Microsoft Word - Gaglio_et_al_Manuscript_Revised_FINALE_DOI.docx Dietary studies in birds: testing a non-invasive 1 method using digital photography in seabirds 2 3 Davide Gaglio1*, Timothée R. Cook1,2, Maëlle Connan3, Peter G. Ryan1 4 and Richard B. Sherley4,5,* 5 1Percy FitzPatrick Institute, DST-NRF Centre of Excellence, University of Cape Town, 6 Rondebosch 7701, South Africa; 2Institute of Ecology and Environmental Sciences, 7 Evolutionary Eco-physiology Team, University Pierre et Marie Curie, Bâtiment A–7ème étage, 7 8 quai, St Bernard, 75005 Paris, France; 3DST-NRF Centre of Excellence at the Percy FitzPatrick 9 Institute, Department of Zoology, Nelson Mandela Metropolitan University, PO Box 77000, Port 10 Elizabeth 6031, South Africa; 4Environment and Sustainability Institute, University of Exeter, 11 Penryn Campus, Penryn, Cornwall, TR10 9FE, United Kingdom; 5Animal Demography Unit, 12 Department of Biological Sciences, University of Cape Town, Private Bag X3, Rondebosch 13 7701, South Africa. 14 15 *Corresponding authors, email: (DG) swift.terns@gmail.com; (RBS) r.sherley@exeter.ac.uk 16 Summary 17 1. Dietary studies give vital insights into foraging behaviour, with implications for understanding 18 changing environmental conditions and the anthropogenic impacts on natural resources. Traditional diet 19 sampling methods may be invasive or subject to biases, so developing non-invasive and unbiased 20 methods applicable to a diversity of species is essential. 21 2. We used digital photography to investigate the diet fed to chicks of a prey-carrying seabird, and 22 compared our approach (photo-sampling) to a traditional method (regurgitations) for the greater crested 23 tern Thalasseus bergii. 24 3. Over three breeding seasons, we identified >24,000 prey items of at least 47 different species, more 25 than doubling the known diversity of prey taken by this population of terns. We present a method to 26 estimate the length of the main prey species (anchovy Engraulis encrasicolus) from photographs, with an 27 accuracy < 1 mm and precision ~0.5 mm. Compared to regurgitations at two colonies, photo-sampling 28 produced similar estimates of prey composition and size, at a faster species accumulation rate. The prey 29 compositions collected by two researchers photo-sampling concurrently were also similar. 30 4. Photo-sampling offers a non-invasive tool to accurately and efficiently investigate the diet 31 composition and prey size of prey-carrying birds. It reduces biases associated with observer-based 32 studies and is simple to use. This methodology provides a novel tool to aid conservation and 33 management decision-making in light of the growing need to assess environmental and anthropogenic 34 change in natural ecosystems. 35 36 Key-words: diet, digital photography, non-invasive monitoring, prey-carrying birds, rarefaction curves, 37 Thalasseus bergii, regurgitation 38 Introduction 39 Dietary studies are essential to understand animal ecology, temporal changes in the environment, and to 40 establish sustainable management strategies for natural resources (Jordan 2005). In complex natural 41 systems, top-predators can act as indicators of environmental conditions, and their diet, in particular, can 42 provide important information on prey species abundance, occurrence and size, which may reflect 43 processes over short time-frames (e.g. Suryan et al. 2002; Parsons et al. 2008). As such, outcomes from 44 diet studies are important tools for monitoring changes in demographic parameters or behaviour, 45 themselves a product of changing diet (Sherley et al. 2013). Moreover, dietary studies can provide 46 powerful indicators of anthropogenic impacts and environmental change on food-webs (e.g. Piatt et al. 47 2007; Green et al. 2015), facilitating conservation biology and ecosystem-based management (Grémillet 48 et al. 2008; Sherley et al. 2013). The importance of monitoring diet thus demands the development of 49 simple, efficient, non-invasive methods applicable to a diversity of species. 50 Numerous techniques exist to investigate bird diets (Jordan 2005; Inger & Bearhop 2008; Karnovsky, 51 Hobson & Iverson 2012). Invasive techniques include induced regurgitations (Diamond 1984), stomach 52 flushing of live birds (Wilson 1984), application of neck-collars on chicks (Moreby & Stoate 2000) and 53 the dissection of birds collected specifically for this purpose (Doucette, Wissel & Somers 2011). These 54 methods describe short-term diet composition accurately (González-Solís et al. 1997), despite some 55 errors introduced by differential prey regurgitation or digestion (e.g. Jackson & Ryan 1986). More recent 56 biochemical methods involving isotopic, lipid and DNA analyses provide complementary approaches, 57 but generally cannot be used alone due to their coarse taxonomic resolution (Karnovsky, Hobson & 58 Iverson 2012). Moreover, these approaches typically require disturbance or capture of birds, which can 59 impact their physiology and behaviour (e.g. Ellenberg et al. 2006; Carey 2009). 60 Accurate, non-invasive diet sampling is therefore required to give fine-scale indicators of prey 61 availability or prey selection. One of the least non-invasive methods is to observe birds carrying visible 62 prey with binoculars or video recording systems, from a safe distance. This typically involves birds 63 feeding offspring or incubating partners (e.g. Safina et al. 1990; Redpath et al. 2001; Tornberg & Reif 64 2007). Such studies are generally limited to assessing chick diet, but have the potential to reveal changes 65 in prey communities (Anderson et al. 2014). However, observer-based diet studies are subject to several 66 methodological limitations (Cezilly & Wallace 1988; González-Solís et al. 1997; Lee & Hockey 2001) 67 calling for further development of this approach. 