Quality control of the data is an important step in all scientific processes, and also very important for citizen science projects (Wiggins et al. 2011). The data validation procedure in the ‘waarnemingen.be’-database combines species specialists (experienced volunteers) assigning a validation status to observations and an algorithm automatically evaluating observations. This multi-step process depends on the proof presented (not mandatory but possible), species status (common vs rare), location and time (was there already a proven record of presence within a species group dependent defined range of space and time) of the observation (Swinnen et al. 2018). Species specialists can assign a validation status to an observation: (a) ‘Approved (based on evidence)’, evidence can be a picture or sound, (b) ‘Approved (based on expert judgement)’, the additional information or the knowledge of the observer makes it highly likely this is a correct observation, (c) ‘Under review’, temporary status, no decision has been taken yet, (d) ‘Cannot be assessed’, proof or explanation does not allow for a decision to be made, (e) ‘Rejected’, observation was wrong and user does not correct it. The algorithm can also assign a validation status: (f) ‘Automatic validation’, for a record to be automatically validated, there need to be a number of earlier observations of the species supported by proof (at least one or two), within a certain radius (ranging from 100 m to 10 km) within a specified time range (60–3000 days). Remaining observations are classified (g) ‘unverified’. The validation process is an interactive process where users can be contacted for additional information or suggested to change the species name or other details in case of an error. We investigate the possible error ratio by calculating the percentage of approved observations (based on photographic evidence) which was initially wrong but corrected by the user after interaction with a validator.

To allow standardised data collection and a quantifiable measure of search effort, two options for data registration are offered to users. In 2013, the option to gather standardised transect data was added to the website. Users were asked to choose a specific route, draw it online and check it at least once every two weeks, but not more than once a day. They were asked to fill in the survey, even if no roadkill was detected. These type of transects are called fixed transects in this manuscript. Since 2018, smartphone users can allow their app to register their transect while observing nature and registering observations. When finished, users indicate per species group if their transect can be used as a roadkill monitoring transect. Since there are no requirements for transects to be identical, or to be repeated over time, we call them variable transects.

For the fixed transects, users register the transport modus (on foot, by bike, by car). For the variable transects, the transect is recorded by the smartphone and we derived the speed from the track length and duration, and classified transects as 0–7 km/h as on foot, 7–25 km/h by bike and >25 km/h by car (although another motorised vehicle is also possible). This distinction according to speed is important because speed affects detection probability and it is known that searching on foot is more effective than counting while driving (Slater 2002). Data collected during standardised monitoring contains more information but it is also more demanding for volunteers resulting in a smaller number of participants (Bonney et al. 2009).

Waarnemingen.be is mainly used as a personal notebook by naturalists to register and document their sightings. Although some users are aware of the additional scientific advantages standardised data collection offers, the majority of all observations in waarnemingen.be are presence only records (also known as roving records) (Vercayie and Herremans 2015). Given the correct identification of the species, presence is confirmed but search effort is unknown. The absence of a record can have multiple causes: no roadkill present, no observer present or both present but not registered by the observer. We show a summary of the transect data including transect characteristics and top 10 of recorded bird and mammal species and calculate the average distance that needs to be covered to encounter a roadkill. For the presence only data, a top 20 for bird and mammal casualties is presented and we compare the results with the data collected during transect counts. While herpetofauna is also an important species group, e.g. because of their worldwide threatened status (Heigl et al. 2017), we do not discuss them here since they are only recorded at lower driving speeds, and a larger (roadkill) database, separate from waarnemingen.be is available, calling for a specific analysis.

The number of new observations (of all organisms) submitted to waarnemingen.be continues to increase year after year, from 400,000 in 2008 to over 6,000,000 in 2020 (and over 8.7 million in 2021). For 2010–2020 we investigate by means of a linear regression (R Core Team 2016): (a) is there an increase in mammal roadkill observations? (b) is there an increase in mammal observations (excluding all automated observations by camera traps and bat-detectors since they do not represent human search effort)? (c) are both correlated?

The large majority of roadkill data is collected as presence only data. Since search effort is unknown, absolute roadkill trends per species cannot be calculated. However, relative trends can be calculated and give an indication of the increase or decrease of roadkill abundance of a specific species compared to the other species killed on the road. For this analysis all mammal roadkill records were combined (presence only and transect data), excluding observations where observers indicated they were uncertain of species determination (1.5% of observations), and only species with a minimum of 50 roadkill individuals were withheld, resulting in 17 species (only species level records were considered). Per species, the percentual abundance per year from these 17 species was calculated. By using a linear regression, we determine per species if the relative proportion per year changed significantly between 2010 and 2020. Graphs were made using ggplot 2 (R Core Team 2016; Wickham 2016). Based on unstructured, presence only, citizen science data on roadkill, we propose the relative change in proportion of roadkill victims as a means to gain insight in relative population changes as roadkill numbers are expected to be strongly and positively associated with the local abundance of living animals (Baker et al. 2004; George et al. 2011; Pettett et al. 2018; Schwartz et al. 2020).

