro- temperatures and air humidity throughout the bulks. Product tem- vided using a halogen security lamp fitted with a 600 W IP65 rated perature at the surface of the bulk was measured continuously using a halogen lamp, which was attached on the interior wall of the dryer and pyrometer (Bartec Type R300-123). All values, except of the combined focused at the region directly beneath the camera system (Fig. 1). The temperature/humidity probes, were transferred directly to a computer protective casing was constructed of steel, with a rockwool insulation and stored. lining of 30 mm thickness. The casing also had a viewing window covered with a clear acrylic window sheet (Fig. 1). This allowed the 2.2.2. Sampling camera to view the hops situated directly below. This protective casing Standard process settings used were 65 and 70 °C drying air tem- also provided ventilation to the camera during operation to prevent perature and 0.35 m/s air velocity. For the simulation of real condi- overheating. Temperature within the casing was measured con- tions, the bulk weight of the samples was used rather than the bulk tinuously using a thermocouple. The maximum temperature within the height. Three different bulk weights (12, 20 and 40 kg/m2) were in- casing was 27 °C. This protective casing was attached to the interior of vestigated at 65 °C, resulting in 25, 30 and 33 measurements across the an experimental dryer above the top drying tray as outlined by triplicate experimental run. Furthermore, at 70 °C for 0.35 m/s 22 Hofmann et al. (2013). measurements were conducted and for the purpose of hyperspectral The lens was of a manual focus type and was focused at a specific

3 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Fig. 1. Location of imaging box in chamber, alongside CCD camera and illumination. hop depth, which led to an increase in blur towards the lowest hop hyperspectral imagining. It attempts to uncover whether any of the depths. Hyperspectral cubes were recorded using the Specim retrieved wavelengths vary in a linear manner with respect to the SpectralDAQ software (Specim Spectral Imaging Ltd., Finland), with all metric investigated. Previous investigations have looked into moisture processing occurring within Matlab using software previously utilized prediction (ElMasry et al., 2007; Huang et al., 2014; Pu et al., 2015), in Crichton et al. (2015) and Crichton et al. (2016). Noise removal was chromaticity (Iqbal et al., 2013; Wu et al., 2012), maturity (Rajkumar carried out in the same manner as presented in Crichton et al. (2015). et al., 2012) and pH levels (He et al., 2014) amongst many other me- Due to the nature of product shrinkage during drying, the height of trics. Monte-Carlo uninformative variable elimination (MCUVE) the hops changed which resulted in changes in the light intensity over (Centner et al., 1996) and the competitive adaptive reweighted sam- the area. To take these changes into account, the measurement of the pling (CARS) (Li et al., 2009) are different wavelength reduction incident illumination at a number of different heights was carried out. methods, noted from here onward as MCUVE–PLSR and CARS–PLSR, The resulting illumination field was smoothed with a small window respectively. In some cases, they are good alternatives to the PLSR operator to remove the influence of the non-perfectly uniform surface of prediction model (Amigo et al., 2013). the white reference tile. This height dependent illumination data was For the statistical investigations, the measured hops were split into then utilized in combination with hop height measurements carried out the calibration and test sets on a random 2:1 ratio, on the basis of re- at each sample period. Fig. 2 below illustrates the increase in light in- plicate number. Furthermore, each model was cross validated five-fold tensity at the 669 nm wavelength as drying occurs. Each of the 3 col- to ensure all combinations of calibration and test sets were tested. umns from left to right shows that as drying occurs, and the depth of hop bed decreases the intensity variation across the viewing field 2.2.5. Modelling of drying curves changes. This increase in intensity as the hop moves further from the Weight and color changes were measured intermittently, while reference point occurs due to the focal point of the light source. The process and product temperature, as well as relative air humidity, were upper row illustrates the reference white image for each of the hop bed measured continuously. To ensure a meaningful correlation between depths, with the lower plots illustrating the intensity map across the those values, the modelling of the drying curves based on the measured white reference in the picture. Images were then processed with the values was necessary. Two drying models, the Page model (Doymaz, corresponding white reference image to produce the reflectance image 2004) and the Henderson & Pabis model (Doymaz, 2004), were used to on a pixel-by-pixel basis. fit the drying curves for the experimental moisture content of hops – The nature of the imaged hops and singular illumination direction Eqs. (1) and (2) below represent the Page and Henderson & Pabis Model led to challenges in image processing, which required further use of pre- respectively. processing methods. Both factors meant that, when hops across the n region of interest were analyzed, a great variation in spectral intensity MRKt=−exp( ) PageModel (1) ff existed. In order to counteract this, two di erent normalization pre- where K, t and n are the drying constant, drying time in min and the processing methods (Amigo et al., 2013) were used; Normalization and drying exponent respectively Standard Normal Variate (SNV). Using these two methods had the effect of minimizing the baseline variation between pixels related to hop MRKt=−aexp( ) Henderson&PabisModel (2) surface topology, as seen in Fig. 3 below. where a & k are drying constants and t is the drying time in min. fl After the retrieval of the average pre-processed re ectance spectra The moisture ratio MR [–] for hops was calculated using Eq. (3): from each image, Partial Least Squares Regression (PLSR) analysis was MM− performed upon the retrieved spectra against the measured moisture MR = e content, and CIELAB chromatic co-ordinates retrieved from the cali- MMoe− (3) brated RGB images. PLSR analysis was performed upon the normalized where M [gW/gDM] is the moisture content at any given time, M0 [gW/ and SNV data. PLSR is the most common method for investigating the gDM] is the initial moisture content and Me [gW/gDM] is the equilibrium feasibility of metric prediction in combination with spectroscopy and moisture content. As Me is relatively small compared to M and M0

4 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Fig. 2. An illustration of how the intensity (at 669 nm) and spatial distribution of illumination varies upon the hops as the hop height changes from a reference point at a) 5 cm, b) 10 cm and c) 15 cm.

(Doymaz, 2004), the moisture ratio was simplified to Eq. (4) experimental data for each of the two models. In addition, the basic requirements for the application of nonlinear regression in the form of M MR = residual analysis were examined. The following assumptions were M (4) o tested: Variance homogeneity, normal distribution of residuals and Non-linear regression analysis was performed using GraphPad Prism lack-of-fit of the fitted model. The independence of the measurements 2 2 8.0 software (GraphPad Software, 2018). Adjusted R (Radj), Root mean was ensured by the test design. square error (RMSE) and Akaike Information Criterion (AICc) were A comparison of fits between the two models was also performed used to evaluate the goodness of fit between the predicted and the using the AICc, rather than F test due to the assumption of non-nested

Fig. 3. Illustrating the variation in irradiance due to the non-uniform topology of hops in a) the original image, and b) the eﬀect of SNV pre-processing used to remove it.