68 Digital photography represents an excellent alternative tool to study the diet fed to chicks of prey-69 carrying birds, because 1) there is virtually no limit to the number of pictures that can be taken, 2) 70 species identification is possible in most cases, 3) prey can potentially be measured accurately and 71 precisely, 4) images can be re-analysed without loss of data quality, i.e. samples do not deteriorate over 72 time and 5) storage is simple. Over the last decade, the use of digital photography for dietary studies has 73 included the use of camera-traps to investigate the diet of nesting raptors (García-Salgado et al. 2015; 74 Robinson et al. 2015), and the combined use of digital compact cameras with spotting scopes 75 (digiscoping) to assist prey identification (made primarily by observations) for Caspian terns 76 (Hydroprogne caspia) and common murres (Uria aalge) (Larson & Craig 2006, Gladics et al. 2015). 77 However, both techniques have limitations including poor image quality and difficultly in capturing 78 images of birds carrying prey in flight or during fast delivery to chicks (see Larson & Craig 2006, 79 García-Salgado et al. 2015). 80 Recent advances in performance and price reductions of digital single lens reflex (DSLR) cameras 81 combined with autofocus telephoto lenses makes digital photography an affordable option for prey 82 identification, even for birds in flight. In the last few years, DSLRs have been used opportunistically to 83 identify items carried by a variety of birds (e.g. Woehler et al. 2013; Gaglio, Sherley & Cook 2015, Tella 84 et al. 2015) but a systematic approach and an accurate method to estimate prey dimensions are lacking. 85 We developed a standardised application of digital photography using DSLR cameras and telephoto 86 lenses to investigate chick diet composition and prey size in prey-carrying birds. We tested the method 87 on the colonial breeding greater crested tern Thalasseus bergii in South Africa. We compared the 88 efficacy of photo-sampling to the more traditional used regurgitations (Walter et al. 1987) using prey 89 identified to species level collected from chicks, and assessed the accuracy and precision of length 90 measurements of the main prey made from photographs. We also evaluated the potential for observer 91 bias in this system. Finally, we discuss the validity of applying our non-invasive approach to any prey-92 carrying bird and the potential to develop a simple and effective tool-box to accurately identify and 93 estimate the size of any carried item. 94 95 Methods 96 STUDY SPECIES AND SITES 97 The greater crested tern (hereafter ‘tern’) is distributed from the Namibian coast eastwards to the central 98 Pacific. It feeds mostly at sea by dipping onto the surface or plunge diving up to ca 1 m (Crawford, 99 Hockey & Tree 2005). During breeding, adults usually return from foraging with a single prey item, 100 which is either offered to the partner during courtship or delivered to the offspring (Crawford, Hockey & 101 Tree 2005). In South Africa, the sub-species Thalasseus bergii bergii breeds mostly on islands in the 102 Western Cape (Crawford 2003). Since 2008, Robben Island (33°48’S, 18°22’E), Table Bay, has hosted 103 the largest southern African colony, reaching ~13,000 breeding pairs in 2010 (Makhado et al. 2013). A 104 few hundred pairs breed in the Eastern Cape, mostly on Seal Island (33°50’S, 26°17’E), Algoa Bay 105 (Makhado et al. 2013). We studied their diet at both Robben and Seal Islands. 106 107 PHOTO-SAMPLING 108 We investigated the diet of breeding terns at Robben Island during 2013 (February–June), 2014 109 (January–June) and 2015 (February–June) and at Seal Island during June 2015. Adult terns returning 110 with prey were photographed from a vantage point of 50–80 m from the edge of their colony (Fig. 1a). 111 At Seal Island (~300 pairs) we were able to photograph all adults returning to the colony during our 112 photo-sampling sessions. At Robben Island, colonies were much larger (> 6,000 pairs) so we could not 113 photograph all individuals. However, every attempt was made to not bias selection to individuals 114 carrying particularly conspicuous prey items. The distance to the flying birds ranged between 6.5 and 25 115 m. Total sampling effort represented ~ 50 h of photography per year. For each individual, we typically 116 took a sequence of 3 photos (a “photo set”) for identification and prey measurements (Fig. 1b). We found 117 by trial and error that 3 images provided the best trade-off to balance processing time with obtaining at 118 least one sharp image. To avoid biasing the results and maintain independence among photo sets, ad-hoc 119 image analysis was performed for each sampling session to discard repeated photo sets of the same 120 adults carrying the same prey item. Recurrent birds were identified using distinguishable feather patterns, 121 presence of colour or metal rings, type and position of prey in the bill while flying, and distinctive 122 markings on the prey. 123 Photos were taken using Canon 7D and 7D Mark II cameras, fitted with Canon EF 100–400 mm 124 f/4.5-5.6L IS USM zoom lenses. We set the cameras to (i) shutter speed priority (1/2500 s); (ii) 125 automatic ISO (or aperture priority mode that provided shutter speeds of at least 1/2500 s); (iii) high-126 Speed Continuous Shooting; (iv) Autofocus on AI Servo (for moving subjects) using the AF point 127 expansion; and (v) large Jpeg file format for high-speed recording. We set the telephoto lens to 128 autofocus, the image stabilizer to on and the closest focal point to 6.5 m to increase autofocus speed. 129 130 IDENTIFICATION OF PREY SPECIES 131 All blurred or otherwise non-identifiable images (due to e.g. distance, an unfavourable position of prey in 132 the bill or lighting) were discarded. From the remaining photographs (e.g. Fig. 1), we determined the 133 numerical abundance (Duffy & Jackson 1986) of prey (usually at species level) using fish guides (Smith 134 & Heemstra 2003; Branch et al. 2010) and assistance from experienced observers (see 135 Acknowledgements). In some instances, good quality photographs contained prey that could not be 136 identified (< 0.01% of total prey items). For example, some adults returned with pieces of fish flesh, 137 possibly originating from kleptoparasitism disputes or scavenging. These images were excluded from our 138 analyses. Approximately 45% of photo sets were suitable for prey identification; there was no evidence 139 of bias towards particular prey types among discarded images. 