Trophic interactions play a key role in maintaining ecological balance. In urban environments, avian predation has been demonstrated to be particularly important due to its effects on community structure, pest control, and nutrient cycling. As humanity relies upon ecosystem services for sustenance, and with 70% of the global population projected to reside in urban areas by 2050, understanding the impact of urbanization on avian predation is becoming increasingly important. This study investigates the impacts of urban microclimates, shaped by impervious surfaces and green/blue infrastructure, and human-induced disturbances, on avian predation in urban areas, with the aim of testing the increased disturbance hypothesis. To assess the avian predation rate, plasticine caterpillars were placed in Quercus robur trees in the city of Amsterdam for a period of two months. The analyses evaluated the impact of artificial lighting at night, human population density, the urban heat island effect, impervious surfaces, vegetation, noise pollution, and water bodies on predation rates. The results indicated a substantial increase in predation during the second month, which was likely caused by an increase in naïve fledglings or elevated ambient temperatures. Noise pollution was identified as the most frequent and robust predictor of predation, consistently leading to a reduction in predation rates, possibly due to avoidance behavior. Other predictors exhibited substantial temporal and spatial variability. The variables related to urbanization increased predation in the initial month, suggesting that insectivorous birds prey on areas with higher illumination and temperature. However, the effect diminished in the subsequent month, potentially due to the increased daylight hours or a reduction in heating effects. During the second month, all predictors exhibited a negative effect on predation, thereby supporting the increasing disturbance hypothesis. These findings underscore the complex relationship between urban factors and avian predation, emphasizing the necessity for mitigation efforts in urban planning.

Study Site

The city of Amsterdam (The Netherlands) was selected as the study site for this experiment due to several factors. The city has a high human population density with 5.336 people per km² (CBS 2023), a significant vegetation presence, numerous water canals, and a negligible elevation difference throughout the city that might affect trophic interactions (Dean et al. 2024). In consideration of the findings presented by Hernández-Agüero et al. (2020), which demonstrated predation differences between tree species, this study exclusively utilized Quercus robur trees. The trees were selected from a georeferenced list of all trees in Amsterdam. The selection criteria included species and a maximum height of twelve meters, which was necessary for us to reach the branches. A total of 2,882 Q. robur trees were identified as eligible from a list of 259,431 trees in Amsterdam. The human population density, NDVI index and water percentage surrounding each tree were estimated, and the resulting ranges were divided into five categories. Only one tree per category of each was randomly selected. Ultimately, 38 trees were chosen based on variations in HPD, vegetation, and water presence (Figure 1).

Avian Predation Data Collection

The field experiment involved the placement of artificial plasticine caterpillars (N = 114) of three distinct colors as a proxy for prey organisms in Q. robur trees (N = 38) throughout Amsterdam. The coloration of prey organisms can influence the detection and selection of prey by avian species, as these animals primarily detect prey through visual cues (Ruxton et al. 2018). By using three different colors in our experiment, we ensured sufficient variability in predation pressure among study sites, at least for one color, independent of the time interval between tree visits, as is common in similar studies (e.g. Alonso-Crespo & Hernández-Agüero 2023, Hernández-Agüero et al 2024b). This approach permitted the comparison of predation levels between trees, even in instances where predation for a particular color was either nearly absent or so high as to preclude the observation of differences in the number of attacks. All colors demonstrated sufficient variability among trees, and thus no color was excluded from the analysis.

The models were placed in week 15 (2024), with two reviews occurring at four-week intervals. This was done to account for potential temporal variability, with avian predation rates being highest in the summer months, thereby strengthening the potential for seasonal variability (Hernández-Agüero et al. 2020). During the reviews, we identified attack marks at the coarse taxonomic level with the assistance of the standardization proposed by Low et al. (2014). First, the avian predation marks were recorded, and the caterpillars were molded back to their original shape. The methodology employed was similar to that used by Alonso-Crespo and Hernández-Agüero (2023) (see Appendix Figure A1), with the caterpillars attached to tree branches with metal wires (diameter 0.5 mm). The average length of each caterpillar was approximately 30mm, with a diameter of 4mm. Given that invertebrate predation rates are higher when placed near the ground (Lövei and Ferrante 2017), and our objective was to observe avian predation, the caterpillars were placed at heights ranging between 1.5 and 2 meters. The plasticine caterpillar models were non-toxic and unscented, consisting of a mixture of waxes, inert substances, and colored pigments (STAEDTLER MARS GmbH & Co KG 2017). This material addresses the concerns raised by of Rößler et al. (2018) regarding the avoidance of polyvinyl chloride due to potential ingestion hazards. All models were molded exclusively by hand, resulting in slight differences between the models. However, this is the most reasonable method for creating caterpillar-like shapes with plasticine (Bateman et al. 2017).

