Jonn R. Hll , Olivi Mtthews , Timothy N. Volonkis , Eri Liggins ,Krl P. Lymer , Roln Beley , Innes C. Cuthill , Nihols E. Sott-Smuel ,**

a School of Psychological Science, University of Bristol,12a Priory Road, Bristol, BS8 1TU, UK

b School of Biological Sciences, University of Bristol, 24 Tyndall Avenue, Bristol, BS8 1TQ, UK

c QinetiQ Ltd, Cody Technology Park, Farnborough, Hampshire, GU14 0LX, UK

d Defence Science and Technology Lab, Physical Sciences Group, Porton Down, Salisbury, Wiltshire, SP4 0JQ, UK

Keywords:Camouflage Detection Military Civilian Expertise Training

ABSTRACT The human visual system is still an important factor in military warfare; military personnel receive training on effective search strategies,and camouflage that can effectively conceal objects and personnel is a key component of a successful integrated survivability strategy. Previous methods of camouflage assessment have,amongst others,used psychophysics to generate distinctiveness metrics. However, the population from which the human observers are drawn is often not well defined, or necessarily appropriate. In this experiment we designed a new platform for testing multiple patterns based on a camouflaged object detection task, and investigate whether trained military observers perform better than civilians. We use a two-alternative forced choice paradigm, with participants searching images of woodland for a replica military helmet displaying Olive Green, Multi Terrain Pattern, US Marine Pattern or,as a conspicuous control,UN Peacekeeper Blue.Our data show that there is no difference in detection performance between the two observer groups but that there are clear differences in the effectiveness of the different helmet colour patterns in a temperate woodland environment. We conclude that when tasks involve very short stimulus presentation times,task-specific training has little effect on the success of target detection and thus this paradigm is particularly suitable for robust estimates of camouflage efficacy.

1. Introduction

Exploited by nature and the military alike, camouflage aids in the prevention of target acquisition(detection and recognition)[1].Assessing the effectiveness of camouflage is therefore an essential component of any military procurement decision, but also central to studies of biological camouflage which use humans as model‘predators’ (e.g. Refs. [2-5]). It is therefore a concern if tests of camouflage are not robust to differences in expertise,particularly if they use participants who lack specialist skills for target acquisition, either acquired through training, in the case of military personnel, or evolution, in the case of non-human animals.

Current methods used by the military for ecologically valid testing of the effectiveness of camouflage patterns are highly expensive in terms of finance and manpower,as new patterns have to be physically produced in fabric and then tested in the field,requiring the time of multiple trained personnel. This means that relatively few patterns are ever tested, and the best possible patterns may never be identified. A test that would allow a larger number of new patterns to be compared to each other (and to existing patterns)in a simple manner involving non-experts would be a radical change to current procurement methods.Such a testing platform is not required to be the final test of camouflage patterns:field trials are still the best option for this. Rather, we aimed to provide an initial method for testing multiple different patterns quickly and easily, with the best options then passing through to the next round of more stringent testing.Our goal,therefore,was to develop a camouflage detection paradigm that would be robust to variations in training/expertise effects and to evaluate its performance with trained military and na?ve civilian observers.

Natural scenes can provide a complex background for search and it is often difficult to know whether various parts of the scene might act as background or distractors. A target displayed on a complex background is more difficult to find due to imperfect segmentation of the objects in the scene[6,7].Prinzmetal&Banks[8] showed that identification accuracy decreases with increasing number of distractors, with reduced similarity of distractors, and with increased similarity between target and distractors. In a natural woodland scene, all the different components of the image could be considered to be distractors meaning there will be many distractors with large differences between them. Search for a specific target will therefore take longer than it would in a simpler scene. If that target is also coloured to resemble the scene then in effect it will be more similar to the distractors and so detection will be even slower.

There are a number of different camouflage strategies in the visual domain, of which background matching and disruptive colouration are two of the best known. Background matching, as the name suggests,involves displaying colours and textures that mimic those found in the environment in which the target will be viewed[9,10]. Disruptive colouration may also involve elements of background matching but with the addition of high contrast patches that interfere with perceptual grouping mechanisms by breaking up the outline or other distinctive features of the target [3,5,11].Greater contrast or colour differences between adjacent patches have been found to increase the strength of the false internal edges[12-14] but, if the patches are too different from the background,the object can be detected more quickly[2,15].This means that the effectiveness of a specific pattern will depend upon the background on which it is viewed. Direct mimicry of an irrelevant background object (‘masquerade’ [16]) is another strategy common in nature and restricted military applications [17,18]. All of these strategies reduce signal to noise ratio, whether in detection or object classification [19].

