Following the launch of a national roadkill awareness campaign in 2013, the EWT gathered roadkill data using four different approaches: (i) developing a form-based reporting system on a website (ii) soliciting historical records of roadkill incidents (iii) developing a smartphone-application called “EWT Road Watch” and (iv) using social media such as LinkedIn, Twitter and Facebook to attract interest. All of these methods relied on data collected by lay people, satisfying the criteria for citizen science. From these four methods, two main types of data collection strategies emerged: occasional/random observations (187 individuals), and data from regular observers (nine individuals) that each provided >50 roadkill reports. Regular observers were not trained in data collection but were working in the Zoology/Conservation sector and thus their data were considered to be accurate. Road patrol staff received annual training (conducted by the EWT) in species identification, collection of roadkill data, and the taking of photographs. Data could only be verified if a photograph was submitted with the report.

In parallel to citizen scientist data collection, road patrol agencies were trained by EWT staff to conduct regular road transects on set routes. Transects were conducted on the N3 highway from Johannesburg to Durban along a total length of 431 km, driven four times a day every day (twice in each direction); teams were allocated to six shorter sections to ensure the whole distance was covered effectively. Once discovered, carcasses were removed from the road to avoid recounts (Collinson et al., 2014; Guinard et al., 2015). Road patrols took place from July 2011 to November 2014.

Due to the small number of reports pre-2011 and in 2015, we conducted the following analyses on the data from 2011 to 2014 only.

Within each group of reporters (regular vs. occasional and road patrol vs. citizen scientists), we defined “repeated sampling” as a roadkill that was reported for the same species within a 2-day period, <150 m from one another. This resulted in two reports from the occasional reporter's dataset and 22 from the regular reporters' datasets being removed from the final dataset. There were no repeated samples by citizen scientists on the N3 since the road patrol removed carcasses from the road once detected.

The final dataset comprised a total of 2,642 roadkill reports, 1,647 from regular reporters, 786 from occasional reporters and 209 from the N3 road patrols (see Figure 1 for spatial distribution and Supplementary Materials S1, S2). A further 183 roadkill reports from citizens were located on the N3 surveyed by the road patrol.

Each species in the dataset was categorized according to taxonomic order. Domestic species (cat Felis catus, dog Canis lupus familiaris and livestock) were pooled into a group labeled “domestic,” and unknown/ unidentifiable mammal species were grouped into “Mammalia”. Three body mass classes were defined to account for carcass detection probability, using the average adult female body mass for African mammals (Skinner and Chimimba, 2005): very small (<2 kg), small (2–10 kg) and medium to large (>10 kg). Unknown species were assigned to a size class labeled “unknown.” Based on the National Red List of mammals in South Africa (Child et al., 2016) each species in the dataset was assigned to a Red List threat category. Domestic species and generic Mammalia were assigned a Not Determined (ND) Red List status.

Along the same section of the N3 highway, trained road patrols and citizen scientists (from both occasional and regular reports totaling 183 reports) reported a total of 31 and 35 mammalian species, respectively, across eight and nine taxonomic groups (Supplementary Material S2). The vast majority of the road patrol reports were of small and medium sized species (Figure 2A), in NT and ND Red List categories (Figure 2B). The distribution of Red List category (χ² = 21.7, df = 2, p < 0.001) from road patrol reports was significantly different from that reported by citizen scientists. Citizen scientists reported more roadkill of medium-size and species falling in the Red List category EN (Figures 2A,B). Additionally, they reported less very-small and NT species (Figure 2B). The frequencies of taxonomic groups reported differed significantly between road patrols and citizen scientists (χ² = 26.9, df = 8, p < 0.001), the latter reporting more Carnivora and domestic mammals and fewer Rodentia and unknown mammals than road patrols (Figure 2C).

The top three roadkill species reported by road patrols (unknown rabbit sp Leporidae sp., 22.1%; domestic dog Canis lupus familiaris, 11.5%; and black-backed jackal Canis mesomelas 10.5%; Supplementary Material S2) were among the top four species reported by citizen scientists, followed by serval (Leptailurus serval, 14.8%).

Using the KDE+ method, we identified eight and 11 significant clusters of roadkill from the road patrol and citizen scientist reports respectively. Using a threshold strength of 0.4 to define biologically relevant clusters, seven road patrol clusters were relevant as were eight citizen scientist clusters. These clusters represent 10.9% (n = 15) and 12.1% (n = 21) of the roadkill reported by road patrol and citizens respectively. Most of these clusters (four from road patrols and seven from citizens) were located within a section of ~41 km along the northern part of the surveyed road (Figure 3), representing ~10% of the total distance surveyed (431 km).

Data quality gathered by citizen scientists can be compromised (Floridi, 2012) due to inaccuracies in reporting and lack of scientific understanding (Batini and Scannapieca, 2006) and thus should be interpreted with caution. Our results suggest that data collected by ad hoc citizen scientists (i.e., occasional reporters) can be as accurate as those obtained by trained or informed reporters in terms of broad spatial patterns and species identification. Citizen scientists (i.e., occasional and regular reporters) who gathered roadkill data in South Africa between 2011 and 2014, largely agreed with the data from that of trained road patrols. Furthermore, the regular, informed reporters conveyed similar roadkill patterns as the occasional reporters.