140 141 ESTIMATION OF PREY STANDARD LENGTH 142 Dietary studies of piscivorous birds commonly measure the standard length (SL) of the fish (length from 143 the tip of the snout to the posterior edge of the hypural plate) to compare prey size (Barret 2002, Smith & 144 Heemstra 2003). We estimated SL from photographs for anchovy Engraulis encrasicolus, the most 145 common species in the tern’s diet. As prey tended to flex to differing degrees in the adults’ bills, direct 146 SL measurement from the image underestimates fish length. Thus, we estimated SL from measurements 147 of individual body parts (eye diameter, operculum width and head width, all measured dorsoventrally), 148 which were less distorted in the image and generally in a plane parallel to the bird’s bill and the camera 149 (Figs 1b and 2). 150 To do this we first assessed the accuracy of predicted SLs based on these morphological 151 measurements using cross-validation by fitting log-linear allometric regressions to a training dataset (n = 152 50) and comparing model predictions to a test dataset (n = 20) of anchovies measured by hand (see 153 Appendix S1). Next, we measured 37 additional anchovies with Vernier callipers (to the nearest 0.1 mm) 154 and then photographed them held in the bill of a dead tern, for which the culmen length was known (Fig. 155 2 in Appendix S1). For each image, we used the ‘line selection tool’ in ImageJ (Schneider et al. 2012) to 156 estimate eye diameter (𝐸), operculum width (𝑂) and head width (𝐻) for each fish by scaling the pixel 157 length in the image to (1) the length of the dead tern’s culmen (62.1 mm; measured with Vernier 158 callipers), (2) the mean culmen length for this species (61.2 mm, n = 128; Crawford, Hockey & Tree 159 2005) and (3) the minimum and maximum recorded culmen lengths (range: 54.5–67.6 mm, Crawford, 160 Hockey & Tree 2005). We used the estimates of 𝐸, 𝑂 and 𝐻 to obtain three estimates of SL (𝑆𝐿) using 161 the log-linear allometric regressions (see also Appendix S1), and calculated their arithmetic mean 162 (combined 𝑆𝐿) and used this value in further analyses (since it was generally most accurate; Appendix 163 S1). 164 To determine the accuracy (γ) of the combined 𝑆𝐿 estimates from the images, we compared them to 165 the known SL of each fish. We defined the mean percentage accuracy (𝛾) of the combined 𝑆𝐿 estimates 166 as: 167 𝛾 = 100 𝑛 1 − 𝑆𝐿! − 𝑐𝑜𝑚𝑏𝑖𝑛𝑒𝑑 𝑆𝐿! 𝑆𝐿! ! !!! (eqn 1) 168 where i indexes each of the n = 37 fish. As the absolute difference was computed, both overestimates and 169 underestimates of e.g. 2% would yield γ = 98%. In addition, we assessed the mean difference between 170 the known SLs and the combined 𝑆𝐿 estimates using permutation tests with 10,000 Monte Carlo 171 iterations (perm library v. 1.0-0.0 for R). 172 To determine the precision (or repeatability) of the method, we repeated the measurement process in 173 ImageJ to obtain six 𝐸, 𝑂 and 𝐻 values and the corresponding combined 𝑆𝐿 values for 17 of the 37 fish 174 (using a known length on the ruler in each photograph). We calculated the combined 𝑆𝐿 as above and 175 used this to assess precision. Precision (τ) was defined as: 176 𝜏!,! = 1 𝑛 𝑐𝑜𝑚𝑏𝑖𝑛𝑒𝑑 𝑆𝐿!,! ! !!! − 𝑐𝑜𝑚𝑏𝑖𝑛𝑒𝑑 𝑆𝐿!,! (eqn 2) 177 where j indexes each of the n = 6 combined 𝑆𝐿 values for the f = 17 fish. We report mean precision (in 178 mm) of all (6 x 17 = 102) values of τf,j. 179 In addition, we examined whether either precision or accuracy were influenced by the SL of a fish. 180 For accuracy, we used a linear model of the form: 181 logit(γi) = α + β × SLi + εi 182 (eqn 3) 183 where α and β are estimated from the data, γi are the accuracy estimates (as proportions), SLi the known 184 standard length for fish i and εi ~ N(0,σ) the residual error, with σ estimated from the data. For precision 185 we used a linear-mixed model (LMM: lme4 library for R) of the form: 186 τf,j = β × SLf,j + δf,j × ηj + εf,j 187 (eqn 4) 188 where β is the fixed effect parameter, ηj ~ N(0, 𝜍) the random effect parameter, εf,j ~ N(0, σ) the residual 189 error, δf,j the vector of fish IDs, τf,j the vector of precision values and SLf, the vector of known standard 190 lengths for each measurement j of fish f, with β, σ and 𝜍 estimated from the data. 191 Finally, we used the above approach to estimate SL of the prey in a subset of the digital images 192 collected in the field where the bird’s bill and the head of the prey were clearly visible and approximately 193 parallel to the camera (Fig. 1b). For each image, we used combined 𝑆𝐿 and assumed the length of the 194 bird’s culmen to be 61.2 mm (see above). 195 196 COMPARISON BETWEEN PHOTO-SAMPLING AND REGURGITATION-SAMPLING 197 To compare photo-sampling and regurgitation-sampling, we collected images of adults carrying prey and 198 regurgitations from chicks concurrently on 18 and 19 April 2015 at Robben Island (photo-sampling 199 effort: 600 min) and on 9 June 2015 at Seal Island (photo-sampling effort: 132 min). Regurgitates were 200 collected from the ground, while chicks were inside a pen during ringing operations (chicks often 201 regurgitate when disturbed). Prey were later identified from whole-prey or diagnostic prey remains 202 resistant to digestion such as otoliths and squid beaks using Clarke (1986), Smith & Heemstra (2003), 203 Smale, Watson & Hecht (1995), Branch et al. (2010) and the Port Elizabeth Museum’s reference 204 collection. Prey items that were not identified mainly consisted of fish flesh and were excluded from our 205 analysis. The SL of whole anchovies collected from regurgitations was measured using a ruler. 