Since disruptive colouration breaks up cues to identity rather than simply ‘blending with the background’, it also potentially allows for more flexibility by giving some measure of concealment across multiple backgrounds. However, because disruptive colouration still requires matching of at least some colours to the background, that flexibility is constrained. So, for example, a single pattern might provide good concealment across woodlands in Western Europe but it would be unlikely to be as effective in deserts. This is an issue for military camouflage where the economic and logistical costs of generating many different camouflage patterns for use in different environments must be balanced against the human cost of relying on locally sub-optimal patterns.Another strategy would be to adopt a biologically inspired adaptive or active camouflage system that changes to match any environment. This type of technology would, however, be expensive and the effectiveness of the patterns would still need to be assessed before the system could be relied on to increase survivability.

Since the success of concealment through camouflage is of great importance to the military, it is a priority that the effectiveness of different patterns should be accurately assessed for the different environments in which they may be used. While technological advances mean that sensors now play an increasing role in the detection of personnel and hardware, camouflage must still routinely be effective against human observers;many conflicts are asymmetric. Even when sensors are available, suspect areas are often detected by the human eye, through visual search, prior to more detailed surveillance being employed. Thus better camouflage can increase time to detection and hence confer a survival benefit. It is therefore imperative that any assessment methods take this into account and provide valid methods to measure concealment from the human visual system.Recent advances have also led to the design of a computer-based human observer model whose behaviour has been shown to correlate highly with that of civilian (non-expert) participants [20]. Whilst this system has the potential to increase the efficiency of camouflage testing, particularly in situations where it may be impractical to use human participants,its efficacy currently also relies on the assumption that it behaves in a manner that is similar to experts as well as the civilians to which it was compared.

Military personnel receive training on how to search for targets,and the exact details of this training are classified. If our goal of designing a platform for efficient early stage testing of camouflage patterns is to be successful, then it needs to be robust to any training that military personnel might have received. While segmentation is more difficult in cluttered scenes, the visual system can gain benefits from top-down information such as that gained through perceptual training [21]. Learning effects in visual search tasks have been found to be highly task specific(e.g.Refs.[22-24]),even in experiments that mimic real world applications such as searching for objects in airport-xray images [25]. However, in experiments using a flight simulator, Guznov [26] tested different training strategies for unmanned aerial vehicle pilots performing target search, and showed that training can improve performance when compared to a control condition.There is also evidence from laboratory-based visual search tasks that performance in camouflage search can be improved with training and this increased performance can be transferred to the search for similar, novel,camouflaged objects(e.g.Refs.[21,27-29]).This effect may be due to an improvement in search efficiency via enhanced segmentation[21]. However, compared to real world situations, these studies involve narrow transfer,where training and test conditions are very similar, and it has been shown that as task difficulty increases perceptual training becomes more specific and transfer is more limited[22].Therefore in order for our testing platform to be robust to expertise,we designed our task with a high level of difficulty and tried to reduce the chance that our conditions were of a close match to those of any training received by military personnel. While the details of the training received by military personnel are classified,our limited understanding is that it likely involves scanning search strategies. Therefore a two-alternative forced choice paradigm,specifically with short duration times to limit the impact of scanning strategies, was chosen. For any level of scanning strategy training to impact performance on this type of task would require a much wider transfer of training than has so far been reported in the literature.

Psychophysics, which measures perceived as opposed to physical attributes, has previously been used to generate and test metrics of‘target distinctiveness’.These metrics can then be used to make judgments of the effectiveness of camouflage (e.g.Refs.[30-35]).By generating a metric for each camouflage pattern,the similarity of each to a specific background can be compared,thus allowing the best option to be identified. Psychophysics is a useful approach because it exploits a relevant visual system-that of human observers - in a controlled test environment. However,there are two main criticisms that have been levelled at this approach. First, is the use of unrepresentative samples: in this particular area of research, the populations from which observers have been sampled have not always been specified (e.g.Ref. [31,32,34]) or have included only civilians (e.g. Ref. [35]). The second is that visual search on a computer screen,as is standard in psychophysics, may lack ecological validity, because of the lack of immersion and potential bias toward close attention and foveation of targets[14].To overcome this, Egan and colleagues used a large display screen to increase the validity of their study on visual search for snake camouflage patterns. We follow their example in our study by displaying all images on a 1.90× 1.07 m projector screen with a high contrast range projector. Participants were seated at a distance from the screen that meant all objects in the images were the same physical size as they would have appeared in the real world.