5 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547 models. The comparison of fits not only provides a difference in the AICc but also the likelihood a certain model is right on a percentage scale. 2 2 Radj values rather than R values were used for evaluation as these values have been adjusted to the number of predictors in the model. AICc is a model selection criterion that estimates the most likely model for the selected data. These values are not always negative, however regression analysis can potentially lead to negative values of AICc, as is the case for the selected hop data (Burnham and Anderson, 2002). For non-linear regression, homoscedasticity or test for appropriate weighting (P value) assumes that ‘average distance from the curve is the same for all parts of the curve’ (GraphPad Software, 2018). Therefore, a 2 model is estimated fit for the data if the values for Radj and % likelihood are high, RMSE & AICc are low, and homoscedasticity (P value) > 0.05.

ff 2.2.6. Chemical analysis and determination of the hop storage index Fig. 4. Drying curves of hop samples dried with di erent bulk weights. For the chemical analysis, the frozen samples were freeze dried (LABONCO FreeZone 2.5 Plus). Determination of HSI, α and ß acids The ratios between spindle and body of the cone depends strongly was conducted according to the EBC Method 7.13 (EBC, 2007). A total on the variety in question, and the stage of development of the bulk and of 134 samples were analyzed. Samples were extracted with toluene the individual hop cones. This in turn has a significant impact on the followed by dilution of the extract in methanol and alkaline methanol. drying characteristics and the required conditioning time. Münsterer The diluted solution was measured spectroscopically at 275 nm, (2015) investigated 7 varieties to determine the ratio which was found 325 nm and 355 nm using a Thermo Scientific Evolution 201 photo to vary from 6.3 to 7.5 % (Hallertauer Magnum) to 10.0–12.0% (Hal- spectrometer. Hop Storage Index and α and ß acids’ contents were lertauer mfr.). Furthermore, as demonstrated in Fig. 5, size and shape of determined using the following Eqs. (5)–(7): individual hop cones can vary greatly. Thus, both the average size, as A275 well as the size distribution, of cones directly impacts on the bulk HopStorageInd = density and, thus, the fluid dynamic characteristics of the bulk. These in A325 (5) turn directly influence the drying behavior of the bulk. Consequently, α =−( 19.07·AAAF 275 + 73.79· 325 − 51.56· 335)· (6) bulk density varied from 153.9 kg/m3 to 166.7 kg/m3 for the trials included in the data analysis. β =+( 5.10·AAA 275 − 47.59· 325 + 55.57· 335)·F (7) As the trials were conducted under real conditions, i.e. the harvest For the tests presented F was 0,667. To make the results compar- has to be conducted irrespective of the weather, it was possible to assess able, the results for all water contents were normalized to 10% final the impact of weather condition (heavy rain at harvest (174.0 kg/m3), water content. morning dew (166.7 kg/m3), dry samples from a sunny day (160.0 kg/ For determination of aromatic compounds, samples were also freeze m3)) for the samples with a bulk weight of 40 kg/m2. It was shown that dried and afterwards further processed using a statistic headspace- these differences significantly influence both the duration of the drying method (Agilent Headspace Sampler 7697 A) developed inhouse with process with the associated air conditions (Fig. 6) and the achievable the settings purge gas: He 5.0; loop volume: 3 ml; transfer line dia- product quality in terms of color retention (Fig. 7). meter: 0.53 mm; transfer line length: 1.00 m; incubation time: 60 min; incubation temperature: 80 °C; injection time: 30 s. 3.2. Model comparison for drying behavior Due to a malfunction of the freeze dryer, which only became visible after the chemical analysis, only 73 out of 134 samples could be used Due to the nature of the experimental and system set-up, it was for the model development. The test train split between the sample was necessary to model the development of moisture content with the goal 49:24 of conducting a clear observation of product temperature and air humidity development as a function of moisture content. 3. Results and discussion A comparison of two models (Page and Henderson & Pabis) was conducted for hops dried at 65 °C and at different bulk weights (12, 20 3.1. Drying behavior & 40 kg) and for 20 kg bulk weight dried at 70 °C and 0.35 m/s (Table 1). In both scenarios, R2 , RMSE, AICc, Homoscedasticity and % For all drying conditions there is a characteristic development of adj likelihood were used to estimate the goodness of fit. Higher R2 and drying velocity with a rapid loss of moisture in the initial phase, fol- adj lower RMSE values were obtained for the Page model. However, the lowed by a phase of reduced drying velocity and another phase of in- difference between the R2 and RMSE values obtained from the increased moisture loss thereafter before eventually transitioning into a adj dependent models is relatively close to a conclusive prediction. A second phase of reduced drying velocity (Fig. 4). This could be due to considerable difference was observed in the AICc and homoscedasticity the structure of the hop cones which consist of a spindle, bracteole, values with the values ranging from −50.2 to 88.9 and 0.1634 to bract and lupilin glands (Fig. 5). It is very likely that the spindle only 0.4302, respectively. The lower the AICc values, the better the con- starts drying once the leaves have reached a certain moisture content formity of the model (Raut et al, 2019). Furthermore, through the and, thus, develop a sufficiently high suction tension to draw the water comparison of the fits, the % likelihood is considerably high for the out of the spindle. Page model. Thus, based on the values obtained, the Page Model was As already described, at the end of the drying process (at 10%) the determined to be the more representative model for hops drying. leaves are over dried to a moisture content of 6–8%, whilst the spindle still has a moisture content of around 12% (Fig. 5). This fact also supports the theory regarding suction tension as a driving force for 3.3. Color changes moisture removal and, thus, the very specific development of the drying curve. Fig. 6 depicts the total color differences ΔE for the samples dried at

6 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Fig. 5. Exemplary hop cones of diﬀerent sizes, cross section through a hop cone and individual leaves and spindles.

different bulk weights and temperatures. An increase in bulk weight impacted the color changes of the different samples significantly. The changes can be represented through linear regression with a high degree of accuracy. The differences between the bulk weights can be directly correlated to the decrease of water removal potential of a given amount of air with increased bulk thickness (Münsterer, 2015; Sturm et al., 2016). This is of particular importance during the first phase of drying. However, an increase of temperature at identical bulk weight, and thus a reduction of specific air flow, leads to a lower initial color change but does not impact on the final value. Due to the damage in the bulk dried at 65 °C with a bulk weight of 40 kg after harvest in rainy conditions, the batch had to be discarded as the quality was severely affected and the values were removed from the above comparison. For explanation, Fig. 7 depicts the total color changes ΔE for the different weather conditions for 40 kg bulk weight, while Fig. 8 shows Fig. 6. Total color difference ΔE for samples dried at 60 °C with 12 kg/m2, the rH of the air above the bulk and product surface temperature. 20 kg/m2, and 40 kg/m2 bulk weight and at 70 °C with 20 kg/m2 bulk weight. It is evident that changes in color significantly depended on the weather conditions. Moderately high moisture (morning dew) did not have an impact on final color changes. However, rain wet hop cones deteriorated very quickly with an overall ΔE almost twice as high as the other samples. Air exchange evidently was not sufficient to remove the water quickly enough during drying in the rain wet batch, leading to a rapid increase in browning. Furthermore, within the first 45 min of the process, relative humidity over the bulk was continuously at 100%, making drying impossible and the recondensation of water transported through the bulk very likely due to the wet bulb temperature difference (Fig. 8). It can be clearly seen that, when surface water was present (rain and dew), the initial rH of the air above the bulk was at 100% while the bulk surface temperature remained almost constant. As soon as the rH decreased below 100% and, thus, surface water was removed com- pletely from the hop cones, product temperature increased according to the resulting bulb temperature. These results are in line with the find- ings of Sturm et al. (2012, 2014) for apples. Thus, showing a direct relationship between environmental conditions and thermal product Fig. 7. Impact of weather conditions on color changes (65 °C, 40 kg). characteristics. The somewhat erratic development of both the rH of air

7 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Table 1 Comparison of models for goodness of ﬁt evaluation for an air velocity of 0.35 m/s.

