Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A generalized approach for producing, quantifying, and validating citizen science data from wildlife images

A generalized approach for producing, quantifying, and validating citizen science data from... Special Section: Moving from Citizen to Civic Science to Address Wicked Conservation Problems A generalized approach for producing, quantifying, and validating citizen science data from wildlife images ∗ ∗ ∗ Alexandra Swanson, † ¶ Margaret Kosmala, ‡ Chris Lintott,† and Craig Packer Department of Ecology, Evolution and Behavior, University of Minnesota, Saint Paul, MN 55108, U.S.A. †Department of Physics, University of Oxford, Denys Wilkinson Building, Oxford, OX1 3RH, U.K. ‡Current address: Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA 02138, U.S.A. Abstract: Citizen science has the potential to expand the scope and scale of research in ecology and con- servation, but many professional researchers remain skeptical of data produced by nonexperts. We devised an approach for producing accurate, reliable data from untrained, nonexpert volunteers. On the citizen science website www.snapshotserengeti.org, more than 28,000 volunteers classified 1.51 million images taken in a large-scale camera-trap survey in Serengeti National Park, Tanzania. Each image was circulated to, on average, 27 volunteers, and their classifications were aggregated using a simple plurality algorithm. We validated the aggregated answers against a data set of 3829 images verified by experts and calculated 3 certainty metrics—level of agreement among classifications (evenness), fraction of classifications supporting the aggregated answer (fraction support), and fraction of classifiers who reported “nothing here” for an image that was ultimately classified as containing an animal (fraction blank)—to measure confidence that an aggregated answer was correct. Overall, aggregated volunteer answers agreed with the expert-verified data on 98% of images, but accuracy differed by species commonness such that rare species had higher rates of false positives and false negatives. Easily calculated analysis of variance and post-hoc Tukey tests indicated that the certainty metrics were significant indicators of whether each image was correctly classified or classifiable. Thus, the certainty metrics can be used to identify images for expert review. Bootstrapping analyses further indicated that 90% of images were correctly classified with just 5 volunteers per image. Species classifications based on the plurality vote of multiple citizen scientists can provide a reliable foundation for large-scale monitoring of African wildlife. Keywords: big data, camera traps, crowdsourcing, data aggregation, data validation, image processing, Snapshot Serengeti, Zooniverse Una Estrategia Generalizada para la Produccion, ´ Cuantificacion ´ y Validacion ´ de los Datos de Ciencia Ciudadana a partir de Im´agenes de Vida Silvestre Resumen: La ciencia ciudadana tiene el potencial de expandir el alcance y la escala de la investigacion ´ en la ecolog´ıa y la conservacion, ´ pero muchos investigadores profesionales permanecen esc´epticos sobre los datos producidos por quienes no son expertos. Disenamos ˜ una estrategia para generar datos precisos y fiables a partir de voluntarios no expertos y sin entrenamiento. En el sitio web de ciencia ciudadana www.snapshotserengeti.org mas ´ de 28, 000 voluntarios clasificaron 1.51 millon ´ de imagenes ´ que fueron tomadasenuncenso agranescaladecamaras ´ trampa en el Parque Nacional Serengueti, Tanzania. Cada imagen llego, ´ en promedio, hasta 27 voluntarios, cuyas clasificaciones se conjuntaron mediante el uso de ¶Address for correspondence: Department of Physics, University of Oxford, Denys Wilkinson Building, Oxford OX1 3RH, U.K. email ali@ zooniverse.org Paper submitted February 27, 2015; revised manuscript accepted August 19, 2015. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Conservation Biology, Volume 30, No. 3, 520–531 2016 The Authors. Conservation Biology published by Wiley Periodicals, Inc. on behalf of Society for Conservation Biology. DOI: 10.1111/cobi.12695 Swansonetal. 521 un algoritmo de pluralidad simple. Validamos el conjunto de respuestas frente a un juego de datos de 3, 829 imagenes ´ verificadas por expertos y calculamos tres medidas de certeza: nivel de concordancia entre las clasificaciones (uniformidad), fraccion ´ de clasificaciones que apoyan al conjunto de respuestas (fraccion ´ de apoyo) y fraccion ´ de clasificadores que reportaron “nada aqu´ı” en una imagen que al final se clasificoc ´ omo que s´ıten´ıa un animal (fraccion ´ en blanco). Estas medidas se usaron para estimar la confianza de que un conjunto de respuestas estuviera en lo correcto. En general, el conjunto de respuestas de los voluntarios estuvo de acuerdo con los datos verificados por los expertos en un 98 % de las imagenes, pero la certeza variosegun la ´ ´ ´ preponderancia de la especie, de tal forma que las especies raras tuvieron una tasa mas alta de falsos positivos y falsos negativos. El analisis de varianza calculado facilmente y las pruebas post-hoc de Tukey indicaron que ´ ´ las medidas de certeza fueron indicadores significativos de si cada imagen estuvo clasificada correctamente o si era clasificable. Por esto, las medidas de certeza pueden utilizarse para identificar imagenes ´ para una revision ´ de expertos. Los analisis ´ de bootstrapping indicaron mas ´ a fondo que el 90 % de las imagenes ´ estuvieron clasificadas correctamente con solo ´ cinco voluntarios por imagen. Las clasificaciones de especies basadas en el voto de pluralidad de multiples ´ cient´ıficos ciudadanos puede proporcionar un fundamento fiable para un monitoreo a gran escala de la vida silvestre africana. Palabras Clave: c´amaras trampa, conjunto de datos, crowdsourcing, datos grandes, procesamiento de im´agenes, Snapshot Serengeti, validacion ´ de datos, Zooniverse son et al. 2010). However, these procedures take time and Introduction may waste potentially valuable information and volunteer effort. Alternatively, eBird and FeederWatch ask volun- Modern citizen science, the engagement of the general teers to report bird sightings, flag implausible reports, public in the process of science, has enormous potential and engage experts in validation of flagged contributions to expand the scope and scale of research in ecology (Bonter & Cooper 2012). Although this approach can and conservation. These fields have long benefited from reduce false positives of unusual sightings, it leaves no volunteer contributions to, for example, the Audubon way to verify plausible but erroneous entries. Society’s Christmas Bird Count, which dates back more than 100 years (Silvertown 2009). In the last decade, tech- Successful citizen science projects in astrophysics, nological advances have rapidly accelerated the number such as Galaxy Zoo (Lintott et al. 2008; Willett et al. and diversity of projects that include public participation 2013), Space Warps (Marshall et al. 2016), Milky Way (Silvertown 2009; Dickinson et al. 2010, 2012; Tulloch Project (Simpson et al. 2012; Beaumont et al. 2014), and et al. 2013). Andromeda Project (Johnson et al. 2015) rely on the judg- Online projects engage people to contribute data on ments of multiple volunteers to classify satellite and tele- scope imagery. Cyclone Center, a meteorological project, an extraordinary array of taxa around the world (e.g., asks multiple users to identify features in infrared satellite Firefly Watch, HerpMapper, International Waterbird Cen- images of storms (Hennon et al. 2014). Each of these sus, and Road Watch) and on weather and climate (e.g., projects applies algorithms to aggregate the responses Citizen Weather Observer Program). Increased internet and produces expert-quality data sets. Similar approaches connectivity now allows volunteers to upload species can be applied to wildlife and conservation-based citizen sightings on websites such as iSpot.org and immediately science projects that ask volunteers to identify animals in interact with dozens of other naturalists (Silvertown et al. photographs taken with camera traps. 2015). Integrating volunteer effort and emerging tech- Digital image collection from camera-trap surveys is a nologies expands the range of possibility in both basic rapidly expanding method (O’Connell et al. 2011) that is and applied research.However, broad-scale implementa- used to study rare and elusive species worldwide (e.g., tion of citizen science for research is hindered by con- Karanth & Nichols 2002; Dillon & Kelly 2007; Kelly et al. cerns about data quality. Many professional researchers 2008) and to survey animals across large spatial extents are skeptical of data produced by nonexperts, which (e.g., O’Brien et al. 2010; Kinnaird & O’Brien 2012; lowers publication rates and grant funding of citizen Bischof et al. 2014). However, as inexpensive digital science projects (Foster-Smith & Evans 2003; Dickinson cameras have proliferated, such surveys are increasingly et al. 2010; Bonter & Cooper 2012). Although individ- limited by human processing capacity. Engaging citizen ual contributors can be measurably worse than trained professionals (Foster-Smith & Evans 2003; Galloway et al. scientists to classify images can dramatically increase the 2006; Delaney et al. 2007; Gardiner et al. 2012), solutions amount of information researchers can extract from large are available for assuring quality control of volunteer data. data sets. Some projects train volunteers or require volunteers to We devised an approach to produce accurate, re- pass a competency test, whereas others discard data liable data from multiple untrained, nonexpert volun- teers classifying images from the Snapshot Serengeti from inexperienced or unreliable contributors (Dickin- Conservation Biology Volume 30, No. 3, 2016 522 Citizen Science Data Quality (www.snapshotserengeti.org) camera-trapping study. overused and thus reduced the efficiency of volunteer We framed our analyses in terms of accuracy and effi- contributions. ciency to provide guidelines for optimizing the trade-off Each image was circulated to multiple users and retired between effort (of volunteers and experts) and accuracy. after meeting the following criteria: the first 5 classi- Because conservation studies often target rare species, fications were nothing here (hereafter blank); 10 non- we also evaluated how measures of accuracy and effi- consecutive nothing-here classifications (hereafter blank ciency differed across species of contrasting rarity. Our consensus); or 10 matching classifications of species or overarching goal was to provide straightforward tools and species combinations, not necessarily consecutive (here- quantifiable metrics to test and validate citizen science after consensus). If none of these criteria were met, the data, thus providing biologists with a generalizable ap- image was circulated until it accumulated 25 species clas- proach for engaging citizen scientists to produce reliable sifications (hereafter complete). These values were cho- data for large-scale conservation and wildlife research. sen based on volunteer performance on existing Zooni- verse projects. Volunteers classified Snapshot Serengeti data faster than images were produced, and images were Methods recirculated for use in classrooms. As a result, the number of classifications for images containing animals ranged The Snapshot Serengeti Interface from 11 to 57 (mean = 26, median = 27). Snapshot Serengeti is hosted by the Zooniverse citizen science platform (www.zooniverse.org), which engages 1.5 million volunteers worldwide to participate in a broad Data Aggregation and Validation array of projects. Zooniverse volunteers are motivated largely by a desire to contribute to science (Raddick et al. We implemented a simple plurality algorithm to trans- 2010, 2013) and engage with research teams and one form the volunteer classifications for each image into a another in high-level scientific discussions via the Zooni- single aggregated classification. As described in Swanson verse discussion forums (Mugar et al. 2014). et al. (2015), we first evaluated the median number, n, On www.snapshotserengeti.org, members of the gen- of different species reported by all classifiers for that im- eral public viewed and classified images from a large- age. For simplicity in interpreting volunteer accuracy, we scale camera survey in Serengeti National Park, Tanzania limited our analyses here to the 94% of collected images (see Swanson et al. 2015 for survey details). From June with n = 1. We then identified the species present as the 2010 to May 2013, the camera survey