206 We compared the number of prey items from different taxa between methods using χ2 tests and 207 assessed differences in the estimated anchovy SLs using permutation tests (10,000 iterations) for each 208 island separately as the SL variance between islands was heterogeneous (Levene’s test: W(1,164) = 5.8, p = 209 0.017). 210 We examined prey diversity using sample-based rarefaction curves as these allow for standardized 211 comparison across collections that differ in sample size (Gotelli & Colwell 2001). Using 1,000 random 212 permutations of both the photo-samples and regurgitations from 18 and 19 April 2015, we produced 213 curves of the mean (± asymptotic 95% confidence intervals, CI) species accumulation rate (species 214 identified per sample made). We then compared this rate at samples sizes of n = 190. In addition, by 215 fitting a Generalised Additive Model (GAM) to the photo-sample means and by assuming equal 216 accumulation rates for extrapolation, we also compared the predicted species accumulation rate for 217 regurgitations to the mean rate for photo-sampling at n = 1500. The chosen sample sizes approximate 218 those obtained in the field. 219 Finally, to evaluate any possible observer effect on photo-sampling, two different researchers 220 (observer-A and observer-B) simultaneously collected photographs at Robben Island on 18 and 19 April 221 2015. The two observers used the same equipment (Canon 7D Mark II camera, Canon 100–400 mm lens) 222 and had similar experience in wildlife photography. All other procedures were the same as described 223 above. We compared the samples from the two observers using χ2 tests. Unless otherwise stated, all 224 means are presented ± 1 SD and all statistics were performed using R v.3.2.1. 225 226 Results 227 PHOTO-SAMPLING VS. REGURGITATION-SAMPLING 228 In total ~160,000 photos were taken during the three breeding seasons on Robben Island, yielding images 229 of 24,211 prey items identifiable to species (96%, 48 species) or family (98%, 49 families) level (total of 230 51 prey taxa; Table 1). During the regurgitation comparison trial at Robben Island, we identified 27 231 species from 1,510 photo-samples compared to 11 species from 198 regurgitated prey items. At Seal 232 Island, we identified 11 species from 157 photo-samples and 6 species from 103 regurgitated prey items 233 (Appendix S2). The mean species accumulation rate at 190 samples was 0.075 (95% CI: 0.058–0.089) 234 for photo-sampling and 0.057 (95% CI: 0.053–0.058) for regurgitations; however, at this sample size, the 235 95% CIs overlapped (Fig. 3). The number of species predicted from 1,500 regurgitations was 23.4 (based 236 on the GAM extrapolation) versus 27.0 for photo-sampling (Fig. 3). The diet composition of main prey 237 did not differ significantly between the two methods for Robben Island (χ2 = 47, d.f. = 42, p = 0.26) or 238 Seal Island (χ2 = 18, d.f. = 15, p = 0.26; Table S3 in Appendix S2). 239 240 ACCURACY AND PRECISION IN ESTIMATING ANCHOVY STANDARD LENGTH 241 Mean SL of the 50 anchovy used to calculate the allometric regressions between the morphometric 242 measurements (training set) was 109.6 ± 13.5 mm (range = 83.3–130.5 mm), similar to the 20 anchovy in 243 the test set (SL 112.8 ± 3.0 mm; range = 107.6–116.8 mm). The predicted 𝑆𝐿s of the test set 244 predominantly fell within the 95% prediction intervals for all three specific body part models (Fig. S1, 245 Appendix S1). The mean accuracy (𝛾) for the combined 𝑆𝐿 was 97.9 ± 1.7% (range 93.0–99.9%) for the 246 training set and 97.3 ± 1.8% (range 92.5–100%) for the test set. Accuracy was not affected by SL in 247 either case (linear models: p > 0.05, see Appendix S1). 248 The mean SL of the 37 photographed anchovy was 113.4 ± 6.7 mm. With the culmen length of the 249 dead tern (62.1 mm) as the reference, mean accuracy (𝛾) for the combined 𝑆𝐿 was 98.3 ± 1.5% (range 250 93.8–100%), yielding a mean combined 𝑆𝐿 of 114.0 ± 7.1 mm (Table S2 in Appendix S1). With the 251 species’ mean culmen length (61.2 mm) as the reference, the mean combined 𝑆𝐿 = 112.7 ± 7.0 mm (𝛾 = 252 98.1 ± 1.5%, range 92.2–99.9%; Fig. 4, Table S2). The length of a fish (actual SL) did not influence the 253 accuracy in either case (linear models: p > 0.05, Fig. 4) and neither of the combined 𝑆𝐿s differed 254 significantly from the actual SL (permutations tests: p > 0.05). The mean accuracy (𝛾) reduced to 88.9 (± 255 3.3)% and 91.3 (± 3.2)% for the minimum (54.5 mm) and maximum (67.6 mm) recorded culmen lengths 256 respectively (Table S2) and these combined 𝑆𝐿 series did differ significantly from the actual SLs 257 (permutations tests: p < 0.001; see Appendix S1). 258 The mean precision of the combined 𝑆𝐿 estimates was 0.52 (± 0.38) mm or 99.6 (± 0.3)%, with an 259 absolute range of 0.02–1.58 mm or 98.6–99.99%. Precision was not related to the actual SL of the fish 260 being measured (LMM: χ2 = 0.02, p = 0.89). 261 262 COMPARISONS OF PREY SIZE BETWEEN PHOTO-SAMPLING AND REGURGITATIONS 263 At Robben Island, 116 anchovy from photo-samples (10% of anchovy photographed) and 20 from 264 regurgitates (12%) could be measured, while at Seal Island, the corresponding values were 21 (18%) and 265 nine (9%) respectively. Overall, the anchovy were longer at Seal Island (mean = 120.3 ± 8.2 mm, n = 30) 266 than at Robben Island (91.2 ± 13.2 mm, n = 136; p < 0.001; Fig. 5). For Robben Island, the mean 267 combined 𝑆𝐿 of anchovy in the photo-samples was 91.3 ± 13.6 mm compared to 90.8 ± 11.1 mm for 268 regurgitates (Fig. 5). At Seal Island, they were 121.6 ± 9.3 mm and 117.4 ± 3.6 respectively. The SL 269 estimates from the two methods did not differ statistically for either Robben Island (p = 0.85) or Seal 270 Island (p = 0.21). 271 272 COMPARISON BETWEEN OBSERVERS 273 We identified 1,510 prey items of 22 species from the photographs taken by observer-A and 1,625 of 21 274 species from observer-B. Prey composition did not differ significantly between the two (χ2 = 72, d.f. = 275 64, p = 0.23). However, three species were not recorded in common; observer-A photographed one 276 horsefish Congiopodidae sp. and one eel Ophichthidae sp., while observer-B recorded three individuals 277 of Cape hake Merluccius capensis. 