2 Temperature Bulk Weight of Radj RMSE AICc P Value % Likelihood hops (kg) Page Henderson & Page Henderson & Page Henderson & Page Henderson & Page Henderson & Pabis Pabis Pabis Pabis Pabis

65 °C 12 0.987 0.987 0.035 0.035 −50.33 −50.20 0.291 0.193 51.62 48.38 20 0.997 0.997 0.015 0.018 −65.43 −62.34 0.354 0.179 82.44 17.56 40 0.998 0.968 0.012 0.065 −88.90 −51.72 0.163 0.430 99.99 0.01 70 °C 20 0.972 0.972 0.053 0.054 −140.30 −140.00 0.450 0.434 54.42 45.58 and the product temperature for the dry condition might be due to problem with overfitting to the calibration set, which may have been environmental conditions (e.g. incoming air humidity changes) and caused by any variation in initial hop quality. differences in size distribution in comparison to the other tests. How- CIELAB a* prediction can be seen to be better serviced using the ever, in this trial the data between the two sensors is also clearly cor- normalized spectra with the MCUVE-PLS model with test set results of related. r2 = 0.81 and RMSE = 2.79 using only 5 wavelengths ([620, 665, 902, 912, 1010] nm). Whilst for CIELAB b* prediction, use of the SNV spectra using the CARS-PLS method led to the best performing model 3.4. Hyperspectral prediction of water content, chromaticity and aromatic with test set performance of r2 = 0.73 and RMSE = 1.94. This was components achieved using 13 wavelengths. Moisture content and chromaticity prediction are very useful The first stage of results is that of the retrieved average spectra for a quality metrics to allow the monitoring of hop changes during the given batch of hops during the drying process. Fig. 9 illustrates the drying process. However, hops, in a chemical sense, are also very variation in normalized reflectance for the 12 kg hop batch dried at complex. The resulting chemical content of hops after drying de- 65 °C with an air speed of 0.35 m/s. termines the characteristics of the beer brewed with them. As such, Fig. 9a illustrates that the most usable wavelengths were in the prediction of specific chemical contents during drying, whether to keep region from 500 to 1010 nm. The trials were conducted under quasi- changes within a set limit, or to achieve a specific level of chemical production conditions using halogen lights which have a low irradiance components after drying is possible. Upon analysis of chemical contents below 500 nm and as can be seen from Fig. 9a the utilization of this during drying, three chemicals which uniformly increased or decreased region would introduce a significant level of noise to a small data set during drying were noted and chosen for algorithm development, and be detrimental to model performance. The low light emission issue namely Myrcene, β-Ocimen and 1-octen-3-ol The same calibration to is further enhanced by the lower SNR of the V10E sensor in this range. test set split method was utilized for the analysis of these chemicals. It can also be noted that the retrieved reflectance information is greater However, due to errors during chemical analysis, the overall dataset above 700 nm than in the region between 500 nm and 700 nm. With was reduced from 134 measurements to 73 (49:24 split). this is mind, the PLSR analysis was undertaken in the 500–1010 nm The best predictive models for all 3 of these chemicals were gained range. The green peak present at roughly 550 nm (Fig. 9a), known as using the normalized MCUVE-PLS models. Myrcene prediction on the the “green gap” (Gundlach et al., 2009) between the absorption peaks test set achieved r2 values of 0.83, 0.73 and 0.64, respectively. The of chlorophyll a and b at 430 and 664 nm and 460 and 647 nm re- RMSE errors can also be seen to be within acceptable margins, when spectively, is reduced in size as drying occurs. This indicates a change in considering the variation of these chemical components. anthocyanin concentration, a sub-group of flavonoids, which have an The current predictive models are at a usable level of performance. absorption peak in the region of 530 nm (Kim et al., 2008). Also of great However, there is room for improvement. A number of future changes interest is how the peaks and troughs present in the 700–1010 nm re- will include altering the illumination utilized for hyperspectral imaging gion are removed as drying occurs. This development might be due to to an illumination which includes a greater output within the blue- the change in concentration of other polyphenolic compounds. These green region of electromagnetic space (400–600 nm) in an attempt to include flavanols, flavan-3-ols, phenolic carboxylic acids and others as equalize it. This will give more information for both moisture content prenylflavonoids and have both a hydroxyl group (O–H) and an aro- and chromatic prediction models. However, due to the low SNR of the matic hydrocarbon (C–H). The changes could thus be due to the third V10E sensor in the range between 400 and 500 nm the noise to signal overtone of the third overtone of OH and the fourth overtone of CH in ratio might still be high. Investigations into the prediction of other the region of 740 and 760 nm (Amodio et al., 2017; Clement et al., chemical components would also be of great interest, in combination 2008; Dusek et al., 2016; Kokaly and Skidmore, 2015; Steinhaus and with the illumination addition. Schieberle, 2000). The trough at roughly 970 nm is known to be directly related to the stretching of the H–O–H bonds (Clement et al., 2008) within water molecules. 3.5. Bitter components and hop storage index The results of the PLSR (Table 3), MCUVE-PLSR (Centner et al., 1996)(Table 3) and CARS-PLSR (Li et al., 2009)(Table 4) models de- Table 5 displays the hop storage index, α and ß acid contents for veloped show that basic prediction of hop moisture content, CIELAB samples dried with 12, 20 and 40 kg/m2 bulk weight. The results show chromaticity and chemical estimation can indeed be carried out. that the drying process has no significant influence on the acid contents. Moisture content estimation can be achieved with a lower error The Hop Storage Index as a measure for freshness of the samples is using the normalized retrieved spectra during drying. This produces a low, as expected, because the samples were harvested at optimum ri- model which uses only 6 wavelengths and achieves a test set r2 = 0.95 pening stage and processed within 60 min of harvest. Values between and RMSE = 0.24 with the MCUVE-PLS model. This level of perfor- 0.21 and 0.24 indicate that the samples used were of constant quality mance can be seen to be useful up until the last stages of drying. and not influenced by the picking date. However, the consistent RMSE performance between the calibration The results indicate that, for the components investigated, the and test sets for the reduced wavelength set shows the importance of drying conditions are not critical. This in turn might help to explain the selected wavelengths. The large variation between the full wave- why farmers have not struggled to meet the requirements on quality of length set PLS models, shown in Table 2, does, however, point to a bitter hops in terms of valuable components, as the α and ß acids’

8 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

contents seem to be highly thermostable. However, it is to be expected that this is not true for volatile components.