Our testing platform focused on the detection stage of target acquisition and, using psychophysical techniques, we designed a test paradigm that was intended to be immune to training effects,but still produce robust detection metrics.This would establish that civilians could be used as models for military observers during initial screening of new camouflage patterns for effectiveness. We evaluated the success of this paradigm by comparing the performance of trained military observers and civilians at detecting three different military camouflage patterns and a control pattern in a wooded environment. The four patterns used in the experiment were UN Peacekeeper Blue(UNPKB),Olive Green(British Standard BS381C-220), Multi-Terrain Pattern (MTP, the current UK Army issue) and the US Marine Pattern woodland pattern (MARPAT; see Fig.1).Whilst Olive,MTP and MARPAT are all camouflage patterns,UNPKB is designed to be a conspicuous signal and was included as a control for performance differences unrelated to detectability.Since Olive contains no texture information to break up the surface of the helmet, we hypothesised that this design would be less effective against the complex background of woodland than MTP and MARPAT. Thus we predict that the UNPKB will be the easiest pattern to detect and the camouflage patterns involving texture will be the most difficult to detect.

In summary, we designed a new platform for camouflage pattern testing which would allow a greater number of patterns to be included in the initial testing stages, with the most effective patterns being identified for further evaluation. To assess the effectiveness of camouflage in military and civilian contexts, we require a visual detection task that is robust and immune to effects of training. Otherwise we would be compelled to use only trained participants, and results from visual search experiments using untrained participants (which is the norm) would be of dubious value.A 2AFC task with short viewing time(250 ms)is appropriate,because it ensures a maximum of two fixations on each image and thus should be resistant to scanning strategies that might be affected by training.However,this needs testing,which is the goal of our study.

2. Methods

2.1. Stimuli

The target was a replica Personnel Armor System for Ground Troops(PASGT)helmet on to which one of four differently coloured covers was fitted (as shown in Fig.1).

Fig.1. Military patterns tested. Top left: UN Peacekeeper Blue, top right: Olive Green, bottom left: MTP, bottom right: MARPAT.

Photos of the replica helmet were taken in Leigh Woods(North Somerset, UK, 51°26′57′′N and 2°38′25′′W) between May and July 2013. Leigh Woods is a mixed deciduous woodland environment,dominated by English Oak (Quercus robur) and European Beech(Fagus sylvaticus) and is therefore broadly representative of woodland environments across Western Europe. Images were taken along tracks in the woods culminating in a set of images with 27 different background locations. At each of these background locations, nine images of each helmet pattern were taken: one at each distance of 3.5,5 and 7.5 m,in the centre,left and right of the visual field (see S1-S4 Supplemental Images for examples of photographs).This resulted in a data set that contained 243 images per helmet pattern. Orientation of the helmet was randomised in 45-degree increments, using pre-generated random sequences.Images of each background location without a target were also taken, along with images of the background locations with a Gretag-Macbeth Colorchecker chart(X-Rite,Grand Rapids,MI,USA)for use in colour calibration.

Images were taken using a Nikon D80 SLR (Nikon Corp., Tokyo,Japan) on a tripod (Manfrotto 190 MT190XPRO3; Manfrotto S.p.A.,Cassola, Italy) at approximate human eye level (1.75 m). Images were taken in RAW (Nikon NEF) format and converted to TIFF for calibration. They were then calibrated and downsampled(3872×2592 to 1936×1296 pixels).Calibration was performed by using a cubic polynomial fitted to the measured relationship between the colours of the colour chart at the location, and those of the images when displayed on the experimental set up [36,37].

