Accessing the Accuracy of Citizen Science Data Based on iNaturalist Data

The number of science projects with citizen (volunteer) participants is increasing nowadays. Despite some advantages citizen science have which directly influence science fields, the largest weakness of citizen science is the issue of the reliability of data collected by non-specialists. In this study, we assess the reliability of data collected by citizens and identify the factors that impede the credibility. We selected two species, Pelophylax chosenicus and Dryophytes suweonensis, which have specific habitat, breeding and brumation positions. The location data for both species were collected by the global citizen engagement project “iNaturalist”. As a result, 89.3% for P. chosenicus and 37.1% for D. suweonensis of the location data were estimated to be erroneous. No difference was observed in the data accuracy between experts and citizens for D. suweonensis because the mating call is the main classification key for the species. On the other hand, a significant difference was confirmed in the case of P. chosenicus, whose external characteristics are its classification key. Our study shows that efforts to improve reliability, including appropriate species selection and survey methods, are necessary to use citizen participation data in science research.


Introduction
Citizen science is an expression that denotes a scientific project conducted with the participation of citizens or non-specialists and mainly includes repeated and long-term mapping and monitoring [1,2]. Although the concept of citizen science was first used by Alan Irwin in 1995, citizens have been involved in scientific projects long before then [3]. The most famous example is The Christmas Bird Count (CBC) [4]. CBC is the oldest citizenled monitoring project, which was started in 1900 by the National Audubon Society (NAS). Each year, more than 50,000 observers from 50 regions participate, and the collected data have been published over hundreds of cases [4]. In addition, there are numerous citizenparticipating science projects around the world, such as Stall Catcher, eBird, iNaturalist, and Nurturing (details can be found in Scistarter.org and [2]).
Citizen science has the advantage of being able to obtain vast amounts of data over multiple regions over long periods of time [1,5,6]. Therefore, citizen science has been utilized in various research fields, such as climate change, biological monitoring, conservation, and modeling [2,7]. On the other hand, the low reliability of the data collected by non-specialists is one of the problems with citizen science [2,6]. For example, the misidentification of species or location errors caused by non-experts undermine the reliability of such data. Until recently, this reliability issue served as a major barrier to applying citizen participation data to the scientific field [6].
Several methods have been attempted to increase the accuracy and reliability of the data from citizen science. A typical example is prior education in which, with just one day of education, citizens' ability to classify species improved [8]. In addition, when citizens selected "easy" research topics, the accuracy of the data was 17% higher than that of the data collected for "difficult" topics [8]. Moreover, research design, model building, and calibration methods were developed to ensure the reliability of the data collected by nonexperts [1,9]. However, the accuracy of the data collected by citizens was still far lower than that of experts [8].
In this study, we analyze the accuracy of the data collected to evaluate the credibility of citizen science. First, the accuracy of the data uploaded to "iNaturalist", a representative citizen science project, is analyzed. After that, the accuracy, reliability, and usability of the data are analyzed and compared with those of the experts. Beyond simply evaluating the reliability of such data, we aim to provide important evidence that could contribute to the advancement of citizen science.

Species Selection
The most important factor for a reliability analysis study is the selection of subjects. We focused on selecting target species whose natural and mis-recorded habitats could be differentiated to determine accuracy. Specifically, we selected species whose (1) habitat, breeding, and brumation sites are constant throughout the year and (2) can be easily observed. Among the 16 amphibians (Caudata: 4; Salientia: 12) inhabiting in South Korea [10], the black-spotted pond frog (Pelophylax chosenicus) and the Suweon tree frog (Dryophytes suweonensis) are species only found in rice paddies and related environments in the western lowlands [11][12][13][14][15]. Since the habitat of the two species can be specified as rice fields, they were selected as target species.

Distribution Data Collection
iNaturalist is a representative international social networking resource launched in 2008, to which people from all over the world can easily upload and share data, such as locations, photos, videos, and sounds of numerous creatures [16,17]. We collected location data of P. chosenicus and D. suweonensis uploaded to iNaturalist to verify the accuracy and reliability of the data collected through citizen participation [18]. We only used data that included photos and could be accurately identified. In addition, we excluded from the analysis multiple records documented at the same site. Since data are uploaded and modified in real time on information sharing sites and, therefore, subject to continuous change as research progresses, we collected and analyzed data uploaded only by April 2021. To verify the reliability of the data collected by iNaturalist, we compared them with the data published in previous research papers, including the author's data [19].