3.6. General discussion

Based on the extensive tests performed in the modified pilot plant system which was equipped with an array of several environmental and optical sensors, it was possible to gain a deeper understanding of dynamic changes in hop cones during convective drying. Impacts of varying process temperature, specific air mass flow, harvesting conditions, bulk density and size distribution were observed on both drying behavior and quality changes for hops. In terms of drying behavior and color retention, best results were achieved for bulks harvested on a dry day and the lowest bulk weights which in this case translates to the highest specific air throughput. These results are in line with previous research for hops (Rybka et al., 2017, 2019; LfL, 2019). Due to the distinctive physical characteristics of hops, i.e. leaves and the spindle, the final moisture content varies different parts of the cone. Similar results have been reported for other aromatic and medicinal plants such as Lemon Balm (Cuervo-Andrade, 2011). Wavelengths [501,508,516,526,899,915,922,961,966,977,984,995,1000,1005] [504,511,516,542,683,779,936,954,977,980,987,995,997,1005] [502,507,522,537,597,706,807,878,935,946,997,1000] Bitter components after drying were not impacted by the drying conditions, bulk weights and weather conditions which is in line with 7 6 the research of Forster and Beck (1985) who showed that losses during drying are very low. However, Zeisig (1970) stated that an increase in 4 temperature from 55 °C to 65 °C leads to an increase in losses which RMSE could not be confirmed by this research as the levels stayed constant. Hop Storage Index was also not impacted by the drying. Again, these results are somewhat in disagreement with previous literature 2 Test R 0.680.590.51 5,33 × 10 6.75 × 10 6x10 (Zeisig, 1970; Doe and Menary, 1979). This might be due to the further development of hop drying systems which allow for gentler drying since

7 6 4 these studies were conducted, as well as the diﬀerence in variety and origin of the raw material (LfL, 2019; Lutz et al., 2016; Mejzr and Hanousek, 2007; Münsterer, 2017a, b).

RMSE HSI was successfully implemented to detect both time series moisture content and chromaticity in a real, scaled up, drying process. No eﬀect of temperature, air humidity and shrinkage on the model 2 Reduced Wavelength Set Calibration R 0.850.850.71 3.18 × 10 3.33 × 10 4.96 × 10 performance were found in this study despite clear evidence from literature towards the contrary (Yao et al., 2013; Romano et al., 2011).

7 6 4 This, amongst others, might be due to the low temperature range (Δ∂ = 5 °C) in which the investigations were conducted. These aspects need to be further investigated in future studies. Due to problems with the

RMSE freeze-drying system it was not possible to investigate time series changes in aromatic components. However, it was possible to princi- pally build sound models for the prediction of the aromatic component 2 Test R 0.750.430.41 3.83 × 10 6.67 × 10 8.81 × 10 contents in question. For all investigations conducted models for reduced wavelength sets performed similar to the full wavelength set.

6 5 3 Nevertheless, as demonstrated by Kamruzzaman et al. (2016) and Su et al. (2020) non-linear methods might show more accurate and robust performance than linear approaches such as PLSR in food quality

RMSE analysis. Moreover, recent studies have demonstrated that in the development of HSI algorithms it is imperative to statistically compare the performance of the developed algorithm with the actual performance of 2 R Calibration the comparison method (Shrestha et al., 2020). Thus, future work needs to pay close attention to finding the most appropriate modelling approaches including locally weighted partial least square regression (LWPLSR) (Su and Sun, 2017; Su et al., 2018) and deep convolutional neural networks (CNN). Furthermore, the accuracy of the underlying methods for chemometric correlation and possible bias caused by ff -Ocimen 1.00 4.45 × 10 -Ocimen 1.00 5,38E + 05 0.55temperature 5.98E + 06 0.85di 3.49E + 06erences 0.60within 5.58E + 06the [504,508,526,542,574,796,873,896,927,957,972,980,990,1003] physical system of HSI or resulting β 1-octen-3-ol 1.00 3.19 × 10 CIELAB a*CIELAB b*Myrcene 0.99 0.98 0.73 1.00 0.61 3.26 × 10 0.67 0.63 3.76 2.29 0.87 0.82 2.51 1.62 0.64 0.57 4.14 2.46 [517,543,596,645,694,716,731,896,917,956,975,995] [501,516,536,586,659,760,768,917,935,943,957,967,975,1002,1008] CIELAB a*CIELAB b*Myrceneβ 1-octen-3-ol 1.00 0.98 0.45 1.00 0.61 1.00 3.19E + 06 3.59E + 03 0.68 0.74 0.56 0.35 3.97 4.33Emulti + 07 2.48 8.77E + 04 spectral 0.84 0.75 0.84 3.43E + 0.75 07 4.56Esystems + 04 2.85 1.97 0.60 0.45 in 5.37E +which 07 6.73E + 04 0.64 0.41 [504,510,526,686,886,896,922,943,957,977,984,995,1005] they [501,530,537,597,714,825,935,959,971,977,985,998,1000] 3.87 2.94 are integrated [508,557,757,766,807,807,839,953,971,985,993] [502,516,536,559,658,686,738,753,814,855,886,943,975,977,993,1000] need to be verified.

4. Conclusions

The present study is a proof of concept of integration of non-invasive visual sensors in drying systems beyond the laboratory scale. By Pre-processing Metric/ Chemical Full Wavelength Set Norm. M.C. 1.00 0.05 0.97 0.20 0.95 0.24 0.93 0.30 [501,560,684,741,798,847,914,936 SNV M.C. 1.00 0.06 0.96 0.22 0.96 0.22 0.93 0.29 [501,557,684,703,799,844,915,969]

Table 2 PLSR model results for full and reduced wavelength sets for prediction of metrics and chemical contents. combination of the information gathered through a multitude of

9 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Fig. 8. Relative humidity rH and product temperature ϑP as a function of time (65 °C, 40 kg).

Fig. 9. Variation in normalized hop reﬂectance across a) 500–1010 nm, and b) 700–1010 nm (exemplary for 65 °C, 12 kg/m2).

Table3 MCUVE-PLSR model results for reduced wavelength sets for prediction of metrics and chemical contents.

Pre-processing Metric/ Chemical Calibration Test Wavelengths

R2 RMSE R2 RMSE

Norm. M.C. 0.96 0.23 0.95 0.24 [658,664,839,901,920,1010] CIELAB a* 0.81 3.25 0.81 2,79 [620,665,902,912,1010] CIELAB b* 0.74 2.23 0.63 1.90 [504,522,588,827,865,901,920,1010] Myrcene 0.89 2.57 × 104 0.83 3.83 × 107 [505,554,575,605,690,883,931,941,946,971,1010] β -Ocimen 0.76 4.12 × 106 0.73 4.61 × 106 [507,557,651,659,9220,946,951,1010] 1-octen-3-ol 0.67 5.78 × 104 0.64 5.24 × 104 [517,536,575,580,659,1010] SNV M.C. 0.92 0.33 0.93 0.29 [611,703,902,1010] CIELAB a* 0.87 2.59 0.55 4.18 [507,574,589,605,634,695] CIELAB b* 0.59 2.60 0.48 2.73 [569,750,769,817,925,1010] Myrcene 0.82 3.33 × 107 0.82 4.84 × 107 [536,565,592,897,982,1010] β -Ocimen 0.79 3.92 × 106 0.66 6.68 × 106 [533,583,597,659,684,1010] 1-octen-3-ol 0.74 4.18 × 104 0.63 6.96 × 104 [546,594,651,670,695,700,719,734,819,910] different methods (sensor and data fusion), it was possible to assess the weight and temperature, harvesting conditions and specific air mass dynamic impacts of drying conditions on hop quality in an integrated flow have a significant influence on both the drying time and the color way. changes of hops during drying at identical conditions. Thus, in practice, The results of the research undertaken show that, besides bulk it is important to find the optimum compromise between process

10 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Table 4 CARS-PLSR model results for reduced wavelength sets for prediction of metrics and chemical contents.