2.2. Experimental set-up

For the experiment, the images were displayed on a projector(Panasonic PT-AE 7000U; Panasonic Corporation, Osaka, Japan).The projector had a spatial resolution of 1920 × 1080 pixels and maximum contrast ratio of 300,000:1.Aspect ratio was set at 16:9 and projections were keystoned to the value + 5. The area of the projector screen used to present stimuli was 1.90 × 1.07 m, had a gain of 1.1 and mean luminance of 47.4 cd/m2. Participants were seated 2.55 m from the projector screen,subtending a visual angle of 41°× 24°. An Apple MacBook Pro (Apple Inc., Cupertino, CA,USA)controlled the projector and an Apple Keyboard with Numeric Keypad was used by participants to indicate their answer. The arrangement was such that objects displayed in the images appeared a similar size to the participant as they would have done if observed in the real world.

Each participant completed 135 trials that were split into five blocks of 27 trials to allow participants a short break between blocks;in practice none took the opportunity.For each participant,trials were randomly selected from the dataset of all targetbackground combinations with the only condition being that there were approximately equal numbers of each pattern condition shown to each participant. The order of the trials was also randomised.

Fig. 2. Trial procedure:two photographs of a woodland scene, one with a helmet and one without, each preceded by a fixation cross on a mean colour background.

The experiment used a temporal two-alternative-forced-choice paradigm. In each trial, participants saw two images (of different woodland locations) in sequence and one of these images, first or second with probability 0.5, contained a helmet. The participants were instructed to view the images and then indicate which one contained the helmet. At the beginning of each block, participants were shown an image illustrating the four different pattern conditions for which they would be searching. They then pressed the space bar to indicate that they were ready to begin.See Fig.2 for the sequence of images displayed in each trial, and their timing.Participants indicated their response (helmet in photograph 1 or 2?) with a key press, after which the next trial started automatically.If participants did not respond within 1000 ms the next trial started anyway,but participants were encouraged to answer within this time and to guess if they really did not know.

2.3. Participants

Fifteen participants(14 male,1 female)were recruited from the UK Army’s Armoured Trials and Development Unit, Bovington(trained military observers),and thirteen(12 male,1 female)from the UK Government’s Defence Science and Technology Laboratory or the multinational defence technology company QinetiQ (civilians). All participants were therefore knowledgeable about, and interested in, military technology, but only half had received specialist training in visual search for camouflaged targets. All participants were na?ve to the purpose of the experiments, had normal or corrected-to-normal vision, and gave written informed consent in accordance with the Declaration of Helsinki. The study was approved by the Research Ethics Committee of the Faculty of Science, University of Bristol.

2.4. Data analysis

The data were analysed using Generalized Linear Mixed Models(function glmer in the package lme4[38])in R 3.2.2 [39] to model the binomial error distribution.Helmet cover pattern and observer group were treated as fixed effects with participant as a random effect. Fixed effects were assessed via the change in deviance and degrees of freedom between models with and without the term in question, tested against a chi-squared distribution. Pair-wise tests were carried out with Tukey-type adjustment for multiple testing(function glht in the multcomp package[40]).

3. Results

There was a significant effect of pattern (χ2= 332.97, df = 3,p < 0.0001) but no effect of observer group (χ2= 0.38, df = 1,p = 0.5399) or interaction between observer group and pattern(χ2= 1.07, df = 3, p = 0.7854). Pair-wise tests indicated that detection probability for UNPKB was significantly higher than for all other patterns (Olive Drab: p < 0.0001, MTP: p < 0.0001, MARPAT:p<0.0001)and detection was significantly lower for MARPAT when compared to all other patterns(Olive Drab:p<0.0001,MTP:p<0.0001).However,there was no clear evidence of any difference in detection success between Olive Drab and MTP (p = 0.075; see Fig. 3).

4. Discussion

Fig.3. Detection success (mean ±95% confidence intervals based on fitted model) for four helmet patterns across civilians and trained military observers.