accumulated 99,241 n species with the most classifications. For example, if camera-trap days and produced 1.2 million image sets an image with n = 1 had 15 total classifications, with 7 (each image set contained 1–3 images taken in a single classifications of wildebeest, 5 classifications of buffalo, burst over approximately 1 s). Within 3 d of launching the and 3 classification of topi, the aggregated classification website, volunteers contributed 1 million species classi- would be wildebeest. If the same image had n = 2, the ag- fications and processed an 18-month backlog of images gregated classification would be wildebeest and buffalo. (Swanson et al. 2015). We calculated the number of individuals present for each Users were asked to identify species, count the number identified species by the median count (rounded up) of of animals(binnedas1,2,3,4,5,6,7,8,9,10, 11–50, all raw classifications for that image and the interquartile and 51+ individuals), and characterize behaviors in each range of counts reported by all classifiers for a given image set (Fig. 1). Volunteers followed a simple tutorial image. that explained the online interface, but they were not We calculated 3 measures of certainty or confidence formally trained or tested for accuracy before contribut- for each image: evenness, fraction blanks, and fraction ing. We designed the interface to help guide people with support. Evenness was calculated from all classifications no background knowledge through the process of animal that were not blank for each image with Pielou’s evenness identification from 48 possible species and taxonomic index (Pielou 1966): −( p ln p )/ ln S,where S is i i i = 1 groups while providing a rapid route to classification for the number of different species reported by all volunteers more knowledgeable participants. New users filtered po- and p is the proportion of classifications received by tential species matches by morphological characteristics species i. When all classifications were in agreement, we such as horn shape, body shape, color, pattern, tail shape, assigned a value of zero. The maximum value for this or a general gestalt (e.g., “looks like an antelope or deer”). index is 1.0, indicating high disagreement among clas- More experienced users could select the species directly sifications. Fraction blank was calculated as the fraction from a list. A “nothing here” button allowed users to of classifiers who reported nothing here for an image that classify images without any visible animals, but we did was ultimately classified as containing an animal. Fraction not provide an I-don’t-know option because previous test- support was calculated as the fraction of classifications ing with this option with undergraduate volunteers on a that supported the aggregated answer (i.e., fraction sup- small-scale prototype indicated that such answers were port of 1.0 indicated unanimous support). Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 523 Figure 1. The Snapshot Serengeti website interface used for classifying species, counts, and behaviors in images from the camera-trapping survey: (a) primary interface with all available species options and (b) filters that help narrow users’ choices when classifying species. We compared overall plurality algorithm perfor- culating the proportion of times the aggregated answer mance to a baseline expert-verified data set of 3829 for species present was confirmed by expert classifica- randomly sampled images. This data set (from Swanson tions (reported as proportion correct) in the baseline et al. 2015) was produced by asking a panel of experts expert-validated data set. We calculated accuracy both for to review these images. The experts were individuals the resolvable images and for all images. We calculated who had undergone extensive formal training, passed species-specific accuracy as the likelihood of the aggre- qualification exams, or had years of experience identify- gated answer being correct (i.e., the likelihood of the ing African wildlife. Of these images, 135 were indepen- expert classifications confirming the given aggregated dently classified by >1 expert. In cases where experts answer) based on images in the expanded expert-verified disagreed with the results of the plurality algorithm or data set.We further evaluated species-specific accuracy had marked an image set as particularly difficult or impos- with respect to 2 types of error: false positives (species sible, A.S. and C.P. made the final authoritative species reported when not present) and false negatives (species identification. For species-specific analyses, we used an not reported when actually present). Because the false- expanded data set of 5558 images that included extensive negative analysis required images to be randomly sam- random sampling of images identified as rare species by pled with respect to the true answer, this analysis was the plurality algorithm to ensure their adequate represen- limited to the baseline expert-verified data. The false- tation (species-specific sample sizes are in Supporting positive analysis was applied to the expanded data set Information). In 0.8% of images, the panel of experts because it only required that images be randomly sam- agreed that no authoritative species identification could pled with respect to the aggregated answer. For each be made. Because the citizen science interface does not species, we evaluated the proportion of photographs allow for an impossible classification, the aggregated an- containing each type of error. We associated a group swers were technically wrong for these images because of species that were easy to identify with zero error, no reliable volunteer answer exists. Additional details of and then compared nonzero error rates to species com- the field study, classification interface, aggregation, and monness (defined as the total number of times a species validation are available in Swanson et al. (2015). was photographed) with simple linear regression on log- transformed variables on the remaining species. To assess the accuracy of counts, we first evaluated Accuracy the agreement among expert counts for the 135 images We compared the results of the plurality algorithm with with multiple expert classifications. We also calculated expert answers for species identification and animal count ranges for each image from the raw classifications counts. We evaluated overall algorithm accuracy by cal- andcomparedexpertagreementwiththese ranges.We Conservation Biology Volume 30, No. 3, 2016 524 Citizen Science Data Quality calculated count range as the interquartile range (i.e., 25th percentile to 75th percentile values) of counts re- ported by all classifiers for a given image. We evaluated how often expert counts fell within this range for all resolvable images in the baseline expert-verified data set. To assess image difficulty for effectively targeting expert review, we evaluated the 3 certainty measures against ac- curacy, classifying images as correct (plurality answer agreed with expert-verified data set), incorrect (plurality answer disagreed with expert-verified data set), or impos- sible (experts could not identify the species). To test the predictive power of these metrics, we performed a one- way analysis of variance (ANOVA) followed by post-hoc Tukey test (package stats in Program R) for each metric to evaluate whether mean evenness, fraction support, or fraction blanks differed significantly across correct (n = 5281), incorrect (n = 225), and impossible (n = 52) images from the extended expert-verified data set. We further used a simple linear regression of log-transformed mean species-specific values of evenness, fraction sup- port, and fraction blanks on the logarithm of the total number of times a given species was photographed to assess the effects of species commonness on the 3 mea- sures of difficulty. Efficiency We determined how many volunteer classifiers were needed to produce reliable species identifications by bootstrapping the plurality answer from raw classifications for images with expert-verified answers. We first excluded the expert answers from the raw data set. Then, for every image, we randomly sampled (with replacement) n classifications from the raw data set (20 iterations each) and applied the plurality algorithm to produce an aggregated answer and calculate an evenness score. In case of ties, one answer was randomly selected from the 2. For each n volunteers, we evaluated accuracy as the Figure 2. Example images from the Serengeti camera proportion of times the plurality algorithm agreed with survey and presented on the Snapshot Serengeti the expert-verified answer. Overall accuracy was cal- website illustrating situations in which: (a) species culated using images from the expert-verified data set. identification is impossible, (b) a precise count of We further characterized species as belonging to 1 of animals is impossible, and (c) animals present in 3 groups: high overall accuracy (i.e., low false-positive foreground and background that leads to a wide and false-negative rates), high false positives, and high range of individual counts by volunteers. false negatives. We calculated species-specific accuracy for species in each of these groups. Species-specific ac- and >0.5 and calculated the proportion of images in each curacy was calculated for the expanded data set as the group (<0.5 and >0.5) that were correct. probability of the aggregated answer being correct. Because images vary in identification difficulty, we fur- ther evaluated the number of classifiers needed to pro- Results vide an evenness score that reliably indicated a correct answer. For every iteration of the bootstrapping analysis Accuracy for every n classifications, we calculated the evenness score for a given image. At every additional classification, Experts identified species in 3800 of 3829 randomly sam- we split images into those with evenness scores of <0.5 pled images (99.2%), labeling 29 images as impossible Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 525 False Negative False Positive rodents(+) aardwolf jackal topi bushbuck(+) G gazelle other bird(+) reedbuck eland impala buffalo hartebeest elephant(+) T gazelle wildebeest waterbuck(*) warthog(*) vervet monkey(*) serval(*) secretary bird(*) porcupine(*) mongoose(*) male lion(*) leopard(*) human(*) hippopotamus(*) giraffe(*) baboon(*) zebra(+) aardvark(+) female lion(+) guinea fowl(+) cheetah(+) ostrich(+) spotted hyena(+) dik dik(+) kori bustard(+) rhinoceros(+) reptiles(*) hare(*) caracal(*) wildcat civet honey badger bat eared fox genet zorilla striped hyena 0.25 0.50 0.75 1.00 0.25 0.50 0.75 1.00 Fraction of Error Figure 3. Species-specific rates of false-negative (calculated relative to the randomly sampled expert-verified data set of 3829 images) and false-positive (calculated relative to the extended expert-verified data set of 5558 images) error for species identifications produced by the plurality algorithm (error bars, standard error calculated for proportions; ∗, species with zero total error; +, species with zero false-positive error). Note that x-axis is plotted on a square-root scale. Sample sizes are in Supporting Information. or unresolvable (e.g., Fig. 2a). The plurality algorithm were perfectly classified regardless of how often they agreed with experts on 3750 images, yielding 97.9% over- appeared in the data set (Fig. 3). Both error types were all agreement and 98.6% agreement on resolvable images. significantly higher for rare species, although stronger for Accuracy differed dramatically by species, ranging from false negatives (p  0.0001, r = 0.71, df = 13) than for 100% accuracy for giraffe (n = 87) and hippopotamus false positives (0.0001, r = 0.45, df = 26). (n = 28) to 70% for jackals and 50% for aardwolves (see Precise counts were unresolvable in many images (e.g., full list in Supporting Information). Fig. 2b). Experts agreed on the number of individuals Species-specific accuracy also varied in the type of er- only 74% of the time (of 135 photos), and the average ror (Fig. 3): species with high false negatives tended to range of expert answers spanned 2.5 bins. The median have fewer false positives. A few species had high rates count reported by the plurality algorithm was just as likely of both false positives and negatives, and these tended to to agree with experts as experts were to agree among be confused with each other. Confusion was typically themselves (75.2% of 3800 images). clustered within groups of species with similar sizes, The interquartile range reported by the plurality al- shapes, and colors (Supporting Information). Algorithm gorithm was generally