Pre-processing Metric/ Chemical Calibration Test Wavelengths

R2 RMSE R2 RMSE

Norm. M.C. 0.97 0.21 0.94 0.25 [551,568,571,585,594,606,683,745,757,846,928,972,982] CIELAB a* 0.87 2.66 0.74 3.37 [537,559,589,619,645,662,670,851,859,897,922] CIELAB b* 0.78 2.01 0.54 2.20 [619,764,785,812,827,855,881,907,928,938,940,946] Myrcene 0.83 3.26 × 107 0.69 5.29 × 107 [507,523,630,888,894,920,956,964,975,980,989,992,1003] β -Ocimen 0.87 3.10 × 106 0.77 5.52 × 106 502,507,540,548,591,634,659,672,763,795,969,972,979,980] 1-octen-3-ol 0.71 4.28 × 104 0.63 6.96 × 104 [516,528,546,572,679,804,946,959,977,985,995,998,1005] SNV M.C. 0.96 0.23 0.92 0.30 [540,542,571,711,720,791,844,846,849,873,914,917] CIELAB a* 0.80 3.29 0.71 3.58 [537,539,559,565,619,662,847,851,859,891,922] CIELAB b* 0.71 2.20 0.73 1.94 [531,758,764,766,777,784,785,806,899,912,928,938,980] Myrcene 0.80 3.47 × 107 0.75 5.70 × 107 [502,507,511752,920,956,962,964,979,980,982,989,992,1003] β -Ocimen 0.82 3.59 × 106 0.65 6.79 × 106 [502,507,525,534,540,556,596,634,905,951,972,998] 1-octen-3-ol 0.74 4.35 × 104 0.54 7.75 × 104 [528,572,662,795,930,944,959,972,977,990,995,998]

Table 5 Development of Hop Storage Index, αand ß acid content of samples throughout the drying process, normalised at 10% water content.

Time [min] 65 °C, 12 kg 65 °C, 20 kg 65 °C, 40 kg 70 °C, 20 kg

HSI [–] α [%DM]ß[%DM] HSI [–] α[%DM]ß[%DM] HSI [–] α[%DM]ß[%DM] HSI [–] α[%DM]ß[%DM]

σσσσσσσσσσσσ

0 0.22 0.004 8.43 0.69 7.79 0.31 0.22 0.009 8.59 0.38 7.85 0.47 0.21 0.002 8.28 0.58 7.56 0.70 0.22 0.010 7.35 0.81 7.45 0.61 15 0.21 0.002 8.18 0.68 7.75 0.35 0.22 0.005 8.41 0.14 7.98 0.10 0.22 0.002 8.90 0.29 7.79 0.09 0.20 0.007 8.08 0.92 8.02 0.36 25 0.22 0.006 8.82 0.71 7.90 0.16 0.22 0.003 8.82 0.77 8.17 0.04 0.22 0.005 9.88 0.40 8.80 0.17 0.22 0.005 8.02 1.07 7.61 0.16 35 0.21 0.013 8.64 1.16 7.73 0.57 0.22 0.001 8.28 0.86 8.14 0.23 0.22 0.006 9.55 0.38 8.45 0.45 0.22 0.008 7.85 1.14 7.56 0.54 45 0.21 0.009 8.97 0.76 8.00 0.50 0.22 0.008 8.67 1.03 7.98 0.14 0.22 0.006 9.94 0.30 8.60 0.36 0.23 0.009 8.16 1.22 8.21 0.24 75 0.23 0.012 8.42 0.84 7.38 0.60 0.23 0.002 8.48 0.68 8.09 0.49 0.21 0.006 9.56 0.85 8.43 0.66 0.23 0.005 8.06 0.99 7.29 0.17 105 0.23 0.002 8.51 0.42 7.70 0.11 0.23 0.004 7.82 0.29 7.62 0.12 0.21 0.006 8.86 0.49 7.67 0.12 0.22 0.005 8.28 1.17 8.09 0.40 135 0.23 0.001 8.64 0.69 7.62 0.32 0.23 0.004 8.56 0.43 7.77 0.26 0.22 0.001 9.06 0.42 7.91 0.18 0.22 0.006 8.17 0.74 7.58 0.46 165 0.23 0.015 8.30 0.89 7.73 0.42 0.24 0.004 8.63 0.66 7.83 0.13 0.22 0.001 9.54 0.41 8.23 0.04 0.23 0.008 7.68 0.77 7.74 0.38 settings and scheduling on the one hand, and acceptable color changes Writing - original draft, Writing - review & editing, Visualization, on the other. The contents of α and ß acid, as well as the Hop Storage Supervision, Project administration, Funding acquisition. Sharvari Index, were not impacted by the drying conditions and, thus, can be Raut: Formal analysis, Data curation, Writing - review & editing, seen as non-critical parameters in terms of process optimization. Visualization. Boris Kulig: Formal analysis, Data curation, Writing - The presented work illustrates the feasibility of moisture content review & editing. Jakob Münsterer: Conceptualization, Methodology, and chromatic information prediction during hop drying using a novel Validation, Formal analysis, Investigation, Resources, Writing - review dryer and imaging system. The developed models have shown an ac- & editing, Project administration, Funding acquisition. Klaus ceptable level of performance, with future improvements likely to en- Kammhuber: Formal analysis, Investigation, Writing - review & hance this further. The design and combination of the hyperspectral editing. Oliver Hensel: Formal analysis, Writing - review & editing, and RGB imaging systems within the hop dryer are also noted to be Funding acquisition. Stuart O.J. Crichton: Conceptualization, novel and have shown a new way hops can be monitored, in real time, Methodology, Software, Validation, Formal analysis, Investigation, during drying. Further information fusion with process data helped to Data curation, Writing - original draft, Writing - review & editing, create a deeper understanding of the underlying processes within the Visualization, Supervision. hop cones. Further work needs to include the deeper investigation of the losses Declaration of Competing Interest of aromatic components which are highly volatile and, thus, susceptible to losses during drying. In addition, the characteristics of the drying fi curves need to be further investigated, and the interactions between the The authors declare that they have no known competing nancial fl leaves and the spindle modelled. Different storage times pre-drying interests or personal relationships that could have appeared to in u- need to be evaluated for the determination of their impact on the drying ence the work reported in this paper. and the final product quality. Lastly, the predictive models need to be investigated for further improvement and development into a real-time, Acknowledgements or near to real-time, observation tool. This also includes, but is not limited to, the investigation of suitability of methods such as LWPLSR The authors wish to acknowledge the HVG and CNN. Building on the above, smart drying systems which take Hopfenverwertungsgenossenschaft e.G for their financial support in the dynamic changes in the product into consideration can be realized by framework of the Bandtrockner project and the Core Organic Plus implementing the relevant information into the control system. Programme for the financial support within the SusOrganic project (Project Number: BLE - 2814OE006). Financial support for this project CRediT authorship contribution statement was also provided by transnational funding bodies, being partners of CORE Organic Cofund, and the cofund from the European Commission. Barbara Sturm: Conceptualization, Methodology, Software, The study is also part of the SusOrgPlus project within the framework of Validation, Formal analysis, Investigation, Resources, Data curation, the CORE Organic Cofund Programme and is supported by funds of the