Our aim was to design a detection task that would act as the first stage of testing for new camouflage patterns. It was specifically intended to be efficient and robust to training effects. In order to reduce the impact of training we designed the task to be as different as possible from training methods that military personnel might encounter and with a high level of difficulty.Consistent with this, we found no evidence for a difference between civilians and military observers in our task. Difficulty was achieved using short stimulus times with the images displayed on the screen for just 250 ms.This is only long enough for one or,at most,two fixations as the mean duration for a single fixation in humans performing visual search is 180-275 ms [41]. While evidence from the visual search literature suggests that training can improve performance at detecting camouflaged objects and that this training can be transferred (e.g. Ref. [21,27-29]), transfer tends to be narrow, with the training and test conditions being very similar.While the details of the training received by military personnel are classified, our understanding is that it likely involves scanning search strategies.Limiting search durations to one or two fixations prevented the possibility of scanning strategies being implemented and thus was predicted to reduce the impact of military training or expertise on performance. This was borne out by the fact that we found no differences between the civilian and military participant groups.Image statistics have also previously been shown to assist in camouflage breaking [28]. However, in real world scenarios, image statistics will not be constant across environments and so to avoid observers using this cue, we used multiple background images without any manipulations of the images, other than basic calibration to ensure real world colours. The important practical implication of our result is that in future, a testing platform involving target detection in very short stimulus displays to assess camouflage can be run using civilians without compromising the results.Another implication of the similar results between observer types,is that the behaviour of the computer-based human observer model described in Volonakis et al. [20] will correlate with that of expert humans as well as civilians for this type of task. This could prove particularly useful in situations where it is not possible to test using human participants.

It is possible that there was no difference between the two observer groups:not because the military observers were unable to exploit their training, as we intended, but rather because their‘expertise’ level was actually no higher than that of the civilian observer group. Empirical evidence of the difference between the two observer groups would strengthen the argument that the paradigm we present is robust to differences in training across a greater range of observers.However,the aim of our experiment was to establish that civilians acted as reasonably accurate models for military observers using our testing platform, and thus the exact level of expertise of the latter group is less important than their being a good representation of the military population in general.Since the military observers group consisted of active members of the military from various different ranks we feel that this is a reasonable assumption.

To illustrate this point further,imagine a scenario where a new camouflage pattern is being sought. Five new experimental patterns are devised and need to be tested. Creating and fully testing all five patterns in the field will be very expensive in terms of time and finances, especially as it is unclear whether any of these new patterns will be more effective than those that already exist. Thus these five new patterns,along with two existing patterns,are tested on civilians using the paradigm described in this paper.Two of the patterns perform poorly compared to the existing patterns and three are more successful.This means that the two least successful patterns can be dropped from testing and the three more successful patterns can be taken into the next round of testing.The next round of testing will involve military observers performing a more ecologically valid task and it is at this point that evidence of the observers’expertise will be critical as they will need to be accurate models for the personnel that will potentially rely on the camouflage pattern in future. However, in the case of the current paradigm being used as an initial test, it is less critical that the exact expertise of the military observers is known.As long as the civilians are a close enough model for military observers, they can provide an accurate guide as to which patterns may potentially be successful in later tests and which patterns stand no chance out in the field and should be dropped from further testing.

While the task was robust to differences between our two observer groups,it was still able to effectively differentiate between the four different helmet patterns. UNPKB fitted with our prediction that,as a signal,it would be significantly easier to detect than the other patterns.We also predicted that the camouflage patterns with texture would be the most difficult to detect. This was borne out in the case of MARPAT, which was the least successfully detected pattern.However,MTP did not follow our prediction and,despite the lack of texture in the Olive Drab pattern, these two conditions produced very similar detection success. We can speculate that MARPAT may be harder to detect than MTP in this wooded environment because it has a luminance value more similar to that of the background (MTP is paler than MARPAT),because the different colour patches more accurately mimic the background environment in their complexity, because the higher contrast patches are more effective at breaking up the surface[42]and/or outline [13] of the helmet or because it is a digital pattern combining a background matching micro pattern with a disruptive macro pattern ([43]; see Supplemental Material for further discussion of this point). These explanations are not mutually exclusive, but further investigations would be required to establish the relative importance of each.