precise. Volunteers agreed on a accuracy was generally related to rarity. Species with single count in 50% of images, and 86% of images had fewer photographs had higher rates of false negatives ranges of 3 bins (e.g., 4–6 animals) (Supporting Infor- and false positives (Fig. 4), although a subset of species mation). In images with a wide range of counts, animals Conservation Biology Volume 30, No. 3, 2016 526 Citizen Science Data Quality was ultimately resolvable (p < 0.0001 for all pairwise 1.00 ErrorType comparisons). Images that evoked a large proportion of False Negative nothing-here responses sometimes contained a partial or False Positive blurred view of an animal, making it particularly difficult to identify. Thus, some classifiers apparently preferred to state there was "nothing here" rather than to guess the animal’s identity. The vast majority of images were classified as easy— 0.10 showing high levels of agreement on species classifica- tion (Supporting Information). For example, half of all images had >87% agreement on the final answer, and only 6% of images did not attain a majority (i.e., no species had >50% of the classifications). Excluding classifications of nothing here, 36% of images had unanimous agree- 0.01 ment on the species classification. As with accuracy, rare species were significantly more difficult to identify than 10 100 1,000 10,000 100,000 Total Captures common species: regression of certainty scores versus log of total pictures showed higher fraction support (p Figure 4. False-positive and false-negative (log 2 2 < 0.0001, r = 0.397), lower evenness (p < 0.001, r = fraction of error) identification of animals versus 0.214, df = 46), and lower fraction blanks (p = 0.0036, species commonness (frequency of species appearance r = 0.17, df = 46) with increasing commonness (Sup- in the overall data set, given as log[total number of porting Information). pictures]). Linear regression was performed on log-log transformed variables and restricted to species with nonzero error rates. Species with an asterisk in Fig. 3 Efficiency were excluded from false-positive and false-negative Accuracy increased asymptotically with the number of analyses, and species in Fig. 3 marked with a plus volunteers (Fig. 6a). Overall, we achieved 90% accuracy were excluded from the false-positive analysis. on species identifications with 5 volunteer classifications and 95% accuracy with 10 classifications. The number of volunteer classifications needed for ac- appeared in both the foreground and background (Fig. curate answers differed with species (Fig. 6b). Images 2c), and the distribution of counts tended to be bimodal with easy species (characterized by low rates of false (Supporting Information), presumably due to some users negatives and false positives) were nearly always correct only counting animals in the foreground and others with just 1–3 classifiers. Images with high rates of false counting everything in the image. negatives (e.g., jackals, aardwolves, and topi) needed Evenness, fraction support, and fraction blank were more classifiers to achieve high accuracy rates, but ad- all excellent predictors of whether aggregated answers ditional improvement declined after about 10 classifica- were likely to be correct (Fig. 5). The post-hoc Tukey test tions. In contrast, for species with high false-positive rates revealed that evenness scores were significantly lower (such as the extremely rare civets, genets, and striped (i.e., answers were all skewed toward a single species) for hyenas), accuracy rates remained low even after 30 clas- images that were classified correctly than for images that sifiers. were incorrectly classified (p < 0.0001) or impossible Evenness scores were excellent predictors of an ac- (p < 0.0001). Similarly, fraction support was higher for curate classification (Fig. 6a & Supporting Information). images that were correctly identified than for images that After 3 classifications, images with evenness 0.5 (e.g., were incorrectly classified (p < 0.0001) or impossible to at least 2 of the 3 users agreed on the species) were97% classify (p < 0.0001). However, evenness stood out as likely to be correct. After 5 classifications, images with an the single best predictor of a correct classification: 99.9% evenness score of 0.5 were 99% likely to be correct. of images with evenness < 0.25 were correct and 99.8% Thus, evenness can quickly be used to identify images of images with evenness < 0.5 were correct. In contrast, requiring additional classifications. only 85% of images with evenness > 0.5 and 71% of images with evenness > 0.75 were correct. Although evenness (p = 0.157) and fraction support Discussion (p = 0.394) for incorrect and impossible images did not differ significantly, the fraction of nothing-here classifica- Snapshot Serengeti provides a case study in engaging tions differed significantly across all 3 groups, and frac- citizen scientists in rapidly and accurately processing tion blank was the best predictor of whether an image large volumes of ecological imagery. Unlike many other Conservation Biology Volume 30, No. 3, 2016 Fraction of Error Swansonetal. 527 Evenness 1 - FractionSupport FractionBlanks abb abb ab c 1.00 0.75 0.50 0.25 0.00 correct incorrect impossible correct incorrect impossible correct incorrect impossible Comparison to Expert Classifications Figure 5. Evenness (support level of agreement among classifications), fraction blanks (fraction of classifiers who reported “nothing here”), and fraction support (fraction of classifications supporting the aggregated answer) for images that were verified by experts and deemed to be correct (aggregated volunteer answer agreed with expert answer), incorrect (aggregated volunteer answer did not agree with expert answer), or impossible (experts could not determine the species present). All metrics are bounded between 0 and 1, fraction support is plotted as “1-fraction support” so that for all 3 metrics, scores closer to 1 reflect greater uncertainty. Boxplots marked with (a) have significantly different means than those marked with (b). citizen science projects (Dickinson et al. 2010), Snapshot ual experts (96.6%) when compared with the consensus Serengeti volunteers were neither trained nor required to expert assessments. Experts can make mistakes: a single demonstrate species identification skills. Instead, we en- field researcher flipping through hundreds or thousands gaged multiple volunteers for every task and aggregated of photographs can become fatigued and miss species or their answers to produce highly accurate data. Snapshot click on the wrong classification. Making a precise count Serengeti data were 97.9% accurate overall, whereas 85– was impossible in many images (Fig. 2b). Calculating 95% accuracy is reported for projects engaging trained a count range from multiple volunteers provided more volunteers (Galloway et al. 2006; Delaney et al. 2007; reliable count data than a single number reported by a Gardiner et al. 2012). Engaging multiple volunteers for single expert (recall that experts disagreed on counts for every image did not necessarily mean many volunteers 26% of images). for every image. By evaluating measures of certainty in When creating the expert-verified data, experts agreed volunteer answers and evaluating the relative contribu- that a small number of photographs (0.8%) were impos- tion of additional volunteer classifications, we provide sible to classify (Fig. 2a). These impossible photographs guidelines for researchers to target volunteer effort and accounted for 36% of the overall error because the Snap- expert effort to balance their data needs and available shot Serengeti interface did not allow users to mark im- human resources for specific projects. ages as such. However, the likelihood of an image truly being impossible to identify can be determined by the fraction of blanks reported in volunteer classifications. Accuracy Furthermore, even guesses provide information, such as distinguishing between a small nocturnal insectivore and Aggregating multiple answers was critical to producing a large ungulate. the high rates of accuracy on Snapshot Serengeti. Individ- As found in many other projects, citizen scientist accu- ual volunteers demonstrated similar levels of accuracy as racy differed by species (Dickinson et al. 2010). A subset volunteers for other projects (approximately 85% [e.g., of species was perfectly classified regardless of how many Dickinson et al. 2010]), and accuracy differed with expe- times they appeared in the data set. These species tended rience (Supporting Information). to be visually striking (e.g., giraffe, porcupine, male lion, Having multiple people classify an image was more and waterbuck) and thus clearly identifiable even to inex- reliable than a single person—even when that single per- perienced volunteers. In contrast, rare species had higher son was an expert. The aggregated volunteer answers rates of both false-positive and false-negative errors were even more accurate (97.9%) than those of individ- Conservation Biology Volume 30, No. 3, 2016 Uncertainty 528 Citizen Science Data Quality 1.0 (a) 0.8 Data 0.6 All Images Difficult Images Easy Images 0.4 010 20 30 Number Classifiers High Accuracy High False Negative High False Positive (b) (c) (d) 1.00 0.75 0.50 0.25 0.00 010 20 30010 20 30010 20 30 Number Classifiers Number Classifiers Number Classifiers giraffe zebra wildebeest aardwolf jackal topi rhinoceros kori bustard dik dik Figure 6. (a) Accuracy of bootstrapped plurality algorithm applied to volunteer classification (20 iterations) for each n volunteer species identifications. Proportion correct is the proportion of times the expert classifications agreed with the aggregated volunteer answer for resolvable images (i.e., images in which the expert could determine the species present) (solid black line, accuracy calculated for all images; dark gray dashed line, accuracy for images characterized as difficult [evenness >0.5], and light gray dashed line, accuracy for images characterized as easy [evenness 0.5], where evenness scores were calculated dynamically after every additional classification). Species characterized by (b) high accuracy, (c) high false-negative identifications, and (d) high false-positive identifications. (Fig. 3), mirroring results from other studies. Rare species camera-trap projects to arise exclusively from failure to present fewer opportunities for learning, and people are photograph animals (MacKenzie et al. 2002, 2006; Royle especially eager to report rare or unique species (Gal- & Dorazio 2008). However, we found that false negatives loway et al. 2006; Delaney et al. 2007). were also caused by a failure of classifiers to identify False positives and false negatives have different impli- an animal, which can be especially problematic when cations for conservation research. False positives are typ- studying rare and elusive species. ically calculated and corrected in citizen science projects The lack of false-negative detection is a critical limita- through expert review or detectors meant to catch un- tion of existing validation protocols in projects that flag likely classifications. False negatives (failure to identify implausible sightings for expert review but overlook plau- an animal that is present) are often implicitly assumed by sible errors, such as misidentifying a rare species as a Conservation Biology Volume 30, No. 3, 2016 Proportion Correct Proportion Correct Swansonetal. 