11 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Federal Ministry of Food and Agriculture (BMEL) based on a decision of near-infrared (Vis-NIR) hyperspectral imaging. Food Chem. 156, 394–401. the Parliament of the Federal Republic of Germany via the Federal Hermanek, P., Rybka, A., Honzik, I., Hoffmann, D., Jost, B., Podsednik, J., 2016. fi ffi Construction and veri cation of an experimantal chamber dryer for hop drying. In: O ce for Agriculture and Food (BLE) under the BÖLN programme 6th International Conference of Trends in Agricultural Engineering, Prague, Czech (Project number: BLE - 2817OE005). Finally, the authors wish to ex- Republic, 7–9 September 2016. press their gratitude to Dr Helen McKee and Dr Arman Arefi for proof Hermanek, P., Rybka, A., Honzik, I., 2018. Determination of moisture ratio in parts of the hop cone during the drying process in belt dryer. Agronomy Res. 16 (3), 723–727. reading of the manuscript for language and content respectively. Hirschfelder, G., Trummer, M., 2016. Bier: Eine Geschichte von der Steinzeit bis heute (in German). Konrad Theiss Verlag, Sturttgart, Germany. References Hofmann, R., Weber, S., Rettenberg, N., Thörner, S., Garbe, L.-A., Folz, R., 2013. Optimization of the hop kilning process to improve energy efficiency and recover hop oils. Brewing Sci. 66, 23–30. Aberl, A., 2016. Methodenentwicklung der Headspace-Trap-GC-MS Analyse zur schnellen Huang, M., Wang, Q., Zhang, M., Zhu, Q., 2014. Prediction of color and moisture content quantitativen Bestimmung flüchtiger Verbindungen in Hopfen und Bier – mit for vegetable soybean during drying using hyperspectral imaging technology. J. Food Korrelationsanalyse der Hopfenölkomponenten (in German), Dissertation, Eng. 128, 24–30. Technische Universität München. Huck, C., 2015. Chapter 22- Infrared Spectroscopic Technologies for the Quality Control Amigo, J.M., Marti, I., Gowen, A., 2013. Hyperspectral imaging and chemometrics. A of Herbal Medicines, In: Mukherjee, P.K. (Ed.) Evidence-Based Validation of Herbal perfect combination for the analysis of food structure, composition and quality. Data Medicine, Elsevier, Amsterdam, Oxford, Waltham. Handling Sci. Technol. 28, 343–370. Iqbal, A., Sun, D.-W., Allen, P., 2013. Prediction of moisture, color and pH in cooked, pre- Amodio, M.L., Capotorto, I., Chaudhry, M.M.A., Colelli, G., 2017. The use of hyper- sliced turkey hams by NIR hyperspectral imaging system. J. Food Eng. 117, 42–51. spectral imaging to predict the distribution of internal constituents and to classify Kaltner, D., 2000. Untersuchungen zur Ausbildung des Hopfenaromas und technologische edible fennel heads based on harvest time. Comput. Electron. Agric. 134, 1–10. Maßnahmen zur Erzeugung hopfenaromatischer Biere (in German). Dissertation, AOAC, 1995. Method 991.42 & 993.19. Official methods of analysis, 16th ed., Association Technische Universität München. of Official Analytical Chemists, Washington, DC. Kammhuber, K., 2018. Vortrag: Was beeinflusst das Hopfenaroma? (in German), available Aral, S., Bese, A.S., 2016. Convective drying of hawthorn fruit (Crataegus spp.): effect of online https://www.lfl.bayern.de/mam/cms07/ipz/dateien/hopfen_kammhuber_ experimental parameters on drying kinetics, color, shrinkage and rehydration capa- 2018.pdf (accessed 2018-08-28). city. Food Chem. 200, 577–584. Kamruzzaman, M., Makino, Y., Oshita, S., 2016. Parsimonious model development for Ashtiani, S.H.M., Sturm, B., Nasirahmadi, A., 2018. Effect of hot-air and hybrid air-mi- real-time monitoring of moisture in red meat using hyperspectral imaging. Food crowave srying on drying kinetics and textural quality of nectarine slices. Heat Mass Chem. 196, 1084–1091. Transf. 54 (4), 915–927. Kavalier, A.R., Litt, A., Ma, C., Pitra, N.J., Coles, M.C., Kennelly, E.J., Matthews, P.D., Barth, 2018. Der Barth-Bericht (in German), Joh. Barth & Sohn GmbH & Co KG. 2011. Phytochemical and morphological characterization of hop (Humulus lupulus Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference A L.) cones over five developmental stages using High Performance Liquid Practical Information-Theoretic Approach, second ed. Springer, New York. Chromatography coupled to Time-of-Flight mass spectrometry, Ultra-high Centner, V., Massart, D., de Noord, O., de Jong, S., Vandeginste, B., Sterna, C., 1996. Performance Liquid Chromatography photodiode array detection, and light micro- Elimination of uniformative variables for multivariate calibration. Anal. Chem. 68 scopy techniques. J. Agric. Food Chem. 59, 4783–4793. (21), 3851–3858. Kaya, A., Orhan, A., 2009. An experimental study on drying kinetics of some herbal Clement, A., Dorais, M., Vernon, M., 2008. Nondestructive measurement of fresh tomato leaves. Energy Convers. Manage. 50 (1), 118–124. lycopene content and other physicochemical characteristics using visible – NIR Kim, M.K., Kim, H., Koh, K., Kim, H.S., Lee, Y.S., Kim, Y.H., 2008. Identification and spectroscopy. J. Agric. Food. Chem. 56, 9813–9818. quantification of anthocyanin pigments in coloured rice. Nutrition Res. Practice 2, Cocuzza, S., Lutz, A., Müller Auffermann, K., 2013. Influence of picking date on the initial 46–49. hop storage index of freshly harvested hops. Master Brewers Assoc. Am. Tech. Quart. Kokaly, R.F., Skidmore, A.K., 2015. Plant phenolics and absorption features in vegetation 50 (2), 66–71. reflectance spectra near 1.66 µm. Int. J. Appl. Earth Obs. Geoinf. 