Apart from the local abundance, timing within the year does influence the number of victims found. Animals are sensitive to wildlife vehicle collisions during movement. This can be daily movement while foraging or patrolling home ranges, or seasonality in mating, juvenile dispersal or migration (Taylor and Goldingay 2010; Garriga et al. 2017; Schwartz et al. 2020). For all roadkill data combined (presence only and transect data) we plot species specific density functions using ggplot 2 (R Core Team 2016; Wickham 2016). For this, the number of records was used, and not the number of individuals. Overall, 98.7% of records comprises a single individual, but more than one individual is also sometimes reported. This can reflect reality, multiple individuals killed at once or, sometimes, users combine a number of observations from a timespan from the same location and add a single observation to the database. Analysing these ‘combination records’ as if all individuals were killed at the same time would introduce errors in this seasonal pattern and to avoid this, the number of observations was used. For species with more than 1,000 records, we show the annual seasonal pattern in roadkill data. When fewer data are available, a single density plot combining the data from 2010–2020 is shown.

In total, 38.9% of records were approved based on different procedures (Table 2). For all observations approved based on the presented photographic evidence, only 139 out of 7,687 recordings needed to be corrected by the validator. This results in an error rate of 1.8%. In only a very small percentage of cases, users do not respond to suggestions to change the species and the observation is then rejected.

All observations which were rejected, under review or which cannot be assessed are removed in the following analyses.

In order to determine which species is killed the most in traffic, standardised monitoring is necessary. Our results indicate that for birds and mammal species, searching at an intermediate speed from 7 to 25 km/h results in the highest number of carcasses found. This is somewhat unexpected given that a slower speed should increase detection rates (Slater 2002). We suggest that the searching for roadkill carcasses was fitted into the routine of a number of people in the past years and that biking happens more frequently next to busy roads, where more carcasses are present compared to walking, which is more likely along calmer roads. Driving by car resulted in roughly the same encounter rates of birds and mammals carcasses compared to walking, however due to the higher speed, corpses not identified to species level are more numerous. Stopping safely to identify the species is often not possible in Belgium and stopping on motorways is forbidden (and dangerous) (minimum speed 70 km/h, maximum speed 120 km/h). At this speed, identification at species level is frequently impossible.

The quality of transect data (with a standard protocol) is higher but it is more difficult to find volunteers to collect them (Bonney et al. 2009; Vercayie and Herremans 2015). As a consequence, they only represent 6% of all available roadkill data from Flanders. Although informative for local situations, currently, this is too sparse for region-wide analysis. The variable transects are promising in this respect because they can be monitored anywhere and anytime, but they are currently not yet widely enough adopted by the user community. It is also too early for a detailed analysis since they were only launched in 2018. Additional promotion and awareness in the user community of the applicability could boost the popularity of these variable transects.

There is a clear difference between the rank list of most observed species during transects and the rank list of most observed species in the opportunistic data. When comparing data collected by car and bike, it is clear that only larger species are registered from cars and a higher proportion was not identified on species level. For the mammal data, all rank lists of most observed species are led by Hedgehogs (with the exception of unidentified mammals which outrank them in species lists collected from cars). Hedgehogs are frequently reported as traffic victims in Western Europe (Huijser and Bergers 2000; Pettett et al. 2018) and road mortality of Hedgehogs is expected to be an important factor in their decline (Wright et al. 2020). Common blackbirds are ranked third by monitoring from the car, but first in the other lists. This is not unexpected since they had the highest predicted roadkill rate, 12 individuals/km/year, in the model of Grilo et al. (2020) and are among the most frequently killed bird species in Western Europe (Erritzoe et al. 2003). Even transect data must be interpreted with care. Carcass persistence times and detection depend on size, with smaller animals being removed faster by scavengers (Santos et al. 2011; Teixeira et al. 2013; Ratton et al. 2014). Detection probability of larger mammals can also be influenced since they are more likely to be removed by maintenance workers or during police intervention at the site of an accident. Data collected by these services can be an important addition to the data collected by citizen scientists. Although proven to be a valuable data source (Grilo et al. 2009) additional steps need to be taken in Flanders to collect and centralise this data.

As expected, the ranking of victims collected as presence only data differs from the rankings in the transect data: presence only data show a clear bias to larger species, but possibly also species which are perceived as more interesting. Number two in the presence only data ranking is Red fox, which ranks only 6^th in transects by car, and 9^th in transects by bike. Foxes are infrequently seen alive, so, an encounter with a dead fox is for many people special enough to report. The number three, Red squirrel ranked 6^th in transects by car and 5^th in transects by bike. The Brown rat, the species encountered most frequently as roadkill (with exception from the Hedgehog) in transects by bike was only ranked 8^th in the presence only data list. This indicates that due to reporting bias the presence only data should not be used to determine which species are killed the most in traffic.