529 common one (Bonter & Cooper 2012). The multiple- how best to produce accurate, reliable data (Tulloch volunteer approach specifically addresses this limitation et al. 2013; Wiggins et al. 2014). Based on lessons from by providing simple metrics that reflect the likelihood of Snapshot Serengeti, we recommend the following. First, an image being correctly classified (Fig. 5). Because errors engage multiple classifiers for each image and calculate tend to be clustered between species of similar morphol- an aggregated answer from multiple classifications. Send- ogy (Supporting Information), false negatives can be ad- ing an image to 10 volunteers, instead of 1, increased dressed by reviewing all images with low certainty scores accuracy from 85% to 95%. Second, do not allow answers, reported as an animal similar in appearance to the target such as I don’t know or impossible, that lack information species. about a correct classification. The variation in allowable Evenness and fraction support both provide simple, answers can indicate whether an image is likely to be dif- reliable metrics that reflect the likelihood of an image ficult or impossible. Third, produce expert-verified data being correctly identified. Researchers can set threshold to validate aggregated classifications and to measure cer- values below which images can be targeted for review tainty for every project. Accuracy should be assessed on or exclusion. The volume of images for expert review a per-project basis because it depends on the ecological depends on the certainty threshold required by the re- system, taxa, volunteer interface, and the volunteer base. search questions as well as the frequency of the target Fourth, balance effort between experts and volunteers species in the data set. according to project needs and capacity. Set baseline levels of accuracy, determine the necessary number of volunteer classifications, and estimate the effort that must Efficiency be devoted by experts to review difficult or ambiguous For projects that are limited by classifier effort, minimiz- images. ing per-image effort is critical to timely data processing. Engaging multiple citizen scientists in image classifi- Maximizing efficiency requires balancing volunteer ef- cation is not limited to remote camera surveys. Many fort, expert effort, and levels of acceptable error for a current projects ask people to essentially collect and an- particular analysis. alyze data on the spot (e.g., identifying animals or plants In Snapshot Serengeti, images achieved approximately seen and submitting a written record). Volunteers could 90% accuracy at 5 classifiers, 95% accuracy at 10 classi- instead collect data as photographs that can then be fiers, and asymptotically approached 98% accuracy after analyzed by multiple volunteers for validation, thereby increasing quality of the data. 20 classifiers (Fig. 6a). Efficiency differed according to A multiple-classifier approach to engaging citizen sci- species (Fig. 6b). Whereas common or easily recogniz- ence has dramatic implications for ecology and conser- able species were almost always correctly classified with vation biology—both for basic and applied research. For 2 or 3 volunteers, rare species sometimes needed 10 or more classifiers to achieve similar levels of accuracy. example, in conservation monitoring, single researchers For some rare species, accuracy rates improved only neg- or small teams often deploy camera traps to study specif- ligibly with additional classifiers. ically a single rare species (Karanth 1995; Dillon & Kelly For a study focusing on relatively rare species, even 2008; O’Connell et al. 2011). To keep pace with high- small rates of error could have a substantial effect on volume data production, conservationists often resort results. Thus, an expert should inspect all images with to classifying only those images containing their target evenness scores of >0.5 that have been identified as species—discarding enormous amounts of image data the target species (to eliminate false positives) and im- that could otherwise be used for multispecies monitor- ages with morphologically similar species (to eliminate ing. By engaging citizen scientists in the processing of false negatives). Adopting a less cautious threshold (say, every image and limiting expert review to only those evenness >0.75) would reduce the number of images to photos with low certainty scores, multispecies camera review. surveys could dramatically expand the scope of conser- Certainty metrics can be used dynamically to assess vation monitoring. whether an image needs additional volunteer classifica- tions. The evenness metric quickly becomes a reliable indicator of whether an image is likely to be correct or in- Acknowledgments correct: with 2 volunteers, 90% of images with evenness <0.5 are correct, and with 5 volunteers, 97% of images We thank members of the Serengeti Lion Project, partic- with evenness <0.5 are correct. The decision to send an ularly D. Rosengren, G. Gwaltu Lohay, and S. Mwampeta, image to more volunteers, flag it for expert review, or the Zooniverse staff, and the 28,040 volunteers exclude it from analysis ultimately depends on the needs who contributed to Snapshot Serengeti classifications and resources of a given project. (complete list at www.snapshotserengeti.org/#/authors). With the widespread engagement of the general public This work was supported by NSF grant DEB-1020479 in scientific research, projects must carefully consider to C.P. for maintenance of the long-term Lion Project, Conservation Biology Volume 30, No. 3, 2016 530 Citizen Science Data Quality by Google’s Global Impact Award to C.L., by private Gardiner MM, Allee LL, Brown PM, Losey JE, Roy HE, Smyth RR. 2012. Lessons from lady beetles: accuracy of monitoring data from US and donations raised during the Serengeti Live and Save UK citizen-science programs. Frontiers in Ecology and the Environ- Snapshot Serengeti crowdfunding campaigns, and by ment 10:471–476. grants from Explorer’s Club, UMN Thesis Research Hennon CC, et al. 2014. Cyclone center: Can citizen scientists improve Grants, UMN Office of International Programs, American tropical cyclone intensity records? Bulletin of the American Mete- Society of Mammalogists, Minnesota Zoo Ulysses S. Seal orological Society 96. DOI: http://dx.doi.org/10.1175/BAMS-D-13- 00152.1. Conservation fund, and National Geographic. Snapshot Johnson LC, et al. 2015. PHAT stellar cluster survey. II. Andromeda Serengeti website development was funded by awards Project cluster catalog. arXiv preprint arXiv:1501.04966. Avail- to Zooniverse from the Alfred P. Sloan Foundation. able from http://arxiv.org/abs/1501.04966 (accessed February 2015). Karanth KU. 1995. Estimating tiger Panthera tigris populations from Supporting Information camera-trap data using capture—recapture models. Biological Con- servation 71:333–338. Confusion matrix for species identifications (Appendix Karanth K, Nichols JD. 2002. Monitoring tigers and their prey: a man- S1), precision of volunteer-contributed animal counts ual for researchers, managers, and conservationists in tropical Asia. (Appendix S2), distribution of volunteer counts for im- Centre for Wildlife Studies, Bangalore. Kelly MJ, Noss AJ, Di Bitetti MS, Maffei L, Arispe RL, Paviolo A, De ages with large count ranges (Appendix S3), distribution Angelo CD, Di Blanco YE. 2008. Estimating puma densities from of certainty metrics (Appendix S4), volunteer accuracy camera trapping across three study sites: Bolivia, Argentina, and compared to number of contributions (Appendix S5), Belize. Journal of Mammalogy 89:408–418. species-specific sample size and error rates (Appendix Kinnaird MF, O’Brien TG. 2012. Effects of Private-land use, livestock S5), and ANOVA results for certainty measures as predic- management, and human tolerance on diversity, distribution, and tors of whether images were classified correctly, incor- abundance of large African mammals: livestock and large mammals in Kenya. Conservation Biology 26:1026–1039. rectly, or were impossible to classify (Appendix S6) are Lintott CJ, et al. 2008. Galaxy Zoo: morphologies derived from vi- available online. The authors are solely responsible for sual inspection of galaxies from the Sloan Digital Sky Survey. the content and functionality of these materials. Queries Monthly Notices of the Royal Astronomical Society 389:1179– (other than absence of the material) should be directed to the corresponding author. MacKenzie DI, Nichols JD, Lachman GB, Droege S, Andrew Royle J, Langtimm CA. 2002. Estimating site occupancy rates when detection probabilities are less than one. Ecology 83:2248–2255. MacKenzie DI, Nichols JD, Royle JA, Pollock KH, Bailey LL, Hines JE. Literature Cited 2006. Occupancy estimation and modeling: inferring patterns and Beaumont CN, Goodman AA, Kendrew S, Williams JP, Simpson R. 2014. dynamics of species. Elsevier, Amsterdam. The Milky Way Project: leveraging citizen science and machine Marshall PJ, et al. 2015. SPACE WARPS - I. Crowdsourcing the discovery learning to detect interstellar bubbles. The Astrophysical Journal of gravitational lenses. Monthly Notices of the Royal Astronomical Supplement Series 214:1. Society 455:1171–1190. Bischof R, Ali H, Kabir M, Hameed S, Nawaz MA. 2014. Being the Mugar G, Østerlund C, Hassman KD, Crowston K, Jackson CB. 2014. underdog: an elusive small carnivore uses space with prey and time Planet hunters and seafloor explorers: legitimate peripheral partic- without enemies. Journal of Zoology 293. DOI: 10.1111/jzo.12100. ipation through practice proxies in online citizen science. Pages Bonter DN, Cooper CB. 2012. Data validation in citizen science: a case 109–119 in Proceedings of the 17th ACB Conference on Computer study from Project FeederWatch. Frontiers in Ecology and the Envi- Proxies in Online Citizen Science. Association for Computing Ma- ronment 10:305–307. chinery. Delaney DG, Sperling CD, Adams CS, Leung B. 2007. Marine invasive O’Brien TG, Baillie JEM, Krueger L, Cuke M. 2010. The Wildlife Picture species: validation of citizen science and implications for national Index: monitoring top trophic levels. Animal Conservation 13:335– monitoring networks. Biological Invasions 10:117–128. Dickinson JL, Shirk J, Bonter D, Bonney R, Crain RL, Martin J, Phillips O’Connell AF, Nichols JD, Karanth KU. 2011. Camera traps in animal T, Purcell K. 2012. The current state of citizen science as a tool for ecology: methods and analyses. Springer, New York. ecological research and public engagement. Frontiers in Ecology Pielou EC. 1966. Species-diversity and pattern-diversity in the study and the Environment 10:291–297. of ecological succession. Journal of Theoretical Biology 10:370– Dickinson JL, Zuckerberg B, Bonter DN. 2010. Citizen science as an ecological research tool: challenges and benefits. Annual Review of Raddick MJ, Bracey G, Gay PL, Lintott CJ, Murray P, Schawinski K, Szalay Ecology, Evolution, and Systematics 41:149–172. AS, Vandenberg J. 2010. Galaxy Zoo: exploring the motivations of Dillon A, Kelly MJ. 2007. Ocelot Leopardus pardalis in Belize: the im- citizen science volunteers. Astronomy Education Review 9. DOI: pact of trap spacing and distance moved on density estimates. Oryx 10.3847/AER2009036. 41:469–477. Raddick MJ, Bracey G, Gay PL, Lintott CJ, Cardamone C, Mur- Dillon A, Kelly MJ. 2008. Ocelot home range, overlap and density: ray P, Schawinski K, Szalay AS, Vandenberg J. 2013. Galaxy comparing radio telemetry with camera trapping. Journal of Zoology Zoo: motivations of citizen scientists. arXiv:1303.6886 [astro-ph, 275:391–398. physics:physics]. Available from http://arxiv.org/ftp/arxiv/papers/ Foster-Smith J, Evans SM. 2003. The value of marine ecological data 1303/1303.6886.pdf (accessed October 2015). collected by volunteers. Biological Conservation 113:199–213. Royle JA, Dorazio RM. 2008. Hierarchical modeling and inference in Galloway AWE, Tudor MT, Haegan WMV. 2006. The reliability of citizen ecology: the analysis of data from populations, metapopulations science: a case study of Oregon white oak stand surveys. Wildlife and communities. Academic, Burlington, MA. Society Bulletin 34:1425–1429. Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 531 Silvertown J. 2009. A new dawn for citizen science. Trends in Ecology Tulloch AIT, Possingham HP, Joseph LN, Szabo J, Martin TG. 2013. & Evolution 24:467–471. Realising the full potential of citizen science monitoring programs. Silvertown J, Harvey M, Greenwood R, Dodd M, Rosewell J, Rebelo T, Biological Conservation 165:128–138. Ansine J, McConway K. 2015. Crowdsourcing the identification of Wiggins A, Lagoze C, Wong W-K, Kelling S. 2014. A sensor net- organisms: a case-study of iSpot. ZooKeys 480:125–146. work approach to managing data quality in citizen science. Sec- ond AAAI Conference on Human Computation and Crowdsourc- Simpson E, Roberts S, Psorakis I, Smith A. 2012. Dynamic Bayesian ing. Available from http://www.aaai.org/ocs/index.php/HCOMP/ combination of multiple imperfect classifiers. ArXiv e-prints HCOMP14/paper/view/9263 (accessed March 7, 2016). 1206:arXiv:1206.1831. Willett KW, et al. 2013. Galaxy Zoo 2: detailed morphological classi- Swanson A, Kosmala M, Lintott C, Simpson R, Smith A, Packer C. 2015. fications for 304,122 galaxies from the Sloan Digital Sky Survey. Snapshot Serengeti, high-frequency annotated camera trap images arXiv:1308.3496 [astro-ph]. Available from http://arxiv.org/pdf/ of 40 mammalian species in an African savanna. Scientific Data 1308.3496v2.pdf (accessed October 2015). 2:150026. Conservation Biology Volume 30, No. 3, 2016 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Conservation Biology Pubmed Central

A generalized approach for producing, quantifying, and validating citizen science data from wildlife images

Conservation Biology , Volume 30 (3) – Apr 25, 2016

Loading next page...