Data Categorization and Analysis
We classified the collected location data to 4 types: (1) "Agriculture" is an environment created for paddy farming, such as paddy fields, waterways, and agricultural ponds; (2) "Forest" is an environment that includes mountains or valleys; (3) "Wetland" was used for reservoirs and rivers; and (4) "Other" included environments unrelated to amphibian habitats, such as urban centers, buildings, and the sea [20]. We analyzed the location type and frequency of the target species using the chi-square test. Prior to analyzing the differences according to the investigator, a normality test was performed on the comparison between the expert's data and the iNaturalist's data, and both showed a result of p < 0.001. Then, we compared the data differences by species and investigators using the Kolmogorov-Smirnov test, which is a non-parametric statistic. We used SPSS 26.0 (IBM, New York, NY, USA) to run statistical tests and the significance level was set to 0.05.

Data Collected by Experts
The number of location points of P. chosenicus and D. suweonensis collected by experts was 79 and 162, respectively ( Figure 1). In P. chosenicus, 91.1% of the locations were "Agriculture", followed by "Wetland" at 8.9% (chi-square test, χ 2 = 53.481, df = 1, p < 0.001). On the other hand, there were no "Forest" and "Other" for the location of P. chosenicus. In the case of D. suweonensis, rice fields accounted for the highest percentage of the locations at 79.6% (χ 2 = 262.346, df = 3, p < 0.001), followed by "Forest" at 12.3%, "Other" at 7.4%, and "Wetland" at 1 site. For the "Other" type, 9 sites in the city center and 3 sites in the sea were included.

Data Collected by Experts
The number of location points of P. chosenicus and D. suweonensis collected by experts was 79 and 162, respectively ( Figure 1). In P. chosenicus, 91.1% of the locations were "Agriculture", followed by "Wetland" at 8.9% (chi-square test, χ 2 = 53.481, df = 1, p < 0.001). On the other hand, there were no "Forest" and "Other" for the location of P. chosenicus. In the case of D. suweonensis, rice fields accounted for the highest percentage of the locations at 79.6% (χ 2 = 262.346, df = 3, p < 0.001), followed by "Forest" at 12.3%, "Other" at 7.4%, and "Wetland" at 1 site. For the "Other" type, 9 sites in the city center and 3 sites in the sea were included.

Data Collected by Experts
The number of location points of P. chosenicus and D. suweonensis collected by experts was 79 and 162, respectively (Figure 1). In P. chosenicus, 91.1% of the locations were "Agriculture", followed by "Wetland" at 8.9% (chi-square test, χ 2 = 53.481, df = 1, p < 0.001). On the other hand, there were no "Forest" and "Other" for the location of P. chosenicus. In the case of D. suweonensis, rice fields accounted for the highest percentage of the locations at 79.6% (χ 2 = 262.346, df = 3, p < 0.001), followed by "Forest" at 12.3%, "Other" at 7.4%, and "Wetland" at 1 site. For the "Other" type, 9 sites in the city center and 3 sites in the sea were included.

Comparison between Expert and Citizen Data
The location data of P. chosenicus differ significantly between experts and iNaturalist (Kolmgorov-Smirnov test, n = 293, z = 6.107, p < 0.001). On the other hand, there was no difference in the location data of D. suweonensis between experts and iNaturalist data (n = 197, z = 0.867, p = 0.440).