From 2010–2020 there is a strong increase in the number of non-roadkill mammal observations registered on waarnemingen.be but no significant increase in registered roadkill mammal observations. It is known that retention of volunteers can be challenging (Pocock et al. 2014; Shilling et al. 2015, 2020) but the number of observers registering roadkill has never been higher than the past 3 years (see Fig. 2) and their retention time on the waarnemingen.be platform exceeds 6 year. Volunteer participation depends also on repeated communication about the project. Over the last 3 years, our own communication channels mentioned the project ‘animals under wheels’ in 23 newsletters, we provided 15 contributions to written magazines, made 2 promotion videos and contributed to 10 national symposiums. Mainstream media wrote 47 articles about the project, and we gave 20 radio and 3 TV interviews (overview in Jacobs et al. (2021)) on the subject. This indicates that the absence in increase in registered roadkill mammals is not due to a reduction in observers/search effort but we believe that this is a strong indication that the number of roadkill is diminishing. Additional standardised collected data could confirm/refute this hypothesis. If this reduction is caused by effective road mitigation such as fencing, when possibly combined with crossing structures or animal detection systems (Rytwinski et al. 2016) this reduction does not reflect a decrease in population but a decrease in wildlife victims due to the mitigation measures. However, it might also reflect a reduction in abundance of (a number of) mammal species in Flanders that are most prone to being killed by vehicles.

Our species specific linear regression models indicate that 8 out of 17 mammal species have a significant change in proportion of roadkill victims through time. The number of reported roadkill victims of Mustela putorius, the Western polecat, declines, with the steepest significant slope of all species (slope = -0.450). The polecat is suspected to be in decline in Belgium, and also in most neighbouring countries (Croose et al. 2018) and there are indications this decline was already present from 1998–2010 (Van Den Berge and Gouwy 2012).

The proportion of victims of the seven other species are increasing over the years. Two species are (recently) recolonising (parts of) Flanders after a period of absence: Eurasian beaver (Swinnen et al. 2017) and Wild boar (Rutten et al. 2019). Roe deer has increased in range and numbers significantly since the 70’s (Casaer and Huysentruyt 2016), Beech marten, is doing the same the last decades (Van Den Berge 2016) and more recently, Badgers are also expanding from their last stronghold (Van Den Berge et al. 2017). Although the increase in population density is not quantified, we assume that this translates in higher relative roadkill numbers. The increase of the Eurasian hare was unexpected since the species was recently added as vulnerable to the red list of the Netherlands (bordering Flanders) (van Norren et al. 2020). However, for Flanders no monitoring scheme is in place. For Wood mouse we have no knowledge of population monitoring. This is a small-bodied species resulting in low carcass retention times (Santos et al. 2011; Ratton et al. 2014) and they were recorded relatively infrequently indicating that these results have to be interpreted cautiously. Remarkable is that the number of reported European hedgehog roadkill remains stable from 2010–2020. Until 2018, a strong decrease was occurring, but in 2019 and 2020 the proportion abruptly increased and was again at the 2010 level. This increase is currently unexplained but a fast recovery of the populations seems unlikely. There are reports of an unknown disease the last few years in Hedgehogs, possibly this also influences behaviour and making Hedgehogs more sensitive to being killed by cars.

Species distribution maps can be consulted at www.waarnemingen.be and additional info in Verkem et al. (2003). Linear regression models were also performed for the period of 2010–2019 since the global pandemic of the coronavirus disease (COVID-19) in 2020 resulted also in Flanders in confinement measures which are expected to have affected the search effort and the number of animals killed (Bíl et al. 2021; Driessen 2021). All trends remained similar, with the exception of the European hare, where the increase became non-significant.

Although the seasonal patterns are based on the rough data, without any correction for search effort within or between years, patterns of the same species are (highly) consistent. We expect that the large amount of data smoothens smaller inter- and within-year variation in search effort of individual observers. However, major events are detectable. In Flanders, there was a strict ban on non-necessary (car)travel from the 18^th of March 2020 to the 8^th of June 2020 due to the COVID-19 pandemic. Apart from the lives of wildlife this would have saved (Bíl et al. 2021; Driessen 2021), also very few observers were on the road to quantify this effect. Determining which of both factors was the most important is not possible using presence only data. For species in which the peak period of kills overlaps with the confinement measures, such as Western polecat, the seasonal pattern of 2020 is clearly affected. Knowing the roadkill patterns can help to protect specific species of interest by using specific warning signs, and (temporal) road closure can even increase habitat quality (Whittington et al. 2019). Although no age or sex of the individuals was recorded in most cases, most peaks in roadkill density are presumed to be linked to increased movement because of mating or juvenile movements and dispersal (Carvalho et al. 2018; Raymond et al. 2021).

We show that roadkill monitoring using citizen scientists can generate informative results. However, this is not the endpoint. Data collected during the ‘animals under wheels’ project also contributed to the mitigation of local mortality hotspots. Furthermore, the data can be consulted by policy makers and a number of questions were asked in the Flemish Parliament concerning wildlife roadkill, indicating that the problem is acknowledged at the political level.