278 279 Discussion 280 Photo-sampling offers an effective, low-impact alternative to traditional diet studies for birds that carry 281 prey items in their bill, with accurate prey identification and size estimates possible. Samples can be 282 acquired quickly and equivalent diet compositions obtained with relatively low effort (Fig. 3). In three 283 breeding seasons, we sampled 24,211 prey items and identified 51 prey taxa (Table 1) with this 284 approach; the most comprehensive diet analysis for terns in southern Africa prior to our study identified 285 25 species from 1,311 regurgitated prey items in 10 breeding seasons (1977–1986; Walter et al. 1987). 286 Despite ~55% of photos being discarded, our approach yielded an order of magnitude more samples and 287 identified twice as many species, with minimal disturbance to breeding birds. 288 The photo-sampling approach has several other advantages over traditional diet sampling. First, terns 289 often regurgitate only the posterior body and caudal fin of a fish, making identification of similar species 290 difficult (McLeay et al. 2009). Photo-sampling records the entire prey, and if there is doubt as to the 291 identification, images can be shared easily with global experts or on specialized websites (e.g. I-spot). 292 Second, photo-sampling can be used in a range of situations (e.g. on land or from a boat), by one 293 individual (collection of regurgitations often involves many people), with minimal training in 294 photography (cameras can be pre-set). Third, the photographic equipment is relatively affordable and 295 once purchased can be used for several years, at multiple colonies and for several species. Also, although 296 processing the photographs can be time-consuming, taking about 30 min for an average of 100 prey 297 identified, the images can be stored and analysed multiple times if needed, without the loss of data 298 quality or metadata (e.g. date and location). 299 Possible drawbacks associated with photo-sampling include the repeated photography of prey items, 300 especially those with long handling times, leading the frequency of these items being over-estimated. 301 This is predominately a problem in larger colonies, where it is difficult to follow the fate of individual 302 prey items, and one that could be countered using delays (e.g. 5 mins) between photosets. When only a 303 subset of prey is sampled, large or conspicuous prey items may induce an observer bias if they are easier 304 to photograph, more readily identified to species level or more interesting to the photographer. Training 305 photographers to randomise the photo-sampling as much as possible should help reduce this potential 306 bias. Differences in photographic experience between different observers could create a potential bias 307 and should be examined in future studies. Photo-sampling is difficult in bad weather (strong wind, rain or 308 mist) and this may also introduce bias in some situations. Finally, one constraint of our study is that 309 photo-sampling was applied to study chick diet. Although this can provide important insights into 310 changes in prey communities (Anderson et al. 2014), it may not always represent adult diet, or diet 311 outside the breeding season (McLeay et al. 2009). We thus suggest implementing indirect methods such 312 as measuring stable isotope ratios in e.g. blood and feathers of adults (Inger & Bearhop 2008) 313 concurrently with photo-sampling. Moreover, applying both methods concurrently on marked individuals 314 would allow the development of trophic discrimination factors in wild animals (Newsome et al. 2010). 315 More broadly, ecologists now use digital photography to study animals across a wide range of taxa 316 (e.g. Morrison et al. 2011; Marshall & Pierce 2012; Gregory et al. 2014). Opportunistic observations 317 have documented novel behaviours and trophic interactions (e.g. Gaglio, Sherley & Cook 2015; Tella et 318 al. 2015), suggesting that standardised approaches to study species bringing items to a known location 319 have great potential for ecological monitoring. This approach could also be applied to a diversity of taxa 320 in addition to birds that carry prey (e.g. carnivores bringing prey to their offspring, or ants and termites 321 carrying items to their nests). In any of these applications photo-sampling could provide high quality 322 photographic data to complement the now extensive use of camera-traps. 323 The ecological information provided by prey size is almost as important as prey species, giving 324 information on the targeted prey cohort and the predator’s energetics. We demonstrated that prey size 325 (anchovy SL) can be estimated accurately (~98%) and precisely (~99%) from images. The approach 326 could be used with a wide variety of predators and prey species to eliminate biases associated with in situ 327 visual observation (Lee & Hockey 2001). Even if photo-sampling is unlikely to obtain measurements as 328 accurately or precisely as regurgitated/dropped prey, the sample size from photo-sampling is always 329 likely to be greater than the number of prey found undigested. A crucial step to estimate absolute prey 330 size is identifying a reference object (e.g. culmen, eye diameter) of known size, to provide a scale for 331 prey measurements. These reference objects should be chosen carefully and the degree to which the 332 selected trait varies within the population assessed to constrain and minimise errors where possible (see 333 Results). Additional studies could photograph birds of known bill length, age and sex (e.g. colour banded 334 individuals) with prey held with different angles to the body and compare larger numbers of observers 335 photo-sampling concurrently to further quantify the errors associated with prey measurements. For prey 336 species that are not distorted in images (e.g. some insects do not bend over a bird’s bill), size can be 337 estimated directly and even when absolute estimates are not possible, the method still can be used to 338 assess changes in relative prey size, allowing for spatial and temporal comparisons. 