 
/lp/pubmed-central/a-generalized-approach-for-producing-quantifying-and-validating-02C0wJuqy2

References (41)

Publisher
Pubmed Central
Copyright
© 2016 The Authors. Conservation Biology published by Wiley Periodicals, Inc. on behalf of Society for Conservation Biology.
ISSN
0888-8892
eISSN
1523-1739
DOI
10.1111/cobi.12695
Publisher site
See Article on Publisher Site

Abstract

Special Section: Moving from Citizen to Civic Science to Address Wicked Conservation Problems A generalized approach for producing, quantifying, and validating citizen science data from wildlife images ∗ ∗ ∗ Alexandra Swanson, † ¶ Margaret Kosmala, ‡ Chris Lintott,† and Craig Packer Department of Ecology, Evolution and Behavior, University of Minnesota, Saint Paul, MN 55108, U.S.A. †Department of Physics, University of Oxford, Denys Wilkinson Building, Oxford, OX1 3RH, U.K. ‡Current address: Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA 02138, U.S.A. Abstract: Citizen science has the potential to expand the scope and scale of research in ecology and con- servation, but many professional researchers remain skeptical of data produced by nonexperts. We devised an approach for producing accurate, reliable data from untrained, nonexpert volunteers. On the citizen science website www.snapshotserengeti.org, more than 28,000 volunteers classified 1.51 million images taken in a large-scale camera-trap survey in Serengeti National Park, Tanzania. Each image was circulated to, on average, 27 volunteers, and their classifications were aggregated using a simple plurality algorithm. We validated the aggregated answers against a data set of 3829 images verified by experts and calculated 3 certainty metrics—level of agreement among classifications (evenness), fraction of classifications supporting the aggregated answer (fraction support), and fraction of classifiers who reported “nothing here” for an image that was ultimately classified as containing an animal (fraction blank)—to measure confidence that an aggregated answer was correct. Overall, aggregated volunteer answers agreed with the expert-verified data on 98% of images, but accuracy differed by species commonness such that rare species had higher rates of false positives and false negatives. Easily calculated analysis of variance and post-hoc Tukey tests indicated that the certainty metrics were significant indicators of whether each image was correctly classified or classifiable. Thus, the certainty metrics can be used to identify images for expert review. Bootstrapping analyses further indicated that 90% of images were correctly classified with just 5 volunteers per image. Species classifications based on the plurality vote of multiple citizen scientists can provide a reliable foundation for large-scale monitoring of African wildlife. Keywords: big data, camera traps, crowdsourcing, data aggregation, data validation, image processing, Snapshot Serengeti, Zooniverse Una Estrategia Generalizada para la Produccion, ´ Cuantificacion ´ y Validacion ´ de los Datos de Ciencia Ciudadana a partir de Im´agenes de Vida Silvestre Resumen: La ciencia ciudadana tiene el potencial de expandir el alcance y la escala de la investigacion ´ en la ecolog´ıa y la conservacion, ´ pero muchos investigadores profesionales permanecen esc´epticos sobre los datos producidos por quienes no son expertos. Disenamos ˜ una estrategia para generar datos precisos y fiables a partir de voluntarios no expertos y sin entrenamiento. En el sitio web de ciencia ciudadana www.snapshotserengeti.org mas ´ de 28, 000 voluntarios clasificaron 1.51 millon ´ de imagenes ´ que fueron tomadasenuncenso agranescaladecamaras ´ trampa en el Parque Nacional Serengueti, Tanzania. Cada imagen llego, ´ en promedio, hasta 27 voluntarios, cuyas clasificaciones se conjuntaron mediante el uso de ¶Address for correspondence: Department of Physics, University of Oxford, Denys Wilkinson Building, Oxford OX1 3RH, U.K. email ali@ zooniverse.org Paper submitted February 27, 2015; revised manuscript accepted August 19, 2015. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Conservation Biology, Volume 30, No. 3, 520–531 2016 The Authors. Conservation Biology published by Wiley Periodicals, Inc. on behalf of Society for Conservation Biology. DOI: 10.1111/cobi.12695 Swansonetal. 521 un algoritmo de pluralidad simple. Validamos el conjunto de respuestas frente a un juego de datos de 3, 829 imagenes ´ verificadas por expertos y calculamos tres medidas de certeza: nivel de concordancia entre las clasificaciones (uniformidad), fraccion ´ de clasificaciones que apoyan al conjunto de respuestas (fraccion ´ de apoyo) y fraccion ´ de clasificadores que reportaron “nada aqu´ı” en una imagen que al final se clasificoc ´ omo que s´ıten´ıa un animal (fraccion ´ en blanco). Estas medidas se usaron para estimar la confianza de que un conjunto de respuestas estuviera en lo correcto. En general, el conjunto de respuestas de los voluntarios estuvo de acuerdo con los datos verificados por los expertos en un 98 % de las imagenes, pero la certeza variosegun la ´ ´ ´ preponderancia de la especie, de tal forma que las especies raras tuvieron una tasa mas alta de falsos positivos y falsos negativos. El analisis de varianza calculado facilmente y las pruebas post-hoc de Tukey indicaron que ´ ´ las medidas de certeza fueron indicadores significativos de si cada imagen estuvo clasificada correctamente o si era clasificable. Por esto, las medidas de certeza pueden utilizarse para identificar imagenes ´ para una revision ´ de expertos. Los analisis ´ de bootstrapping indicaron mas ´ a fondo que el 90 % de las imagenes ´ estuvieron clasificadas correctamente con solo ´ cinco voluntarios por imagen. Las clasificaciones de especies basadas en el voto de pluralidad de multiples ´ cient´ıficos ciudadanos puede proporcionar un fundamento fiable para un monitoreo a gran escala de la vida silvestre africana. Palabras Clave: c´amaras trampa, conjunto de datos, crowdsourcing, datos grandes, procesamiento de im´agenes, Snapshot Serengeti, validacion ´ de datos, Zooniverse son et al. 2010). However, these procedures take time and Introduction may waste potentially valuable information and volunteer effort. Alternatively, eBird and FeederWatch ask volun- Modern citizen science, the engagement of the general teers to report bird sightings, flag implausible reports, public in the process of science, has enormous potential and engage experts in validation of flagged contributions to expand the scope and scale of research in ecology (Bonter & Cooper 2012). Although this approach can and conservation. These fields have long benefited from reduce false positives of unusual sightings, it leaves no volunteer contributions to, for example, the Audubon way to verify plausible but erroneous entries. Society’s Christmas Bird Count, which dates back more than 100 years (Silvertown 2009). In the last decade, tech- Successful citizen science projects in astrophysics, nological advances have rapidly accelerated the number such as Galaxy Zoo (Lintott et al. 2008; Willett et al. and diversity of projects that include public participation 2013), Space Warps (Marshall et al. 2016), Milky Way (Silvertown 2009; Dickinson et al. 2010, 2012; Tulloch Project (Simpson et al. 2012; Beaumont et al. 2014), and et al. 2013). Andromeda Project (Johnson et al. 2015) rely on the judg- Online projects engage people to contribute data on ments of multiple volunteers to classify satellite and tele- scope imagery. Cyclone Center, a meteorological project, an extraordinary array of taxa around the world (e.g., asks multiple users to identify features in infrared satellite Firefly Watch, HerpMapper, International Waterbird Cen- images of storms (Hennon et al. 2014). Each of these sus, and Road Watch) and on weather and climate (e.g., projects applies algorithms to aggregate the responses Citizen Weather Observer Program). Increased internet and produces expert-quality data sets. Similar approaches connectivity now allows volunteers to upload species can be applied to wildlife and conservation-based citizen sightings on websites such as iSpot.org and immediately science projects that ask volunteers to identify animals in interact with dozens of other naturalists (Silvertown et al. photographs taken with camera traps. 2015). Integrating volunteer effort and emerging tech- Digital image collection from camera-trap surveys is a nologies expands the range of possibility in both basic rapidly expanding method (O’Connell et al. 2011) that is and applied research.However, broad-scale implementa- used to study rare and elusive species worldwide (e.g., tion of citizen science for research is hindered by con- Karanth & Nichols 2002; Dillon & Kelly 2007; Kelly et al. cerns about data quality. Many professional researchers 2008) and to survey animals across large spatial extents are skeptical of data produced by nonexperts, which (e.g., O’Brien et al. 2010; Kinnaird & O’Brien 2012; lowers publication rates and grant funding of citizen Bischof et al. 2014). However, as inexpensive digital science projects (Foster-Smith & Evans 2003; Dickinson cameras have proliferated, such surveys are increasingly et al. 2010; Bonter & Cooper 2012). Although individ- limited by human processing capacity. Engaging citizen ual contributors can be measurably worse than trained professionals (Foster-Smith & Evans 2003; Galloway et al. scientists to classify images can dramatically increase the 2006; Delaney et al. 2007; Gardiner et al. 2012), solutions amount of information researchers can extract from large are available for assuring quality control of volunteer data. data sets. Some projects train volunteers or require volunteers to We devised an approach to produce accurate, re- pass a competency test, whereas others discard data liable data from multiple untrained, nonexpert volun- teers classifying images from the Snapshot Serengeti from inexperienced or unreliable contributors (Dickin- Conservation Biology Volume 30, No. 3, 2016 522 Citizen Science Data Quality (www.snapshotserengeti.org) camera-trapping study. overused and thus reduced the efficiency of volunteer We framed our analyses in terms of accuracy and effi- contributions. ciency to provide guidelines for optimizing the trade-off Each image was circulated to multiple users and retired between effort (of volunteers and experts) and accuracy. after meeting the following criteria: the first 5 classi- Because conservation studies often target rare species, fications were nothing here (hereafter blank); 10 non- we also evaluated how measures of accuracy and effi- consecutive nothing-here classifications (hereafter blank ciency differed across species of contrasting rarity. Our consensus); or 10 matching classifications of species or overarching goal was to provide straightforward tools and species combinations, not necessarily consecutive (here- quantifiable metrics to test and validate citizen science after consensus). If none of these criteria were met, the data, thus providing biologists with a generalizable ap- image was circulated until it accumulated 25 species clas- proach for engaging citizen scientists to produce reliable sifications (hereafter complete). These values were cho- data for large-scale conservation and wildlife research. sen based on volunteer performance on existing Zooni- verse projects. Volunteers classified Snapshot Serengeti data faster than images were produced, and images were Methods recirculated for use in classrooms. As a result, the number of classifications for images containing animals ranged The Snapshot Serengeti Interface from 11 to 57 (mean = 26, median = 27). Snapshot Serengeti is hosted by the Zooniverse citizen science platform (www.zooniverse.org), which engages 1.5 million volunteers worldwide to participate in a broad Data Aggregation and Validation array of projects. Zooniverse volunteers are motivated largely by a desire to contribute to science (Raddick et al. We implemented a simple plurality algorithm to trans- 2010, 2013) and engage with research teams and one form the volunteer classifications for each image into a another in high-level scientific discussions via the Zooni- single aggregated classification. As described in Swanson verse discussion forums (Mugar et al. 2014). et al. (2015), we first evaluated the median number, n, On www.snapshotserengeti.org, members of the gen- of different species reported by all classifiers for that im- eral public viewed and classified images from a large- age. For simplicity in interpreting volunteer accuracy, we scale camera survey in Serengeti National Park, Tanzania limited our analyses here to the 94% of collected images (see Swanson