Despite being common and widespread, the diets of Astur cooperii (Cooper’s Hawk) and Accipiter striatus (Sharp-shinned Hawk) have been poorly studied. Although some dietary datasets indicate both feed heavily on birds, most studies have focused on nest site or bird feeder observations, thus precluding comprehensive assessments of the species’ diets. Recently, increased popularity in citizen science programs like iNaturalist and eBird have allowed an unprecedented glimpse into the ecology of both species that may allow dietary analyses. Herein, we used iNaturalist, a photograph-vouchered citizen science platform, to perform an assessment of Astur cooperii and Accipiter striatus diets, examining patterns of prey size and species preferences for each predator. We hypothesized that both species would feed primarily on birds, however, we suspected some mammals might also be selected for. Across >70,000 raptor observations, we found 4,206 (6.79% of records) and 619 (4.75% of records) predation events for Astur cooperii and Accipiter striatus, respectively. After accounting for spatial heterogeneity in observation data, we found Astur cooperii selected for Columbiformes, small Galliformes, Colaptes auratus (Northern Flicker), and invasive Passeriformes, with their preferred prey item matching their most observed prey item, C. livia (Rock Pigeon). Despite their proclivity for avian prey, Astur cooperii also fed on mammals, selecting heavily for Rattus norvegicus (Norway rats). The smaller-bodied Accipiter striatus preferred small Passeriformes, though Columbiformes were also selected for. Both species exhibited bimodal prey size selection, likely driven by differential use of Columbiformes and other prey items. Collectively, these data support the hypothesis that Astur cooperii and Accipiter striatus are avian specialists, despite Astur cooperii’s regular predation on small mammals.

In August, 2023, we downloaded all research-grade records of Astur cooperii and Accipiter striatus from inaturalist.org using the iNaturalist online data download tool. We then visually reviewed each record to remove any misidentifications/duplicate records and to determine the presence or absence of prey within the photograph/series of photographs. Only prey seen in physical contact with the Astur cooperii/Accipiter striatus was counted, to remove uncertainty on the source of the prey item (Figure 1). If a record contained multiple photographs, all were reviewed.

Observed prey were identified to the lowest possible taxonomic group, with birds identified first by the corresponding author, then shared with several other PhD-level ornithologists to confirm identities. Mammals were preliminarily identified by corresponding author and confirmed by a PhD-level mammalogist who studies North American mammal diversity and ecology. All reptiles were identified by a PhD-level herpetologist who studies North American reptile diversity. For avian prey, some individuals could not be identified to the species level (e.g., if silhouetted or blurry), but could be assigned to a coarse size class. For such records, we categorized avian prey items into “small” (≤ 50g), “medium” (51g – 174g), and “large” (≥ 175g) classifications. We did not conduct a similar size class assessment with mammals or reptiles because, with many poor quality images of those taxa, we found it unclear whether we could see the entire prey item or just a portion of the prey item (e.g., a mass of muscle with hair would clearly be of mammalian origin but unclear whether it was intact or from a larger carcass). Any instances of obvious scavenging, in particular, mesocarnivores like Felis domesticus (domestic cat), Procyon lotor (common raccoon) , and Didelphis virginiana (Virginia opossum) , were assumed to be scavenged and, thus, removed from analyses.

Weight assignments

Central to our analysis of prey weights was the development of a “prey weight database”. We initiated our prey weight database by generating a list of all species-level identifications of prey items (prey that could not be identified to species [e.g., “Sciurus sp.”] were not included in the weight analyses, save for birds with size class identities; discussed below). For each species, we obtained a “minimum” and “maximum” weight using Pyle (2022; for birds), Animal Diversity Web (animaldiversity.org), or other published scientific literature found via Google Scholar. For birds of unknown species assigned to a prey weight class, it was assigned a body weight, for analytical purposes, by first assigning it a semi-random “pseudo-species” identity and its weight was generated using that identity. The pseudo-species was assigned by selecting either a random bird of the same size class (small, medium, or large) from within 10 km of the predation event, or, if no other records occurred within 10km, the closest prey item’s species identity within the same size class. For all predation events involving birds (including individuals with pseudo-species identification), mammals, and reptiles identified to species, we assigned each a weight based on the minimum and maximum weights derived for that species from the literature. For each such record, we assigned the prey item a weight by sampling one value from a normal distribution where the mean was the midpoint between the minimum and maximum weight and the minimum weight was -2 standard deviations and the maximum weight was +2 standard deviations. We performed this using the rnorm function in base R (R Core Team 2024).

Weight analyses- spatial balancing

Because iNaturalist data, like most citizen science data, are heavily biased in space around urban centers (Dimson & Gillespie 2023), we corrected for this by thinning the data used for weight summary in a spatially balanced manner that effectively down-weighted records from locations with high densities of records and up-weighted more isolated records. Specifically, we used the terra package in R to map the coordinates of all Astur cooperii/Accipiter striatus predation observations (Hijmans 2025, R Core Team 2024). We then created a 10km moving window across the United States, selecting one predation event per 10km square, regardless of the density of records within. This created a more spatially balanced distribution of prey weights for each species across the United States. We then repeated this process one hundred times to generate uncertainty associated with our sampling procedure.

Prey selection ratios

Finally, to estimate selection ratios for birds used by Astur cooperii and Accipiter striatus, we used the gbif package in R to extract every research-grade iNaturalist observation for birds, mammals, and reptiles in the United States from the GBIF database (Chamberlain & Boettiger 2017, Chamberlain et al. 2025). Each Astur cooperii and Accipiter striatus predation record with known prey species identity was plotted, and a “background” sample of 1000 random GBIF observations respective to each prey class (i.e., 1000 random birds for each avian predation, 1000 mammals for each mammalian predation, etc.) were selected from within a 25km buffer of each point. These samples served as prey that would be “available” to hawks at each predation point from which we calculated the proportion of each species in the background data (m). This “available prey” analysis was repeated 1000 times, and the mean proportion of each species in the Astur cooperii / Accipiter striatus prey base (n) and standard deviation of n/m (σ) were calculated. Next, we generated selection ratios with the equation (m – n) / σ (Beavis & Mahama 2023). These selection ratios are typically interpreted as selection for a resource if that ratio is positive, and selection against a resource if negative. If zero or near zero, the selection ratio is interpreted as random. As an exploratory analysis, we regressed these selection ratios against average prey weights and fit a generalized additive mixed model (GAMM) to visualize trends within Astur cooperii and Accipiter striatus weight selection.