339 Crucially, the photo-sampling method caused little if any disturbance to the nesting birds. Distances 340 from animals can be selected to balance each species’ sensitivity against image quality. The opportunity 341 to record the number and size of prey brought to offspring remotely and in real time without influencing 342 behaviour, allows for accurate monitoring of temporal variability. For threatened or declining species 343 (e.g. many seabirds; Croxall et al. 2012), such non-invasive methods can help elucidate functional links 344 between population dynamics, environmental variability and anthropogenic pressures (Saraux et al. 345 2011). Incorporating these observations into detailed information on species composition and energy 346 content for energetic models offers great potential for indicators of long-term and large-scale ecosystem 347 change (Furness & Cooper 1982). Furthermore, with standardized protocols, digital images can be shared 348 easily using digital platforms (e.g. I-spot, Google Images) to facilitate global collaborations (e.g. 349 González-Solís et al. 2011; Lynch et al. 2015), encourage community involvement in citizen science 350 projects (e.g. Newman et al. 2012), and develop data archives to answer as yet unforeseen questions. 351 Given the growing need to assess environmental changes and human impacts on natural ecosystems 352 (Hobday et al. 2015), our methodology offers a novel tool for collaborative efforts in conservation. 353 354 Acknowledgements 355 Our research was supported by a Department of Science and Technology-Centre of Excellence grant to 356 the Percy FitzPatrick Institute of African Ornithology. SAN Parks and Robben Island Museum provided 357 logistical support and access to the tern colonies. We thank: Pierre Pistorius (Nelson Mandela 358 Metropolitan University), for supporting the research on Seal Island; Barrie Rose, Bruce Dyer, Carl van 359 der Lingen, Rob Leslie, Bryan Maritz, Charles Griffiths, Mike Picker, Jean-Paul Roux and Malcolm 360 Smale for assistance with prey identification; Malcolm Smale for access to the otolith and squid beak 361 reference collections at the Port Elizabeth Museum; Carl van der Lingen and Cecile Reed for samples of 362 anchovy; numerous volunteers who collected regurgitations; and Alistair McInnes (Observer A) and 363 Dominic Rollinson (Observer B). Thanks to Tom Flower and Stephen Votier for constructive comments 364 on an earlier draft. This research was approved by SANParks (CONM1182), the Department of 365 Environmental Affairs (RES2013/24, RES2014/83, RES2015/65) and the animal ethics committee of the 366 University of Cape Town (2013/V3/TC) and NMMU (A14-SCI-ZOO-003). 367 368 Data accessibility 369 All data used in this article are available in the Supplementary materials or Dryad Digital Repository 370 (Gaglio et al. 2016) http://dx.doi.org/10.5061/dryad.j647p. 371 372 References 373 Anderson, H.B., Evans, P.G.H., Potts, J.M., Harris, M.P. & Wanless S. (2014) The diet of Common 374 Guillemot Uria aalge chicks provide evidence of changing prey communities in the North Sea. Ibis, 375 156, 23–34. 376 Barrett, R. T. 2002. Atlantic puffin Fratercula arctica and common guillemot Uria aalge chick diet and 377 growth as indicators of fish stocks in the Barents Sea. Marine Ecology Progress Series, 230, 275–287. 378 Branch, G.M., Griffiths, C.L., Branch, M.L. & Beckley, L.E. (2010) Two Oceans: A Guide to the Marine 379 Life of Southern Africa. Randomhouse/Struik, Cape Town. 380 Carey, M.J. (2009) The effects of investigator disturbance on procellariiform seabirds: a review. New 381 Zealand Journal of Zoology, 36, 367–377. 382 Cezilly, F. & Wallace, J. (1988) The determination of prey captured by birds through direct field 383 observations: a test of the method. Colonial Waterbirds, 11, 110–112. 384 Clarke, M.R. (1986) A Handbook for the Identification of Cephalopod Beaks. Clarendon Press, Oxford. 385 Crawford, R.J.M. (2003) Influence of food on numbers breeding, colony size and fidelity to localities of 386 Swift Terns in South Africa’s Western Cape, 1987-2000. Waterbirds, 26, 44–53. 387 Crawford, R.J.M., Hockey, P.A.R. & Tree, A.J. (2005) Swift Tern Sterna bergii. Roberts Birds of 388 Southern Africa (7th edn) (eds P.A.R. Hockey, W.R.J. Dean& P.G. Ryan), pp 453–455. Trustees of 389 the John Voelcker Bird Book Fund, Cape Town. 390 Croxall, J.P., Butchart, S.H., Lascelles, B., Stattersfield, A.J., Sullivan, B., Symes, A. & Taylor, P.H.I.L. 391 (2012) Seabird conservation status, threats and priority actions: a global assessment. Bird 392 Conservation International, 22, 1–34. 393 Diamond, A.W. (1984) Feeding overlap in some tropical and temperate seabird communities. Studies in 394 Avian Biology, 8, 24–46. 395 Doucette, J.L., Wissel, B. & Somers, C.M. (2011) Cormorant-fisheries conflicts: stable isotopes reveal a 396 consistent niche for avian piscivores in diverse food webs. Ecological Applications, 21, 2987–3001. 397 Duffy, D.C. & Jackson, S. (1986) Diet studies of seabirds: a review of methods. Colonial Waterbirds, 9, 398 1–17. 399 Ellenberg, U., Mattern, T., Seddon, P.J. & Jorquera, G.L. (2006) Physiological and reproductive 400 consequences of human disturbance in Humboldt Penguins: the need for species-specific visitor 401 management. Biological Conservation, 133, 95–106. 402 Furness, R.W. & Cooper, J. (1982) Interactions between breeding seabird and pelagic fish populations in 403 the Southern Benguela region. Marine Ecology Progress Series, 8, 243–250. 404 Gaglio, D., Sherley, R.B. & Cook, T.R. (2015) Insects in the diet of the Greater Crested Tern Thalasseus 405 bergii bergii in southern Africa. Marine Ornithology, 43, 131–132. 