et al. 2015 for survey details). From June with n = 1. We then identified the species present as the 2010 to May 2013, the camera survey accumulated 99,241 n species with the most classifications. For example, if camera-trap days and produced 1.2 million image sets an image with n = 1 had 15 total classifications, with 7 (each image set contained 1–3 images taken in a single classifications of wildebeest, 5 classifications of buffalo, burst over approximately 1 s). Within 3 d of launching the and 3 classification of topi, the aggregated classification website, volunteers contributed 1 million species classi- would be wildebeest. If the same image had n = 2, the ag- fications and processed an 18-month backlog of images gregated classification would be wildebeest and buffalo. (Swanson et al. 2015). We calculated the number of individuals present for each Users were asked to identify species, count the number identified species by the median count (rounded up) of of animals(binnedas1,2,3,4,5,6,7,8,9,10, 11–50, all raw classifications for that image and the interquartile and 51+ individuals), and characterize behaviors in each range of counts reported by all classifiers for a given image set (Fig. 1). Volunteers followed a simple tutorial image. that explained the online interface, but they were not We calculated 3 measures of certainty or confidence formally trained or tested for accuracy before contribut- for each image: evenness, fraction blanks, and fraction ing. We designed the interface to help guide people with support. Evenness was calculated from all classifications no background knowledge through the process of animal that were not blank for each image with Pielou’s evenness identification from 48 possible species and taxonomic index (Pielou 1966): −( p ln p )/ ln S,where S is i i i = 1 groups while providing a rapid route to classification for the number of different species reported by all volunteers more knowledgeable participants. New users filtered po- and p is the proportion of classifications received by tential species matches by morphological characteristics species i. When all classifications were in agreement, we such as horn shape, body shape, color, pattern, tail shape, assigned a value of zero. The maximum value for this or a general gestalt (e.g., “looks like an antelope or deer”). index is 1.0, indicating high disagreement among clas- More experienced users could select the species directly sifications. Fraction blank was calculated as the fraction from a list. A “nothing here” button allowed users to of classifiers who reported nothing here for an image that classify images without any visible animals, but we did was ultimately classified as containing an animal. Fraction not provide an I-don’t-know option because previous test- support was calculated as the fraction of classifications ing with this option with undergraduate volunteers on a that supported the aggregated answer (i.e., fraction sup- small-scale prototype indicated that such answers were port of 1.0 indicated unanimous support). Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 523 Figure 1. The Snapshot Serengeti website interface used for classifying species, counts, and behaviors in images from the camera-trapping survey: (a) primary interface with all available species options and (b) filters that help narrow users’ choices when classifying species. We compared overall plurality algorithm perfor- culating the proportion of times the aggregated answer mance to a baseline expert-verified data set of 3829 for species present was confirmed by expert classifica- randomly sampled images. This data set (from Swanson tions (reported as proportion correct) in the baseline et al. 2015) was produced by asking a panel of experts expert-validated data set. We calculated accuracy both for to review these images. The experts were individuals the resolvable images and for all images. We calculated who had undergone extensive formal training, passed species-specific accuracy as the likelihood of the aggre- qualification exams, or had years of experience identify- gated answer being correct (i.e., the likelihood of the ing African wildlife. Of these images, 135 were indepen- expert classifications confirming the given aggregated dently classified by >1 expert. In cases where experts answer) based on images in the expanded expert-verified disagreed with the results of the plurality algorithm or data set.We further evaluated species-specific accuracy had marked an image set as particularly difficult or impos- with respect to 2 types of error: false positives (species sible, A.S. and C.P. made the final authoritative species reported when not present) and false negatives (species identification. For species-specific analyses, we used an not reported when actually present). Because the false- expanded data set of 5558 images that included extensive negative analysis required images to be randomly sam- random sampling of images identified as rare species by pled with respect to the true answer, this analysis was the plurality algorithm to ensure their adequate represen- limited to the baseline expert-verified data. The false- tation (species-specific sample sizes are in Supporting positive analysis was applied to the expanded data set Information). In 0.8% of images, the panel of experts because it only required that images be randomly sam- agreed that no authoritative species identification could pled with respect to the aggregated answer. For each be made. Because the citizen science interface does not species, we evaluated the proportion of photographs allow for an impossible classification, the aggregated an- containing each type of error. We associated a group swers were technically wrong for these images because of species that were easy to identify with zero error, no reliable volunteer answer exists. Additional details of and then compared nonzero error rates to species com- the field study, classification interface, aggregation, and monness (defined as the total number of times a species validation are available in Swanson et al. (2015). was photographed) with simple linear regression on log- transformed variables on the remaining species. To assess the accuracy of counts, we first evaluated Accuracy the agreement among expert counts for the 135 images We compared the results of the plurality algorithm with with multiple expert classifications. We also calculated expert answers for species identification and animal count ranges for each image from the raw classifications counts. We evaluated overall algorithm accuracy by cal- andcomparedexpertagreementwiththese ranges.We Conservation Biology Volume 30, No. 3, 2016 524 Citizen Science Data Quality calculated count range as the interquartile range (i.e., 25th percentile to 75th percentile values) of counts re- ported by all classifiers for a given image. We evaluated how often expert counts fell within this range for all resolvable images in the baseline expert-verified data set. To assess image difficulty for effectively targeting expert review, we evaluated the 3 certainty measures against ac- curacy, classifying images as correct (plurality answer agreed with expert-verified data set), incorrect (plurality answer disagreed with expert-verified data set), or impos- sible (experts could not identify the species). To test the predictive power of these metrics, we performed a one- way analysis of variance (ANOVA) followed by post-hoc Tukey test (package stats in Program R) for each metric to evaluate whether mean evenness, fraction support, or fraction blanks differed significantly across correct (n = 5281), incorrect (n = 225), and impossible (n = 52) images from the extended expert-verified data set. We further used a simple linear regression of log-transformed mean species-specific values of evenness, fraction sup- port, and fraction blanks on the logarithm of the total number of times a given species was photographed to assess the effects of species commonness on the 3 mea- sures of difficulty. Efficiency We determined how many volunteer classifiers were needed to produce reliable species identifications by bootstrapping the plurality answer from raw classifications for images with expert-verified answers. We first excluded the expert answers from the raw data set. Then, for every image, we randomly sampled (with replacement) n classifications from the raw data set (20 iterations each) and applied the plurality algorithm to produce an aggregated answer and calculate an evenness score. In case of ties, one answer was randomly selected from the 2. For each n volunteers, we evaluated accuracy as the Figure 2. Example images from the Serengeti camera proportion of times the plurality algorithm agreed with survey and presented on the Snapshot Serengeti the expert-verified answer. Overall accuracy was cal- website illustrating situations in which: (a) species culated using images from the expert-verified data set. identification is impossible, (b) a precise count of We further characterized species as belonging to 1 of animals is impossible, and (c) animals present in 3 groups: high overall accuracy (i.e., low false-positive foreground and background that leads to a wide and false-negative rates), high false positives, and high range of individual counts by volunteers. false negatives. We calculated species-specific accuracy for species in each of these groups. Species-specific ac- and >0.5 and calculated the proportion of images in each curacy was calculated for the expanded data set as the group (<0.5 and >0.5) that were correct. probability of the aggregated answer being correct. Because images vary in identification difficulty, we fur- ther evaluated the number of classifiers needed to pro- Results vide an evenness score that reliably indicated a correct answer. For every iteration of the bootstrapping analysis Accuracy for every n classifications, we calculated the evenness score for a given image. At every additional classification, Experts identified species in 3800 of 3829 randomly sam- we split images into those with evenness scores of <0.5 pled images (99.2%), labeling 29 images as impossible Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 525 False Negative False Positive rodents(+) aardwolf jackal topi bushbuck(+) G gazelle other bird(+) reedbuck eland impala buffalo hartebeest elephant(+) T gazelle wildebeest waterbuck(*) warthog(*) vervet monkey(*) serval(*) secretary bird(*) porcupine(*) mongoose(*) male lion(*) leopard(*) human(*) hippopotamus(*) giraffe(*) baboon(*) zebra(+) aardvark(+) female lion(+) guinea fowl(+) cheetah(+) ostrich(+) spotted hyena(+) dik dik(+) kori bustard(+) rhinoceros(+) reptiles(*) hare(*) caracal(*) wildcat civet honey badger bat eared fox genet zorilla striped hyena 0.25 0.50 0.75 1.00 0.25 0.50 0.75 1.00 Fraction of Error Figure 3. Species-specific rates of false-negative (calculated relative to the randomly sampled expert-verified data set of 3829 images) and false-positive (calculated relative to the extended expert-verified data set of 5558 images) error for species identifications produced by the plurality algorithm (error bars, standard error calculated for proportions; ∗, species with zero total error; +, species with zero false-positive error). Note that x-axis is plotted on a square-root scale. Sample sizes are in Supporting Information. or unresolvable (e.g., Fig. 2a). The plurality algorithm were perfectly classified regardless of how often they agreed with experts on 3750 images, yielding 97.9% over- appeared in the data set (Fig. 3). Both error types were all agreement and 98.6% agreement on resolvable images. significantly higher for rare species, although stronger for Accuracy differed dramatically by species, ranging from false negatives (p  0.0001, r = 0.71, df = 13) than for 100% accuracy for giraffe (n = 87) and hippopotamus false positives (0.0001, r = 0.45, df = 26). (n = 28) to 70% for jackals and 50% for aardwolves (see Precise counts were unresolvable in many images (e.g., full list in Supporting Information). Fig. 2b). Experts agreed on the number of individuals Species-specific accuracy also varied in the type of er- only 74% of the time (of 135 photos), and the average ror (Fig. 3): species with high false negatives tended to range of expert answers spanned 2.5 bins. The median have fewer false positives. A few species had high rates count reported by the plurality algorithm was just as likely of both false positives and negatives, and these tended to to agree with experts as experts were to agree among be confused with each other. Confusion was typically themselves (75.2% of 3800 images). clustered within groups of species with similar