43, 55–83. Crichton, S.O.J., Münsterer, J., Kammhuber, K., Sturm, B., 2016. Online measurement of Kurzweiler, P., Hildebrandt, A., Riedmayr, A., 2013. Hopfensorten und ihr Aroma. Nachr. hop moisture content and chromaticity during bulk drying with visual - near infra-red Chem. 61, 656–661. hyperspectral imaging, Proceedings of the CIGR AgEng 2016, Aarhus, Denmark, LfL, 2019. Grünes Heft 2019. Bayrisches Landesamt für Landwirtschaft, Institut für 26–29.06.2016, 1–8 ref.14. Pflanzenzüchtung, Arbeitsbereich Hopfen, Wolnzach. available online: https://www. Crichton, S.O.J., Shrestha, L., Hurlbert, A., Sturm, B., 2018. Prediction of moisture con- lfl.bayern.de/ipz/hopfen/022297/ (accessed 19-10-10). tent and chromaticity of raw and pre-treated apple slices during convection drying Li, H., Liang, Y., Xu, Q., Cao, D., 2009. Key wavelenths screening using competitive using hyperspectral imaging. Drying Technol. 36, 804–816. adaptive reweighted sampling method for multivariate calibration. Anal. Chim. Acta Cuervo-Andrade, S.P., 2011. Quality orented drying of Lemon Balm (Melissa officinalis 648 (1), 77–84. L.). Lutz, A., Kneidl, J., Seigner, E., 2013. Hüller Special Flavor-Hopfen – neue De Keukeleire, D., 2000. Fundamentals of beer and hop chemistry. Quim. Nova 23. Herausforderungen. available online: https://www.lfl.bayern.de/mam/cms07/ipz/ https://doi.org/10.1590/S0100-40422000000100019. dateien/hopfen_hopfenbauversammlungen-special_flavor-hopfen_2013__min.pdf (ac- Doe, P.E., Menary, R.C., 1979. Optimization of the hop drying process with respect to cessed 2018-08-28). alpha acid content. J. Agric. Eng. Res. 24, 233–248. Lutz, A., Kneidl, J., Seefelder, S., Kammhuber, K., Seigner, E., 2015. Mandarina Bavaria Doymaz, I., 2004. Effect of pre-treatments using potassium metabisulphide and alkaline (in German), available online: https://www.lfl.bayern.de/mam/cms07/ipz/dateien/ ethyl oleate on the drying kinetics of apricots. Biosyst. Eng. 89, 281–287. hopfen_mandarina_poster_2015.pdf (accessed 2018-06-27). Dusek, M., Jandovska, V., Cermak, P., Mikyska, A., Olsovska, J., 2016. A novel approach Lutz, A., Kneidl, J., Seigner, E., Kammhuber, K., 2016. Die Hüller Special Flavor- for identification of biologically active phenolic compounds in complex matrices Hopfensorten - aktueller Kenntnisstand (in German), available online: https://www. using hybrid quadrupole-orbitrap mass spectrometer: a promising tool for testing lfl.bayern.de/mam/cms07/ipz/dateien/hopfen_h%C3%BCller_special-fl avor.pdf (ac- antimicrobial activity of hops. Talanta 156–157, 209–217. cessed 2018-08-28). EBC, 2007. EBC Method 7.13: Hop Storage Index of Hops and Hop Pellets. European Lutz, A., Kneidl, J., Seigner, E., Kammhuber, K., 2009. Hopfenqualität - Ernte zum rich- Brewery Convention, Belgium. tigen Zeitpunkt (in German), Bayrische Landesanstalt für Landwirtschaft (LfL). ElMasry, G., Wang, N., ElSayed, A., Ngadi, M., 2007. Hyperspectral imaging for non- Institut für Pflanzenbau und Pflanzenzüchtung, Freising-Weihenstephan, Germany. destructive determination of some quality attributes for strawberry. J. Food Eng. 81, Martynenko, A.I., 2006. Computer-vision system for control of drying processes. Drying 98–107. Technol. 24, 879–888. Fatouh, M., Metwally, M.N., Helali, A.B., Shedid, M.H., 2006. Herbs drying using a heat Martynenko, A., 2017. Computer vision for real-time control in drying. Food Eng. Rev. 9, pump dryer. Energy Convers. Manage. 47, 2629–2643. 91–111. Finlayson, G., Mackiewicz, M., Hurlbert, A., 2015. Colour correction using root-poly- Martynenko, A., Sturm, B., 2019. The shift of paradigm from steady-state to intelligent nomial regression. IEEE Trans. Image Process. 24 (5), 1460–1470. drying, Eurodrying’2019, Torino, Italy, 10–12. July 2019. Fiorentini, C., Demarchi, S.M., Quintero Ruiz, N.A., Torrez Irigoyen, R.M., Giner, S.A., Martynenko, A., Kudra, T., 2015. Non-isothermal drying of medicinal plants. Drying 2015. Arrhenius actoviation energy for water diffusion during drying of tomato Technol. 33, 1550–1559. leathers: the concept of characteristic product temperature. Biosyst. Eng. 132, 39–46. Martynenko, A., Misra, N.N., 2020. Machine learning in drying. Drying Technol. 38, Forster, A., Beck, B., 1985. Der Einfluss von Licht auf den Abbau von Hopfenbitterstoffen 596–609. bei der Lagerung von Pellets. Monatsschrift für Brauwissenschaft 38, 372–375. Mejzr, J., Hanousek, B., 2007. Drying of hop. Res. Agric. Eng. 53 (4), 155–160. Gaston, E., Frias, J.M., Cullen, P.J., O’Donnell, C.P., Gowen, A.A., 2010. Prediction of Münsterer, J., 2017b. Flavour-Hopfen optimal trocknen (in German). Brauwelt 36, polyphenol oxidase activity using visible near-infrared hyperspectral imaging on 1063–1066. mushroom (Agaricus Bisporus) caps. J. Agric. Food Chem. 58 (10), 6226–6233. Münsterer, J., 2006. Optimale Trocknung und Konditionierung von Hopfen (in German), GraphPad Software, 2018. Test for appropriate weighting/homoscedasticity. [Online] Bayerische Landesanstalt für Landwirtschaft, Arbeitsgruppe Hopfenbau, Available at: https://www.graphpad.com/guides/prism/7/curve-. Produktionstechnik, Wolnzach, available online: https://www.lfl.bayern.de/mam/ Gundlach, K., Werwie, M., Wiegand, S., Paulsen, H., 2009. Filling the “green gap” of the cms07/ipz/dateien/hopfen_p_23593.pdf (accessed 2019-03-22). major light harvesting chlorophyll a/b complex by covalent attachment of Münsterer, J., 2012. Hinweise für eine optimale Konditionierung von Hopfen (in Rhodamine Red, Biochimica et Biophysica Acta (BBA). Bioenergetics 1787 (12), German), Bayerische Landesanstalt für Landwirtschaft, Arbeitsgruppe Hopfenbau, 1499–1504. Produktionstechnik, Wolnzach, Germany, available online: https://www.lfl.bayern. He, H.J., Wu, D., Sun, D.W., 2014. Rapid and non-destructive determination of drip loss de/mam/cms07/ipz/dateien/hinweise_f__r_eine_optimale_konditionierung.pdf (ac- and pH distribution in farmed Atlantic salmon (Salmo salar) fillets using visible and cesses 2019-10-09).