406 Gaglio, D., Cook T. R., Connan M., Ryan P. G., Sherley R. B. (2016) Data from: Dietary studies in 407 birds: testing a non-invasive method using digital photography in seabirds. Dryad Digital Repository, 408 doi: 10.5061/dryad.j647p 409 García-Salgado, G., Rebollo, S., Pérez-Camacho, L., Martínez-Hesterkamp, S., Navarro, A. & 410 Fernández-Pereira, J.-M. (2015) Evaluation of trail-cameras for analyzing the diet of nesting raptors 411 using the Northern Goshawk as a model. PLoS One, 10, e0127585. 412 Gladics, A.J., Suryan, R.M., Parrish, J.K., Horton, C.A., Daly, E.A. & Peterson, W.T. (2015) 413 Environmental drivers and reproductive consequences of variation in the diet of a marine predator. 414 Journal of Marine Systems, 146, 72–81. 415 González-Solís, J., Oro, D., Pedrocchi, V., Jover, L. & Ruiz, X. (1997) Bias associated with diet samples 416 in Audouin’s Gulls. Condor, 99, 773–779. 417 González-Solís, J., Smyrli, M., Militão, T., Gremillet, D., Tveraa, T., Phillips, R.A. & Boulinier, T. 418 (2011) Combining stable isotope analyses and geolocation to reveal kittiwake migration. Marine 419 Ecology Progress Series, 435, 251–261. 420 Gotelli, N.J. & Colwell, R.K. (2001) Quantifying biodiversity: procedures and pitfalls in the 421 measurement and comparison of species richness. Ecology Letters, 4, 379–391. 422 Green, D.B., Klages, N.T.W., Crawford, R.J.M., Coetzee, J.C., Dyer, B.M., Rishworth, G.M. & 423 Pistorius, P.A. (2015) Dietary change in Cape Gannets reflects distributional and demographic shifts 424 in two South African commercial fish stock. ICES Journal of Marine Science, 72, 771–781. 425 Gregory, T., Carrasco Rueda, F., Deichmann, J., Kolowski, J., & Alonso, A. (2014) Arboreal camera 426 trapping: taking a proven method to new heights. Methods in Ecology and Evolution, 5, 443–451. 427 Grémillet, D., Pichegru, L., Kuntz, G., Woakes, A.G., Wilkinson, S., Crawford, R.J.M. & Ryan, P.G. 428 (2008) A junk-food hypothesis for gannets feeding on fishery waste. Proceedings of the Royal Society 429 of London, Series B, Biological Sciences, 275, 1149–1156. 430 Hobday, A.J., Bell, J.D., Cook, T.R., Gasalla, M.A. & Weng, K.C. (2015) Reconciling conflicts in 431 pelagic fisheries under climate change. Deep Sea Research Part II, 113, 291–300. 432 Inger, R. & Bearhop, S. (2008) Applications of stable isotope analyses to avian ecology. Ibis, 150, 447–433 461. 434 Jackson, S. & Ryan, P.G. (1986) Differential digestion rates of prey by White-chinned Petrels 435 (Procellaria aequinoctialis). Auk, 103, 617–619. 436 Jordan, M.J.R. (2005) Dietary analysis for mammals and birds: a review of field techniques and animal-437 management applications. International Zoo Yearbook, 39, 108–116. 438 Karnovsky, N.J., Hobson, K.A. & Iverson, S.J. (2012) From lavage to lipids: estimating diets of seabirds. 439 Marine Ecology Progress Series, 451, 263–284. 440 Larson, K. & Craig, D. (2006) Digiscoping vouchers for diet studies in bill-load holding birds. 441 Waterbirds, 29, 198–202. 442 Lee, N.M. & Hockey, P.A.R. (2001) Biases in the field estimation of shorebird prey sizes. Journal of 443 Field Ornithology, 72, 49–61. 444 Lynch, T. P., Alderman, R., & Hobday, A. J. (2015) A high-resolution panorama camera system for 445 monitoring colony-wide seabird nesting behaviour. Methods in Ecology and Evolution, 6, 491–499. 446 Makhado, A.B., Dyer, B.M., Fox, R., Geldenhuys, D., Pichegru, L., Randall, R.M., Sherley, R.B., 447 Upfold, L., Visagie, J., Waller, L.J., Whittington, P.A. & Crawford, R.J.M. (2013) Estimates of 448 numbers of twelve seabird species breeding in South Africa, updated to include 2012. Department of 449 Environmental Affairs, Internal Report. Pp. 1–16. 450 Marshall, A.D. & Pierce, S.J. (2012) The use and abuse of photographic identification in sharks and rays. 451 Journal of Fish Biology, 80, 1361–1379. 452 McLeay, L.J., Page, B., Goldsworthy, S.D., Ward, T.M. & Paton, D.C. (2009) Size matters: variation in 453 the diet of chick and adult crested terns. Marine Biology, 156, 1765–1780. 454 Moreby, S.J. & Stoate, C. (2000) A quantitative comparison of neck-collar and faecal analysis to 455 determine passerine nestling diet. Bird Study, 47, 320–331. 456 Morrison, T.A., Yoshizaki, J., Nichols, J.D. & Bolger, D.T. (2011) Estimating survival in photographic 457 capture–recapture studies: overcoming misidentification error. Methods in Ecology and Evolution, 2, 458 454–463. 459 Newman, G., Wiggins, A., Crall, A., Graham, E., Newman, S. & Crowston, K. (2012) The future of 460 citizen science: emerging technologies and shifting paradigms. Frontiers in Ecology and the 461 Environment, 10, 298–304. 462 Newsome, S.D., Bentall, G.B., Tinker, M.T., Oftedal, O.T., Ralis, K., Estes, J.A. & Fogel, M.L. (2010) 463 Variation in δ13C and δ15N diet - vibrissae trophic discrimination factors in a wild population of 464 California sea otters. Ecological Applications, 20, 1744-1752. 465 Parsons, M., Mitchell, I., Butler, A., Ratcliffe, N., Frederiksen, M., Foster, S. & Reid, J.B. (2008) 466 Seabirds as indicators of the marine environment. ICES Journal of Marine Science, 65, 1520–1526. 467 Piatt, J.F., Harding, A.M.A., Shultz, M., Speckman, S.G., van Pelt, T.I., Drew, G.S. & Kettle, A.B. 468 (2007) Seabirds as indicators of marine food supplies: Cairns revisited. Marine Ecology Progress 469 Series, 352, 221–234. 470 Redpath, S.M., Clarke, R., Madders, M. & Thirgood, S.J. (2001). Assessing raptor diet: comparing 471 pellets, prey remains, and observational data at Hen Harrier nests. Condor, 103, 184–188. 472 Robinson, B.G., Franke, A. & Derocher, A.E. (2015) Estimating nestling diet with cameras: quantifying 473 uncertainty from unidentified food items. Wildlife Biology, 21, 277–282. 474 Safina, C., Wagner, R.H., Witting, D.A. & Smith, K.J. (1990) Prey delivered to Roseate and Common 475 Tern chicks; composition and temporal variability. Journal of Field Ornithology, 61, 331–338. 476 Saraux, C., Le Bohec, C., Durant, J.M., Viblanc, V.A., Gauthier-Clerc, M., Beaune, D., Park, Y.-H., 477 Yoccoz, N.G., Stenseth, N.C. & Le Maho, Y. (2011) Reliability of flipper-banded penguins as 478 indicators of climate change. Nature, 469, 203–206. 