sizes, The interquartile range reported by the plurality al- shapes, and colors (Supporting Information). Algorithm gorithm was generally precise. Volunteers agreed on a accuracy was generally related to rarity. Species with single count in 50% of images, and 86% of images had fewer photographs had higher rates of false negatives ranges of 3 bins (e.g., 4–6 animals) (Supporting Infor- and false positives (Fig. 4), although a subset of species mation). In images with a wide range of counts, animals Conservation Biology Volume 30, No. 3, 2016 526 Citizen Science Data Quality was ultimately resolvable (p < 0.0001 for all pairwise 1.00 ErrorType comparisons). Images that evoked a large proportion of False Negative nothing-here responses sometimes contained a partial or False Positive blurred view of an animal, making it particularly difficult to identify. Thus, some classifiers apparently preferred to state there was "nothing here" rather than to guess the animal’s identity. The vast majority of images were classified as easy— 0.10 showing high levels of agreement on species classifica- tion (Supporting Information). For example, half of all images had >87% agreement on the final answer, and only 6% of images did not attain a majority (i.e., no species had >50% of the classifications). Excluding classifications of nothing here, 36% of images had unanimous agree- 0.01 ment on the species classification. As with accuracy, rare species were significantly more difficult to identify than 10 100 1,000 10,000 100,000 Total Captures common species: regression of certainty scores versus log of total pictures showed higher fraction support (p Figure 4. False-positive and false-negative (log 2 2 < 0.0001, r = 0.397), lower evenness (p < 0.001, r = fraction of error) identification of animals versus 0.214, df = 46), and lower fraction blanks (p = 0.0036, species commonness (frequency of species appearance r = 0.17, df = 46) with increasing commonness (Sup- in the overall data set, given as log[total number of porting Information). pictures]). Linear regression was performed on log-log transformed variables and restricted to species with nonzero error rates. Species with an asterisk in Fig. 3 Efficiency were excluded from false-positive and false-negative Accuracy increased asymptotically with the number of analyses, and species in Fig. 3 marked with a plus volunteers (Fig. 6a). Overall, we achieved 90% accuracy were excluded from the false-positive analysis. on species identifications with 5 volunteer classifications and 95% accuracy with 10 classifications. The number of volunteer classifications needed for ac- appeared in both the foreground and background (Fig. curate answers differed with species (Fig. 6b). Images 2c), and the distribution of counts tended to be bimodal with easy species (characterized by low rates of false (Supporting Information), presumably due to some users negatives and false positives) were nearly always correct only counting animals in the foreground and others with just 1–3 classifiers. Images with high rates of false counting everything in the image. negatives (e.g., jackals, aardwolves, and topi) needed Evenness, fraction support, and fraction blank were more classifiers to achieve high accuracy rates, but ad- all excellent predictors of whether aggregated answers ditional improvement declined after about 10 classifica- were likely to be correct (Fig. 5). The post-hoc Tukey test tions. In contrast, for species with high false-positive rates revealed that evenness scores were significantly lower (such as the extremely rare civets, genets, and striped (i.e., answers were all skewed toward a single species) for hyenas), accuracy rates remained low even after 30 clas- images that were classified correctly than for images that sifiers. were incorrectly classified (p < 0.0001) or impossible Evenness scores were excellent predictors of an ac- (p < 0.0001). Similarly, fraction support was higher for curate classification (Fig. 6a & Supporting Information). images that were correctly identified than for images that After 3 classifications, images with evenness 0.5 (e.g., were incorrectly classified (p < 0.0001) or impossible to at least 2 of the 3 users agreed on the species) were97% classify (p < 0.0001). However, evenness stood out as likely to be correct. After 5 classifications, images with an the single best predictor of a correct classification: 99.9% evenness score of 0.5 were 99% likely to be correct. of images with evenness < 0.25 were correct and 99.8% Thus, evenness can quickly be used to identify images of images with evenness < 0.5 were correct. In contrast, requiring additional classifications. only 85% of images with evenness > 0.5 and 71% of images with evenness > 0.75 were correct. Although evenness (p = 0.157) and fraction support Discussion (p = 0.394) for incorrect and impossible images did not differ significantly, the fraction of nothing-here classifica- Snapshot Serengeti provides a case study in engaging tions differed significantly across all 3 groups, and frac- citizen scientists in rapidly and accurately processing tion blank was the best predictor of whether an image large volumes of ecological imagery. Unlike many other Conservation Biology Volume 30, No. 3, 2016 Fraction of Error Swansonetal. 527 Evenness 1 - FractionSupport FractionBlanks abb abb ab c 1.00 0.75 0.50 0.25 0.00 correct incorrect impossible correct incorrect impossible correct incorrect impossible Comparison to Expert Classifications Figure 5. Evenness (support level of agreement among classifications), fraction blanks (fraction of classifiers who reported “nothing here”), and fraction support (fraction of classifications supporting the aggregated answer) for images that were verified by experts and deemed to be correct (aggregated volunteer answer agreed with expert answer), incorrect (aggregated volunteer answer did not agree with expert answer), or impossible (experts could not determine the species present). All metrics are bounded between 0 and 1, fraction support is plotted as “1-fraction support” so that for all 3 metrics, scores closer to 1 reflect greater uncertainty. Boxplots marked with (a) have significantly different means than those marked with (b). citizen science projects (Dickinson et al. 2010), Snapshot ual experts (96.6%) when compared with the consensus Serengeti volunteers were neither trained nor required to expert assessments. Experts can make mistakes: a single demonstrate species identification skills. Instead, we en- field researcher flipping through hundreds or thousands gaged multiple volunteers for every task and aggregated of photographs can become fatigued and miss species or their answers to produce highly accurate data. Snapshot click on the wrong classification. Making a precise count Serengeti data were 97.9% accurate overall, whereas 85– was impossible in many images (Fig. 2b). Calculating 95% accuracy is reported for projects engaging trained a count range from multiple volunteers provided more volunteers (Galloway et al. 2006; Delaney et al. 2007; reliable count data than a single number reported by a Gardiner et al. 2012). Engaging multiple volunteers for single expert (recall that experts disagreed on counts for every image did not necessarily mean many volunteers 26% of images). for every image. By evaluating measures of certainty in When creating the expert-verified data, experts agreed volunteer answers and evaluating the relative contribu- that a small number of photographs (0.8%) were impos- tion of additional volunteer classifications, we provide sible to classify (Fig. 2a). These impossible photographs guidelines for researchers to target volunteer effort and accounted for 36% of the overall error because the Snap- expert effort to balance their data needs and available shot Serengeti interface did not allow users to mark im- human resources for specific projects. ages as such. However, the likelihood of an image truly being impossible to identify can be determined by the fraction of blanks reported in volunteer classifications. Accuracy Furthermore, even guesses provide information, such as distinguishing between a small nocturnal insectivore and Aggregating multiple answers was critical to producing a large ungulate. the high rates of accuracy on Snapshot Serengeti. Individ- As found in many other projects, citizen scientist accu- ual volunteers demonstrated similar levels of accuracy as racy differed by species (Dickinson et al. 2010). A subset volunteers for other projects (approximately 85% [e.g., of species was perfectly classified regardless of how many Dickinson et al. 2010]), and accuracy differed with expe- times they appeared in the data set. These species tended rience (Supporting Information). to be visually striking (e.g., giraffe, porcupine, male lion, Having multiple people classify an image was more and waterbuck) and thus clearly identifiable even to inex- reliable than a single person—even when that single per- perienced volunteers. In contrast, rare species had higher son was an expert. The aggregated volunteer answers rates of both false-positive and false-negative errors were even more accurate (97.9%) than those of individ- Conservation Biology Volume 30, No. 3, 2016 Uncertainty 528 Citizen Science Data Quality 1.0 (a) 0.8 Data 0.6 All Images Difficult Images Easy Images 0.4 010 20 30 Number Classifiers High Accuracy High False Negative High False Positive (b) (c) (d) 1.00 0.75 0.50 0.25 0.00 010 20 30010 20 30010 20 30 Number Classifiers Number Classifiers Number Classifiers giraffe zebra wildebeest aardwolf jackal topi rhinoceros kori bustard dik dik Figure 6. (a) Accuracy of bootstrapped plurality algorithm applied to volunteer classification (20 iterations) for each n volunteer species identifications. Proportion correct is the proportion of times the expert classifications agreed with the aggregated volunteer answer for resolvable images (i.e., images in which the expert could determine the species present) (solid black line, accuracy calculated for all images; dark gray dashed line, accuracy for images characterized as difficult [evenness >0.5], and light gray dashed line, accuracy for images characterized as easy [evenness 0.5], where evenness scores were calculated dynamically after every additional classification). Species characterized by (b) high accuracy, (c) high false-negative identifications, and (d) high false-positive identifications. (Fig. 3), mirroring results from other studies. Rare species camera-trap projects to arise exclusively from failure to present fewer opportunities for learning, and people are photograph animals (MacKenzie et al. 2002, 2006; Royle especially eager to report rare or unique species (Gal- & Dorazio 2008). However, we found that false negatives loway et al. 2006; Delaney et al. 2007). were also caused by a failure of classifiers to identify False positives and false negatives have different impli- an animal, which can be especially problematic when cations for conservation research. False positives are typ- studying rare and elusive species. ically calculated and corrected in citizen science projects The lack of false-negative detection is a critical limita- through expert review or detectors meant to catch un- tion of existing validation protocols in projects that flag likely classifications. False negatives (failure to identify implausible sightings for expert review but overlook plau- an animal that is present) are often implicitly assumed by sible errors, such as misidentifying a rare species as a Conservation Biology Volume 30, No. 3, 2016 Proportion Correct Proportion Correct Swansonetal. 