Discussion
The accuracy of the data showed a clear difference depending on the target species and investigator. Without professional education, people can easily confuse P. chosenicus with P. nigromaculatus, because the two are very alike in their body size, coloration, and patterns [21]. Despite their external similarities, P. nigromaculatus inhabits forests and their surroundings, so their ecological niche is clearly distinguished from that of P. chosenicus [15]. The high percentage of "Forest" among P. chosenicus data from iNaturalist is presumably because the location data of morphologically similar true frogs were uploaded. On the other hand, the locations collected by experts reflected the habitat characteristics of P. chosenicus well [15]. It shows that expertise in the target species can have a significant impact on the accuracy of the data collected. Therefore, since morphological similarity can be a problem for species classification, actively referring to species characteristics, such as mating calls, for classification and detection-which require no prior training-will help to increase the accuracy of the data [22]. Dryophytes suweonensis is also very similar in color, shape, size, and skeleton to its related species, H. japonica, so even experts have limitations in accurately discerning the two species based on the external shape alone [23]. Although a clear difference was identified between the two species in size [24], it was difficult for citizens to distinguish the two species by size in the field. The authors of [25] discovered, first, a frog making a unique mating call, and then the species was recorded as a new species, D. suweonensis, through additional studies. In our results, there was no significant difference between the data of citizens and experts. Maybe it is due to the use of the mating call, which is an "easy" and "clear" feature for identification. In this case, data collected by citizens can be used for research. On the other hand, mis-recorded locations were also similarly identified in both citizen and expert data. In this case, we determined the mis-recorded data from both citizens and experts may be related to errors in recording and uploading, rather than in species identification.
Many location data of the two species uploaded to iNaturalist were mis-recorded. There are three possible causes for this mis-record, and the first one is related to data uploading. In iNaturalist, location data are based on where the date are uploaded. If data are uploaded from a place other than the location where the species was originally found or from its habitat, there is a possibility of mis-recording. Even in the data collected by experts, there was a difference in the general distribution of the target species, and it is possible that these errors also occurred during the uploading and recording process. Second, if the uploading was attempted from a point where the Internet environment is poor, the location of the point where the Internet conditions improved may be recorded, instead of the original location. The first and second problems can arise because many citizens are not exactly familiar with how GPS works. A study by [8] confirmed that citizens' recording accuracy could be increased by more than 80% through simple training. Errors can also be reduced if the uploader directly records the location on the map. The third cause is the change in the environment of the recorded location points over time. For example, 11.9% of agricultural land in South Korea was converted from 21,444 km 2 in 1985 to 18,888 km 2 in 2000 for other uses, such as cities. In particular, the paddy area decreased by 20.9% from 1,016,000 ha in 2003 to 808,000 ha in 2014 [26,27]. At the time of discovery, the location of the two species was rice paddy field, but there is a possibility that it may have been seen as a mis-record because it is now used for urban or other purposes ( Figure 3). As a solution, we propose a method that captures and stores a photo of the location at the time of data upload, including habitat, so that it can be monitored over time. The biggest advantage of citizen science is that it allows to collect data at the level of big data. However, if reliability is not secured through sufficient verification, the collected data may be limited for use in research. Although the data were collected from two species with a known habitat and location, the reliability was insufficient for scientific application. For species whose habitat cannot be specified, reliability verification and correction of data are also limited. In the end, this problem must be solved in the process of collecting location data. The use of protocols, training, and volunteers suitable for research makes it possible to collect data at the expert level [2,6,8]. For example, professional training is required for participants, and oversight by experts is essential for the data collected afterwards. In addition, easy target selection, GPS training, statistical correction of collected data, and the use of computing tools can be a way to increase data reliability [2,6,8,9].
The merits of citizen science are valuable in terms of the quantity of data required by recent trends. Therefore, if errors that occur in data collection or uploading are identified and efforts to minimize them are continued, it will be possible to secure a large amount of data and further utilize it in various research fields.   The biggest advantage of citizen science is that it allows to collect data at the level of big data. However, if reliability is not secured through sufficient verification, the collected data may be limited for use in research. Although the data were collected from two species with a known habitat and location, the reliability was insufficient for scientific application. For species whose habitat cannot be specified, reliability verification and correction of data are also limited. In the end, this problem must be solved in the process of collecting location data. The use of protocols, training, and volunteers suitable for research makes it possible to collect data at the expert level [2,6,8]. For example, professional training is required for participants, and oversight by experts is essential for the data collected afterwards. In addition, easy target selection, GPS training, statistical correction of collected data, and the use of computing tools can be a way to increase data reliability [2,6,8,9].
The merits of citizen science are valuable in terms of the quantity of data required by recent trends. Therefore, if errors that occur in data collection or uploading are identified and efforts to minimize them are continued, it will be possible to secure a large amount of data and further utilize it in various research fields.