The fact that MTP was as effective as Olive, which contains no texture information, suggests that, in this environment at least,either the texture patterns in MTP are too subtle to impede detection or its average colour is too pale for temperate woodland.Although‘multi-terrain’is now standard issue for the British Army,MTP was procured primarily for deployment in the mixed arid/wooded/urban environments of Helmand,Afghanistan,rather than temperate woodland[44].Furthermore,military camouflage serves a role in identification as well as concealment [45] and procurement decisions should take this into account. Nevertheless, it is possible that MTP’s spatial pattern would be more effective in woodland if the colours and contrast were more similar to those of MARPAT. Stobbe & Schaefer [46] reported that prey with low contrast stripes survived no better than a background matching control and Barnett et al. [47-49] showed that at longer observation distances even high contrast colours,such as black and yellow,will merge into a perceived intermediate colour.The same may be true in the case of MTP and we can speculate that at some of the viewing distances used in this experiment the patches may have started to merge into a perceived pattern similar to Olive Drab(see S5.Supplemental Information for further analysis and discussion of distance effects). In general, high spatial frequency information such as fine details and texture can be perceived only at shorter distances while low spatial frequency information such as variations in global luminance and broad contours dominate at longer distances [50]. It is therefore important that camouflage assessment methods do not only consider the visual system of the observer and the relevant background but also the distance at which the camouflage is likely to be observed.

In this experiment the target was present in one of the images in each trial pair, and thus target prevalence was quite high. This is unlikely to be the case in real world scenarios. Target prevalence has previously been shown to have a strong effect on search behaviour, mostly by affecting decision criteria [51]. When targets are rare,they are also more likely to be missed even when they are present [52,53]. When the target prevalence is higher, as in our experiment, there is an increase in the probability of the target being found. This does not appear to have adversely affected our experiment,since we were still able to identify differences between the various camouflage patterns that were robust to expertise.However, the effect of prevalence does suggest an interesting avenue for subsequent rounds of testing of potential camouflage patterns:a set-up using experts and lower target prevalence could increase ecological validity and help to differentiate the effectiveness of the different patterns further. That said, the increased probability that a camouflage is found due to higher prevalence is not necessarily a bad thing. Given the cost to human life if the effectiveness of a pattern is overestimated,an underestimate of its effectiveness would appear preferable.

4.1. Further work

One obvious limitation of this study is that participants were viewing two-dimensional images rather than being immersed in a real environment.The use of a large projector screen so that objects on the display appeared the same size as if viewed from the relevant distance in the real world should have gone some way to mitigate this issue.However,recent developments in virtual reality technologies and urban terrain simulations (e.g. Refs. [54]) mean that in future, similar testing platforms could exploit 3D immersion,promoting greater ecological validity.This approach could also allow the effects of camouflage patterns to be easily assessed for a greater range of military targets, such as an armoured personnel carrier, battle tank or prone dismounted soldier.

Our goal was to design a testing platform for the first stage of testing for new camouflage patterns that would allow multiple patterns to be tested quickly and easily.We propose that this would be the first stage in a series of rigorous tests that would allow multiple camouflage patterns to be reduced down to one or two patterns that are most effective in a specific environment. At each stage the tests would need to differentiate between the remaining patterns and so later stages would likely involve increasing the range of distances and search times where training would be more likely to have an impact.At this point military personnel would be required to act as observers but the patterns they would be testing would have already been through previous tests, increasing the efficiency of the overall process.Search and detection success could then be compared across a broader range of situations and the use of eye-tracking could provide further details of any search strategies employed. It may also be possible to synthesise images of targets with new camouflage designs, thus allowing initial comparisons to pre-existing patterns without the need to physically create new materials. Field trials could then be used as a final evaluation of the pattern(s) that have made it through the rest of the testing process.

5. Conclusions

We conclude that for this testing platform,with very short(subsaccade-duration) stimulus display times, there were no observed differences between civilians and trained military observers at detecting camouflaged objects.Furthermore,the obtained rankings of detection probabilities for the four colour patterns show that the short-display-time 2AFC paradigm has sufficient sensitivity to discriminate between qualitatively similar camouflage in terms of detectability.

Ethics statement

All experiments were carried out in accordance with the Declaration of Helsinki and with the approval of the University of Bristol Faculty of Science Human Research Ethics Committee(approval code 26,111,528,301).

Availability of data and materials: The dataset supporting the conclusions of this article is available from the University of Bristol ResearchDataRepository,https://doi.org/10.5523/bris.7wp6f2uby3ai29cm3fr379gzz.

Funding

This work was supported by QinetiQ (contract number UoBMASTSUB/1000067064) and the EPSRC (grant number EP/M006905/1).QinetiQ enabled access to military participants,but all data collection and analysis was completed by researchers from the University of Bristol.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.dt.2020.11.004.