529 common one (Bonter & Cooper 2012). The multiple- how best to produce accurate, reliable data (Tulloch volunteer approach specifically addresses this limitation et al. 2013; Wiggins et al. 2014). Based on lessons from by providing simple metrics that reflect the likelihood of Snapshot Serengeti, we recommend the following. First, an image being correctly classified (Fig. 5). Because errors engage multiple classifiers for each image and calculate tend to be clustered between species of similar morphol- an aggregated answer from multiple classifications. Send- ogy (Supporting Information), false negatives can be ad- ing an image to 10 volunteers, instead of 1, increased dressed by reviewing all images with low certainty scores accuracy from 85% to 95%. Second, do not allow answers, reported as an animal similar in appearance to the target such as I don’t know or impossible, that lack information species. about a correct classification. The variation in allowable Evenness and fraction support both provide simple, answers can indicate whether an image is likely to be dif- reliable metrics that reflect the likelihood of an image ficult or impossible. Third, produce expert-verified data being correctly identified. Researchers can set threshold to validate aggregated classifications and to measure cer- values below which images can be targeted for review tainty for every project. Accuracy should be assessed on or exclusion. The volume of images for expert review a per-project basis because it depends on the ecological depends on the certainty threshold required by the re- system, taxa, volunteer interface, and the volunteer base. search questions as well as the frequency of the target Fourth, balance effort between experts and volunteers species in the data set. according to project needs and capacity. Set baseline levels of accuracy, determine the necessary number of volunteer classifications, and estimate the effort that must Efficiency be devoted by experts to review difficult or ambiguous For projects that are limited by classifier effort, minimiz- images. ing per-image effort is critical to timely data processing. Engaging multiple citizen scientists in image classifi- Maximizing efficiency requires balancing volunteer ef- cation is not limited to remote camera surveys. Many fort, expert effort, and levels of acceptable error for a current projects ask people to essentially collect and an- particular analysis. alyze data on the spot (e.g., identifying animals or plants In Snapshot Serengeti, images achieved approximately seen and submitting a written record). Volunteers could 90% accuracy at 5 classifiers, 95% accuracy at 10 classi- instead collect data as photographs that can then be fiers, and asymptotically approached 98% accuracy after analyzed by multiple volunteers for validation, thereby increasing quality of the data. 20 classifiers (Fig. 6a). Efficiency differed according to A multiple-classifier approach to engaging citizen sci- species (Fig. 6b). Whereas common or easily recogniz- ence has dramatic implications for ecology and conser- able species were almost always correctly classified with vation biology—both for basic and applied research. For 2 or 3 volunteers, rare species sometimes needed 10 or more classifiers to achieve similar levels of accuracy. example, in conservation monitoring, single researchers For some rare species, accuracy rates improved only neg- or small teams often deploy camera traps to study specif- ligibly with additional classifiers. ically a single rare species (Karanth 1995; Dillon & Kelly For a study focusing on relatively rare species, even 2008; O’Connell et al. 2011). To keep pace with high- small rates of error could have a substantial effect on volume data production, conservationists often resort results. Thus, an expert should inspect all images with to classifying only those images containing their target evenness scores of >0.5 that have been identified as species—discarding enormous amounts of image data the target species (to eliminate false positives) and im- that could otherwise be used for multispecies monitor- ages with morphologically similar species (to eliminate ing. By engaging citizen scientists in the processing of false negatives). Adopting a less cautious threshold (say, every image and limiting expert review to only those evenness >0.75) would reduce the number of images to photos with low certainty scores, multispecies camera review. surveys could dramatically expand the scope of conser- Certainty metrics can be used dynamically to assess vation monitoring. whether an image needs additional volunteer classifica- tions. The evenness metric quickly becomes a reliable indicator of whether an image is likely to be correct or in- Acknowledgments correct: with 2 volunteers, 90% of images with evenness <0.5 are correct, and with 5 volunteers, 97% of images We thank members of the Serengeti Lion Project, partic- with evenness <0.5 are correct. The decision to send an ularly D. Rosengren, G. Gwaltu Lohay, and S. Mwampeta, image to more volunteers, flag it for expert review, or the Zooniverse staff, and the 28,040 volunteers exclude it from analysis ultimately depends on the needs who contributed to Snapshot Serengeti classifications and resources of a given project. (complete list at www.snapshotserengeti.org/#/authors). With the widespread engagement of the general public This work was supported by NSF grant DEB-1020479 in scientific research, projects must carefully consider to C.P. for maintenance of the long-term Lion Project, Conservation Biology Volume 30, No. 3, 2016 530 Citizen Science Data Quality by Google’s Global Impact Award to C.L., by private Gardiner MM, Allee LL, Brown PM, Losey JE, Roy HE, Smyth RR. 2012. Lessons from lady beetles: accuracy of monitoring data from US and donations raised during the Serengeti Live and Save UK citizen-science programs. Frontiers in Ecology and the Environ- Snapshot Serengeti crowdfunding campaigns, and by ment 10:471–476. grants from Explorer’s Club, UMN Thesis Research Hennon CC, et al. 2014. Cyclone center: Can citizen scientists improve Grants, UMN Office of International Programs, American tropical cyclone intensity records? Bulletin of the American Mete- Society of Mammalogists, Minnesota Zoo Ulysses S. Seal orological Society 96. DOI: http://dx.doi.org/10.1175/BAMS-D-13- 00152.1. Conservation fund, and National Geographic. Snapshot Johnson LC, et al. 2015. PHAT stellar cluster survey. II. Andromeda Serengeti website development was funded by awards Project cluster catalog. arXiv preprint arXiv:1501.04966. Avail- to Zooniverse from the Alfred P. Sloan Foundation. able from http://arxiv.org/abs/1501.04966 (accessed February 2015). Karanth KU. 1995. Estimating tiger Panthera tigris populations from Supporting Information camera-trap data using capture—recapture models. Biological Con- servation 71:333–338. Confusion matrix for species identifications (Appendix Karanth K, Nichols JD. 2002. Monitoring tigers and their prey: a man- S1), precision of volunteer-contributed animal counts ual for researchers, managers, and conservationists in tropical Asia. (Appendix S2), distribution of volunteer counts for im- Centre for Wildlife Studies, Bangalore. Kelly MJ, Noss AJ, Di Bitetti MS, Maffei L, Arispe RL, Paviolo A, De ages with large count ranges (Appendix S3), distribution Angelo CD, Di Blanco YE. 2008. Estimating puma densities from of certainty metrics (Appendix S4), volunteer accuracy camera trapping across three study sites: Bolivia, Argentina, and compared to number of contributions (Appendix S5), Belize. Journal of Mammalogy 89:408–418. species-specific sample size and error rates (Appendix Kinnaird MF, O’Brien TG. 2012. Effects of Private-land use, livestock S5), and ANOVA results for certainty measures as predic- management, and human tolerance on diversity, distribution, and tors of whether images were classified correctly, incor- abundance of large African mammals: livestock and large mammals in Kenya. Conservation Biology 26:1026–1039. rectly, or were impossible to classify (Appendix S6) are Lintott CJ, et al. 2008. Galaxy Zoo: morphologies derived from vi- available online. The authors are solely responsible for sual inspection of galaxies from the Sloan Digital Sky Survey. the content and functionality of these materials. Queries Monthly Notices of the Royal Astronomical Society 389:1179– (other than absence of the material) should be directed to the corresponding author. MacKenzie DI, Nichols JD, Lachman GB, Droege S, Andrew Royle J, Langtimm CA. 2002. Estimating site occupancy rates when detection probabilities are less than one. Ecology 83:2248–2255. MacKenzie DI, Nichols JD, Royle JA, Pollock KH, Bailey LL, Hines JE. Literature Cited 2006. Occupancy estimation and modeling: inferring patterns and Beaumont CN, Goodman AA, Kendrew S, Williams JP, Simpson R. 2014. dynamics of species. Elsevier, Amsterdam. The Milky Way Project: leveraging citizen science and machine Marshall PJ, et al. 2015. SPACE WARPS - I. Crowdsourcing the discovery learning to detect interstellar bubbles. The Astrophysical Journal of gravitational lenses. Monthly Notices of the Royal Astronomical Supplement Series 214:1. Society 455:1171–1190. Bischof R, Ali H, Kabir M, Hameed S, Nawaz MA. 2014. Being the Mugar G, Østerlund C, Hassman KD, Crowston K, Jackson CB. 2014. underdog: an elusive small carnivore uses space with prey and time Planet hunters and seafloor explorers: legitimate peripheral partic- without enemies. Journal of Zoology 293. DOI: 10.1111/jzo.12100. ipation through practice proxies in online citizen science. Pages Bonter DN, Cooper CB. 2012. Data validation in citizen science: a case 109–119 in Proceedings of the 17th ACB Conference on Computer study from Project FeederWatch. Frontiers in Ecology and the Envi- Proxies in Online Citizen Science. Association for Computing Ma- ronment 10:305–307. chinery. Delaney DG, Sperling CD, Adams CS, Leung B. 2007. Marine invasive O’Brien TG, Baillie JEM, Krueger L, Cuke M. 2010. The Wildlife Picture species: validation of citizen science and implications for national Index: monitoring top trophic levels. Animal Conservation 13:335– monitoring networks. Biological Invasions 10:117–128. Dickinson JL, Shirk J, Bonter D, Bonney R, Crain RL, Martin J, Phillips O’Connell AF, Nichols JD, Karanth KU. 2011. Camera traps in animal T, Purcell K. 2012. The current state of citizen science as a tool for ecology: methods and analyses. Springer, New York. ecological research and public engagement. Frontiers in Ecology Pielou EC. 1966. Species-diversity and pattern-diversity in the study and the Environment 10:291–297. of ecological succession. Journal of Theoretical Biology 10:370– Dickinson JL, Zuckerberg B, Bonter DN. 2010. Citizen science as an ecological research tool: challenges and benefits. Annual Review of Raddick MJ, Bracey G, Gay PL, Lintott CJ, Murray P, Schawinski K, Szalay Ecology, Evolution, and Systematics 41:149–172. AS, Vandenberg J. 2010. Galaxy Zoo: exploring the motivations of Dillon A, Kelly MJ. 2007. Ocelot Leopardus pardalis in Belize: the im- citizen science volunteers. Astronomy Education Review 9. DOI: pact of trap spacing and distance moved on density estimates. Oryx 10.3847/AER2009036. 41:469–477. Raddick MJ, Bracey G, Gay PL, Lintott CJ, Cardamone C, Mur- Dillon A, Kelly MJ. 2008. Ocelot home range, overlap and density: ray P, Schawinski K, Szalay AS, Vandenberg J. 2013. Galaxy comparing radio telemetry with camera trapping. Journal of Zoology Zoo: motivations of citizen scientists. arXiv:1303.6886 [astro-ph, 275:391–398. physics:physics]. Available from http://arxiv.org/ftp/arxiv/papers/ Foster-Smith J, Evans SM. 2003. The value of marine ecological data 1303/1303.6886.pdf (accessed October 2015). collected by volunteers. Biological Conservation 113:199–213. Royle JA, Dorazio RM. 2008. Hierarchical modeling and inference in Galloway AWE, Tudor MT, Haegan WMV. 2006. The reliability of citizen ecology: the analysis of data from populations, metapopulations science: a case study of Oregon white oak stand surveys. Wildlife and communities. Academic, Burlington, MA. Society Bulletin 34:1425–1429. Conservation Biology Volume 30, No. 3, 2016 Swansonetal. 531 Silvertown J. 2009. A new dawn for citizen science. Trends in Ecology Tulloch AIT, Possingham HP, Joseph LN, Szabo J, Martin TG. 2013. & Evolution 24:467–471. Realising the full potential of citizen science monitoring programs. Silvertown J, Harvey M, Greenwood R, Dodd M, Rosewell J, Rebelo T, Biological Conservation 165:128–138. Ansine J, McConway K. 2015. Crowdsourcing the identification of Wiggins A, Lagoze C, Wong W-K, Kelling S. 2014. A sensor net- organisms: a case-study of iSpot. ZooKeys 480:125–146. work approach to managing data quality in citizen science. Sec- ond AAAI Conference on Human Computation and Crowdsourc- Simpson E, Roberts S, Psorakis I, Smith A. 2012. Dynamic Bayesian ing. Available from http://www.aaai.org/ocs/index.php/HCOMP/ combination of multiple imperfect classifiers. ArXiv e-prints HCOMP14/paper/view/9263 (accessed March 7, 2016). 1206:arXiv:1206.1831. Willett KW, et al. 2013. Galaxy Zoo 2: detailed morphological classi- Swanson A, Kosmala M, Lintott C, Simpson R, Smith A, Packer C. 2015. fications for 304,122 galaxies from the Sloan Digital Sky Survey. Snapshot Serengeti, high-frequency annotated camera trap images arXiv:1308.3496 [astro-ph]. Available from http://arxiv.org/pdf/ of 40 mammalian species in an African savanna. Scientific Data 1308.3496v2.pdf (accessed October 2015). 2:150026. Conservation Biology Volume 30, No. 3, 2016

Journal

Conservation BiologyPubmed Central

Published: Apr 25, 2016

There are no references for this article.