Caro et al. (2000) noted during roadkill counts in California that small mammals (<10 kg) are more difficult to see when driving at normal speed (~100 km/h⁻¹). Furthermore, these smaller species are often quickly removed from roads by scavengers or become problematic to identify by observers as they become “flattened” by vehicles using the roads (Hels and Buchwald, 2001), thus making the recording of roadkill less accurate. In the case of our study, South African citizen scientists were less likely to report smaller species than the road patrols along the N3 highway. This is in line with our assumption that citizen scientist data would be biased toward larger species (>10 kg). In addition, citizen scientists might also lose motivation or put less effort in to reporting unidentified or “low profile” species, assuming that the information is not scientifically valuable (Lukyanenko et al., 2016). Our data show that road patrols frequently report a high number of very-small and unidentified species. This is likely due to them driving at slower speeds (<60 km/h⁻¹) as well as the training they receive from the EWT in species identification and collecting roadkill data (Collinson, 2016, pers. obs.).

A third of the roadkill clusters identified by occasional reporters were found in locations that agreed with those of regular reporters. The same was true when comparing data from road patrols to both citizen scientist reports (occasional and regular) as well as those of regular and occasional reports. Clusters of roadkill (i.e., roadkill hotspots) were concentrated on the same road sections in South Africa which represents 0.3% of the road network analyzed (~170,000 km). By comparison, 70% of clusters were not close to the clusters identified using data from the other reporting group (road patrol vs. citizens or regular vs. occasional reporters). These most likely resulted from the difference in sampling efforts by both types of citizen reporters. In our opinion, the spatial extent of the area surveyed by the 187 occasional observers (~765,000 km of the entire road network in South Africa; Karani, 2008) is expected to be larger than the one sampled by the nine regular observers (~170,000 km of road); smaller and less utilized roads are likely to have received a different sampling effort from both user groups. Our sample size from the occasional reporters (n = 505) was insufficient in terms of monitoring the country's road network, and consequently, a larger sample size is required to enable cluster analysis of the occasional data. We propose that the clusters identified as “roadkill hotspots” could be the focus of further and more detailed study that concentrate on the fine-scale patterns of roadkill and factors potentially responsible for the high intensity of collisions. Thus, while citizen scientists may not identify all the clusters noted by trained observers, these data can provide an early warning system for potential roadkill hotspots and data management (Shilling et al., 2015).

The role of citizen science in research and monitoring is increasing globally as a data collection tool (Lukyanenko et al., 2016) despite skepticism of the data produced by non-experts (Swanson et al., 2016). Our analysis demonstrates that the assumption that we target only “educated citizens” for the collection of roadkill data (Batini and Scannapieca, 2006) appears to be unnecessary. In this large-scale study of citizen science data collection patterns in South Africa, we conclude that the biases that may be present in our data (Floridi, 2012) are not significant enough to ignore the immense value of citizen science projects. Our study demonstrates that ad hoc citizen science has the potential to map roadkill occurrence and identify hotspots in a reliable and robust manner compared to that of trained road patrols and informed reporters. However, we propose that ad hoc citizen science data should only be used to identify general patterns or trends. As Lukyanenko et al. (2016) states, “Citizen science is about writing a story where citizens contribute to the plot. Experienced researchers should then assume the role of directing the actors and writing the dialogue.”

We outline below recommendations to improve the accuracy, sampling effort and the motivation of the citizen scientist.

(1) Data accuracy: Our dataset did not allow for species identification verification, since not all data were submitted with a photograph. We therefore propose that all similar projects encourage the submission of a photograph that can be verified by an expert, or cross-referenced by other citizen scientists (Swanson et al., 2015, 2016).

(2) Sampling effort: The sampling effort of data collected by both the citizen scientists and the road patrols was not recorded and we were therefore unable to correct for sampling effort bias; this leads to potential bias in both spatial distribution and species reporting. We therefore propose that (where possible) sample effort is encouraged with data submitted on not only where roadkill occurs but also roadkill absence (Shilling et al., 2015). For example, a smartphone application can either automatically record distances driven, or prompt users to report these data.

(3) Training and feedback: Citizen scientists contribute data because they want to make a difference and learn something new (Bonney et al., 2009; Silvertown et al., 2013; van der Wal et al., 2016). Thus, training opportunities (either online or in person), and feedback (not simply in scientific journals) can be invaluable ways of retaining and attracting citizen scientists across the globe. Face-to-face training will also provide the opportunity to explain the value in submitting all data, including the reporting of small and/or unidentifiable carcasses.

(4) Vary the tools: Not all potential citizen scientists are likely to prefer the same method of collecting data. Thus, some individuals will be engaged by an interactive website (Swanson et al., 2015, 2016), while others prefer a smartphone app, and another group may prefer email communication (as was the case in this study).

Despite the limitations associated with reporting efforts, the EWT's citizen science project has established the first national database for animal road mortalities. This will guide future management decisions on mitigating the negative impacts of roads and provide a platform from which future studies can be designed.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.