479 Schneider, C.A., Rasband, W.S. & Eliceiri, K.W. (2012) NIH Image to ImageJ: 25 years of image 480 analysis. Nature Methods, 9, 671–675. 481 Sherley, R.B., Underhill, L.G., Barham, B.J., Barham, P.J., Coetzee, J.C., Crawford, R.J.M., Dyer, B.M., 482 Leshoro, T.M. & Upfold, L. (2013) Influence of local and regional prey availability on breeding 483 performance of African Penguins Spheniscus demersus. Marine Ecology Progress Series, 473, 291–484 301. 485 Smale, M.J., Watson, G. & Hecht, T. (1995) Otolith atlas of southern African marine fishes. 486 Ichthyological Monographs of the JLB Smith Institute of Ichthyology 1. 487 Smith, M.M. & Heemstra, P.C. (2003) Smiths' Sea Fishes. Struik Plublishers, Cape Town. 488 Suryan, R.M., Irons, D.B., Kaufman, M., Benson, J., Jodice, P.G.R., Roby, D.D. & Brown, E.D. (2002) 489 Short-term fluctuations in forage fish availability and the effect on prey selection and brood-rearing in 490 the Black-legged Kittiwake Rissa tridactyla. Marine Ecology Progress Series, 236, 273–287. 491 Tella, J.L., Banos-Villalba, A., Hernández-Brito, D., Rojas, A., Pacifico, E., Diaz-Luque, J.A., Carrete, 492 M., Blanco, G. & Hiraldo, F. (2015) Parrots as overlooked seed dispersers. Frontiers in Ecology and 493 the Environment, 13, 338–339. 494 Tornberg, R., & Reif, V. (2007) Assessing the diet of birds of prey: a comparison of prey items found in 495 nests and images. Ornis Fennica, 84, 21. 496 Walter, C.B., Cooper, J. & Suter, W. (1987) Diet of Swift Tern chicks in the Saldanha Bay Region, 497 South Africa. Ostrich, 58, 49–53. 498 Wilson, R.P. (1984) An improved stomach pump for penguins and other seabirds. Journal of Field 499 Ornithology, 55, 109–112. 500 Woehler, E.J., Saviolli, J.Y., Bezerra-Francini, C.L., Neves, T. & Bastos-Francini, R. (2013). Insect prey 501 of breeding South American Terns. Marine Ornithology, 41, 199–200. 502 503 Supporting Information 504 Additional Supporting Information may be found in the online version of this article. 505 Appendix S1. Additional methods and results for estimation of prey standard length. 506 Appendix S2. Results of the comparison between photo-sampling and regurgitation. 507 508 Table 1. Prey families in the greater crested tern diet identified by photo-sampling on Robben Island during the 509 2013, 2014 and 2015 breeding seasons. N = number of prey items identified. 510 Prey type Family Species N Fish Engraulidae 1 16206 Dussumieriidae 1 2557 Scomberesocidae 1 1658 Syngnathidae 2 866 Clupeidae 1 545 Carangidae 2 409 Gonorynchidae 1 351 Atherinidae 1 198 Mugilidae 1 117 Merlucciidae 1 76 Pomatomidae 1 67 Soleidae Unid. 63 Champsodontidae 1 58 Clinidae Unid. 63 Clinidae 5 25 Holocentridae 1 47 Nomeidae 2 46 Triglidae Unid. 43 Blenniidae Unid. 38 Myctophidae 1 23 Gobiidae Unid. 9 Gobiidae 1 23 Scombridae 1 22 Scyliorhinidae Unid. 16 Macrouridae Unid. 12 Congridae 1 12 Coryphaenidae 1 10 Sebastidae Unid. 6 Gobiesocidae 1 6 Trichiuridae 2 9 Tetraodontidae Unid. 5 Cheilodactylidae 1 4 Ophichthidae Unid. 5 Bregmacerotidae Unid. 4 Ophidiidae 1 3 Ophidiidae Unid. 1 Sparidae 2 3 Congiopodidae Unid. 2 Berycidae 1 2 Centriscidae 1 1 Chlorophthalmidae 1 1 Batrachoididae 1 1 Aulostomidae Unid. 1 Cephalopods Loliginidae 2 54 Sepiidae 1 85 Octopodidae 1 11 Crustaceans Squillidae 1 244 Brachyura* Unid. 2 Portunidae 1 1 Palinuridae 1 3 Insects Gryllidae 1 191 Gryllotalpidae 1 2 Sphingidae 1 2 Sphingidae Unid. 1 Coleoptera ** Unid. 1 *Infraorder, **Order 511 Figure Legeneds 512 Fig. 1 a) Examples of capturing a photo-sample of an adult greater crested terns carrying prey to the colony without 513 causing disturbance to nesting birds and (b) the resulting close-up image of the prey used for identification 514 (anchovy) and standard length measurements. From c to n: Examples of tern prey items: c) sardine Sardinops 515 sagax; d) Atlantic saury Scomberesox saurus; e) multi-prey load (3 anchovy and 1 sardine); f) dolphinfish 516 Coryphaena hippurus; g) snake eel Ophichthidae sp.; h) sole Austroglossus sp.; i) longsnout pipe fish Syngnathus 517 temminckii; l) shyshark Haploblepharus sp.; m) cuttlefish Sepia vermiculata; n) common squid Loligo vulgaris; o) 518 rock lobster Jasus lalandii; p) two-spotted cricket Gryllus bimaculatus . 519 520 Fig. 2 Example of the application (in ImageJ) of the ‘line selection tool’ to measure the linear distances for the 521 three morphometric parameters: (1) eye diameter (E); (2) head width (H) and (3) operculum width (O). 522 523 Fig. 3 Sample-based rarefaction and species accumulation curves for greater crested tern diet at Robben Island. 524 Accumulation curves show the observed species accumulation from 1510 photo-samples (orange points) and 198 525 regurgitations (blue points) collected on 18 and 19 April 2015. Rarefaction curves (solid lines) and 95% asymptotic 526 confidence intervals (shaded areas) are based on 1,000 random permutations (shown as light grey points) of the 527 observed data. The rarefaction curve for regurgitations is extrapolated (blue dashed line) based on a GAM fit to the 528 photo-sampling, assuming an equal species accumulation rate beyond the range of the observed data. Vertical 529 dotted lines show sample sizes of 190 and 1500 used to compare the methods. 530 531 Fig. 4 Accuracy of estimated standard length (𝑆𝐿) (y-axis) compared with actual SL values (x-axis) of anchovy 532 from photographs in ImageJ using allometric regressions based on estimates of eye diameter (𝐸, open orange 533 circles), operculum width (𝑂, open blue circles), head diameter (𝐻, purple open circles) and the mean of all three 534 (mean 𝑆𝐿, black closed circles). The mean culmen length of greater crested terns (61.2 mm) was used as the 535 reference length to scale the pixel-based length estimates in ImageJ. The grey dashed line represents 100% 536 accuracy. 537 538 Fig. 5 Frequency distribution of anchovy standard length from photo-samples and regurgitations (A = Robben 539 Island; B = Seal Island). 540