Spatial Analyses

The averages of the urbanization-related predictors, vegetation presence, and water presence were calculated through zonal statistics in QGIS version 3.34.3 (QGIS 2023). This was conducted for buffers of 200 meters surrounding the trees where measurements were taken, in accordance with the methodology described by Valdés-Correcher et al. (2022). In addition to the 200-meter buffers, larger buffer zones of 400, 600, 800, and 1,000 meters were constructed in order to account for the dynamic nature of avian movement, and to examine the effects of these predictors across spatial scales.

We requested satellite imagery from the Sentinel-II satellite through Copernicus, and the European Space Agency (ESA 2023) processed all satellite imagery. The MultiSpectral Instrument (MSI), comprising 13 spectral bands, and the high spatial resolution of Sentinel-II (10m/20m/60m) facilitate precise remote sensing analyses. The selection of imagery was based on several criteria to ensure its usability for reliable remote sensing. These criteria included the necessity for the imagery to originate from the same satellite, for the sensing periods to fall as close to the measurements as possible, and for cloud coverage to be limited to below 5%. These criteria adhere to the guidelines for remote sensing set forth by Lefsky and Cohen (2003) and Rembold et al. (2020).

As this study incorporated remote sensing analyses for both vegetation and water infrastructures through the Normalized Difference Vegetation Index/NDVI (see Appendix Equation A1)  and the Normalized Difference Water Index/NDWI (see Appendix Equation A2), and given that it was necessary to isolate vegetation and water areas to prevent misguided averages in their respective analyses, a remote sensing analysis on impervious surfaces through the Normalized Difference Built-up Index/NDBI (see Appendix Equation A3) was also conducted. The isolation of water, impervious surfaces, and vegetation was achieved by identifying the predictors and reclassifying them. Subsequently, inversions of the predictors were employed to isolate the respective predictors.

Moreover, we acquired population data for Amsterdam from WorldPop (2018) to assess HPD (people/hectare), with a resolution of 100 x 100 meters. For ALAN, we extracted the modelled light emissions at night in 2022 (Watt/cm²/steradian) provided by the National Oceanic and Atmospheric Administration (NOAA) using the Visible Infrared Imager Radiometer Suite (VIIRS). This model was subsequently modified by the National Institute for Public Health and the Environment (RIVM in Dutch; RIVM 2023). We acquired noise data from RIVM (2020), which provides estimated noise pollution/Lden (level day-evening-night) in 2021 in dB. Anthropogenic noise pollution is defined as artificial noise originating from air traffic, industry, neighborhoods, rail traffic, and road traffic (Radford et al. 2012). The model encompassed all widely accepted primary sources of noise pollution, with the exception of noise from neighborhoods.

Data Analyses

We assessed the associations between standardized averages of the outlined predictors and the avian predation rates recorded with the plasticine caterpillars in R environment version 4.2.2 (R Core Team 2022), including packages ‘car’ (Fox and Weisberg 2019), ‘ggplot2’ (Wickham 2016), ‘Hmisc’ (Harrell 2024), ‘lme4’ (Bates et al. 2015), ‘MuMIn’ (Barton 2021), and ‘visreg’ (Breheny and Burchett 2017). A correlation analysis was conducted between variables within the 600-meter buffer zone, as this is the buffer zone of intermediate size. An additional check for multicollinearity was conducted using the Variance Inflation Factor to confirm the efficacy of the multicollinearity correction (see Appendix Table A2). As HPD, impervious surface, UHI effect, ALAN and NDVI exhibited high correlation coefficients (above ± 0.8; Table A1 in Appendix), and thus were not viable to be included in the model without violating the assumptions of the model, principal component analyses were used to group all those correlated variables with the ‘prcomp’ function. We evaluated the extent to which the predictors exerted a significant influence on avian predation by employing generalized linear mixed models (GLMM) with a Poisson distribution. In each model, we included the first two axes of the PCA analysis, along with noise, water proportion and caterpillar color (to account for color predation differences) as prediction variables. We selected the best fitting GLMMs using the corrected Akaike information criterion (AICc) with the ‘dredge’ function. AICc corrects for small sample sizes and should be used instead of AIC when the ratio between the sample size and the number of parameters (including the intercept) falls below 40 (Burnham & Anderson, 2002), which is the case in our study. Models with a difference in AICc ≤ 2 were selected as best fitting models and the other models were omitted. By using dredge from the ‘MuMIn’ package, we conducted a systematic analysis that runs all possible model configurations and ranks them based on AICc values and model complexity.