12 B. Sturm, et al. Computers and Electronics in Agriculture 175 (2020) 105547

Münsterer, J., 2015. Neueste Erkenntnisse zur Leistungssteigerung und Energieeffizienz Shrestha, L., Crichton, S.O.J., Kulig, B., Kiesel, B., Hensel, O., Sturm, B., 2020. bei der Trocknung von Hopfen. In: 55th Congress of the International Hopgrowers‘ Comparative analysis of methods and model prediction performance evaluation for Convention IHGC, 31.07.–03.08.2015, Bad Gögging, Germany. continuous online non-invasive quality assessment during drying of apples from two Münsterer, J., 2016. Erste Untersuchungen zur Optimierung von Bandtrocknern (in cultivars. Therm. Sci. Eng. Progress. https://doi.org/10.1016/j.tsep.2019.100461. German), Bayerische Landesanstalt für Landwirtschaft, Arbeitsgruppe Hopfenbau, Steinhaus, M., Schieberle, P., 2000. Comparison of the Most Odour-Active Compounds in Produktionstechnik, Wolnzach, Germany, available online: https://www.lfl.bayern. Fresh and Dried Hop Cones (Humulus lupulus L. Variety Spalter Select) Based on GC – de/mam/cms07/ipz/dateien/hopfen_erste_untersuchungen_zur_optimierung_von_ Olfactometry and Odor Dilution Techniques. J. Agric. Food Chem. 48, 1776–1783. bandtrocknern_pdf (accessed 2018-10-22). Sturm, B., 2018. Systemic optimisation and design approach for thermal food processes - Münsterer, J., 2017a. Einfluss von Erntezeitpunkt und Trocknung auf den Brauwert der Increase of quality, process- and resource efficiency in dried agricultural products Flavor-Hopfensorten Hallertau Blanc, Mandarina Bavaria und Polaris (in German), manufacturing. Universität Kassel, Witzenhausen, Habilitationsschrift. available online: https://www.lfl.bayern.de/mam/cms07/ipz/dateien/vortragm Sturm, B., Münsterer, J., Kammhuber, K., Crichton, S., 2016. Impact of bulk weight on %C3%BCnsterer_lfl_2017_.pdf (accessed 2018-08-28). drying behaviour and hop quality after drying. In: Proceedings of the CIGR – AgEng Nagel, J., Culley, L.K., Lu, Y., Liu, E., Matthews, P.D., Stevens, J.F., Page, J.E., 2008. EST 2016, Aarhus, Denmark, 26.–29. June 2016, 1–5 ref.13. analysis of hop glandular trichomes identifies an O-methyltrasferase that catalyzes Sturm, B., Moscetti, R., Raut, S., Massantini, R., Bantle, M., 2018. Feasibility of Vis/NIR the biosynthesis of xanthohumol. Plant Cell 20, 186–200. spectroscopy and image analysis as basis of the development of smart-drying tech- Orphanides, A., Goulas, V., Botsaris, G., Gekas, V., 2017. Influence of air-drying on the nologies. In: Proceedings of the 21st International Drying Symposium, Valencia, quality characteristics of spearmint: effects of air temperature and velocity. J. Food Spain, 11.–14. September 2018, pp. 171–178. Process. Preserv. 41, e12817. Sturm, B., Hofacker, W., Hensel, O., 2012. Optimizing the drying parameters for hot air Priest, F.G., Stewart, G.G., 2006. Handbook of brewing. In: Francis, C.T. (Ed.), Food dried apples. Drying Technol. 30 (14), 1570–1582. Science and Technology, Second edition, Taylor & Francis, Boca Raton, p. 853. Sturm, B., Nunez Vega, A.M., Hofacker, W., 2014. Influence of process control on drying Pu, Y.Y., Feng, Y.Z., Sun, D.W., 2015. Recent progress of hyperspectral imaging on quality kinetics, colour and shrinkage of air dried apples. Appl. Therm. Eng. 62 (2), 456–460. and safety inspection of fruits and vegetables: a review. Compr. Rev. Food Sci. Food Su, W.H., Sun, D.W., 2017. Chemical imaging for measuring the time series variations of Saf. 14, 176–188. tuber dry matter and starch concentration. Comput. Electron. Agric. 140, 361–373. Pu, Y.Y., Sun, D.W., 2015. Vis–NIR hyperspectral imaging in visualizing moisture dis- Su, W.H., Bakalis, S., Sun, D.W., 2020. Chemometric determination of time series tribution of mango slices during microwave-vacuum drying. Food Chem. 188, moisture in both potato and sweet potato tubers during hot air and microwave drying 271–278. using near/mid-infrared (NIR/MIR) hyperspectral techniques. Drying Technol. 38, Raut, S., von Gersdorff, G., Wittkamp, S., Münsterer, J., Kammhuber, K., Kulig, B., Hensel, 806–823. O., Sturm, B., 019. Influence of storage time on essential oil components in dried Su, W.H., Bakalis, S., Sun, D.W., 2018. Fourier transform mid-infrared-attenuated total hops. In: Proceedings of Eurodrying 2019 – 7th European Drying Conference Torino, reflectance (FTMIR-ATR) microspectroscopy for determining textural property of Italy, July 10–12, 2019, pp. 415–420. microwave baked tuber. J. Food Eng. 218, 1–13. Rajkumar, Elmasry, Raghavan, Gariepy, 2012. Studies on banana fruit quality and ma- Wu, D., Sun, D.W., He, Y., 2012. Application of long-wave near infrared hyperspectral turity stages using hyperspectral imaging. Journal of Food Engineering, 108(1), imaging for measurement of color distribution in salmon fillet. Innovative Food Sci. 194–200. http://doi.org/10.1016/j.jfoodeng.2011.05.002. J. Food Eng. 108, Emerg. Technol. 16, 361–372. 194–200. Wu, D., Wang, S., Wang, N., Nie, P., He, Y., Sun, D.W., Yao, J., 2013. Application of time Romano, G., Nagle, M., Argyropoulos, D., Müller, J., 2011. Laser light backscattering to series hyperspectral imaging (TS-HSI) for determining water distribution within beef monitor moisture content, soluble solid content and hardness of apple tissue during and spectral kinetic analysis during dehydration. Food Bioprocess Technol. 6, drying. J. Food Eng. 104 (4), 657–662. 2943–2958. Václav Rybáček (Ed.), 1991. Hop production, Elsevier, Amsterdam, New York. Yang, Y.G., Sun, D.W., Wang, N.N., Xie, A., 2015. Real time evaluation of phenol oxidase Rybka, A., Hermanek, P., Honzik, I., 2019. Hop drying in belt dryer using cooling (PPO) activity in lychee pericarp based on weighted combination of spectral data and chamber. In: Proceedings of the 7th International Conference on Trends in image features as determined by fuzzy neural network. Talanta 138, 198–207. Agricultural Engineering, Prague, Czech Republic, 17–20 September 2019, pp. Yao, Y., Chen, H., Xie, L., Rao, X., 2013. Assessing the temperature influence on the 464–469. soluble solids content of watermelon juice as measured by visible and near-infrared Rybka, A., Hermanek, P., Honzik, I., Krofta, K., 2017. Parameters of the drying medium spectroscopy and chemometrics. J. Food Eng. 119, 22–27. and dried hops in belt dryer. Res. Agric. Eng. 63, S24–S32. https://doi.org/10. Zeisig, H.-D., 1970. Die Trocknung des Hopfens mit erhöhter Heißlufttemperatur unter 17221/35/2017-RAE. der Berücksichtigung ihrer Anwendung in der Darre und im Bandtrockner (in Sharpe, F.R., Laws, D.R.J., 1986. The essential oil of hops – a review. J. Inst. Brew. 87 (1), German). Dissertation. Technische Hochschule München, Germany. 96–107.