Urban greenspaces are a haven for wildlife in densely populated cities. Wildlife use greenspaces for resource acquisition, shelter, and travel across urbanized landscapes. Greenspace metrics such as herbaceous or woody landcover, size, patchiness, and human land use influence species richness. Considering these metrics in urban greenspace design could influence valuable resources for wildlife, therefore the goals of this study were to: 1) determine wildlife communities in different greenspace types 2) identify and quantify greenspace metrics to determine relationships with wildlife at various spatial scales. To monitor wildlife, twenty-six camera traps were set in eastern Los Angeles County, California; greenspace metrics were gathered using 3m landcover supervised classification and FRAGSTATS. Non-metric multidimensional scaling was used to determine wildlife community composition and a generalized linear mixed model was selected to determine influence of greenspace metrics on richness at four scales (200m, 500m, 1km, and 2km). Urban dweller adapted species like the Northern mockingbird, Virginia opossum, and desert cottontail, were commonly found in suburban yards and urban open modified greenspaces like a commercial plant nursery. Whereas species like the band-tailed pigeon, gray fox, and black bear had increased presence in natural greenspaces. Large scale greenspaces, 1km and 2km in size, with high herbaceous cover, either as increasing aggregated patches or increased patchiness, and moderate levels of woody cover, positively influence species richness. At smaller scales, 200m and 500m, low and high levels of herbaceous cover and high levels of woody cover, strongly and positively influence richness. These results suggest fragmented greenspaces with varying levels of herbaceous and woody cover contribute to increasing wildlife richness in urban regions. From the perspective of urban planning, consideration of ecological scale is important to ensure developing greenspaces can support wildlife and ecological functions.

One 30 km transect was created sampling twenty-six sites using camera traps to monitor wildlife from Diamond Bar to the San Gabriel Mountains covering 11 of the 31 cities in San Gabriel Valley (Fig. 1). Camera site selection was based on standardized methodology developed by the Urban Wildlife Information Network (Magle et al. 2019). A 2km buffer was created around the transect and all cameras were placed randomly at least 1km from each other within the buffer. Finalized site selection and placement were based on randomized locations generated initially with available vegetation for camera placement, available site access and approval from public, private, and city landowners.

Camera sites (n=26) were defined within 5 greenspace categories after site selection (Fig. 2) and vary in size, location, amount of greenspace (GS) and human-built surfaces (HB), and average amount of human activity (HA- people/per camera trap night captured during all five seasons). Additional land cover types, such as bareground and water were measured per site but were not metrics of interest in this study. In order to categorize camera sites into the 5 greenspace categories, they were each assessed within a 155 meter buffer (smallest study site size) to standardize greenspace and human-built surface percentages, around each camera site using the land cover classification on ArcMap 10.7 (ESRI, Redlands, CA, USA). These categories include 1) yard space: greenspace, primarily lawns, attached to houses surrounded by human-built structures with low human activity and foot traffic (n=4; GS: 54- 76%, HB: 16-43%, HA: 15.59), 2) open-modified space: greenspace with low to moderate human foot traffic and access, including a landfill, private commercial nursery, and urban farmland (n=5; GS: 42- 75%, HB: 6-55%, HA: 25.39), 3) natural areas: vegetated natural spaces, like trails, with very low human use and access (n=2; GS: 89- 99%, HB: 0%, HA: 0.014), 4) urban-wildland interface: natural greenspace adjacent to houses and human-development with low access and foot traffic (n=7; GS: 70- 99%, HB: 0-11%. HA: 4.67), and 5) recreation: city parks with primarily ornamental vegetation (some with/without patches of natural land) with high access and foot traffic (n=8; GS: 54- 82%, HB: 7-39%, HA: 226.82). All final sites were dependent upon California State Polytechnic University-Pomona’s Risk Management approval and signed agreements with landowner or management personnel, and accessibility during daylight hours.

One Bushnell (Overland Park, KS) Core Low Glow Trail camera (Model- 119936C) was placed at each study site to monitor wildlife presence in the area. Settings for cameras were: image size (4K), capture number (1 photo), interval (30 sec), and sensor level (normal). Cameras were secured to a tree or T-post 0.5 -1 meter above the ground, faced towards any animal sign at initial set-up (e.g., animal trail, scat), and set at an angle to capture approximately 50% of the ground and 50% of the air in the photos to assess for birds and mammals. Cameras were deployed to capture photos for once per month, with a camera check mid-month to swap SD cards and batteries. Data was gathered for a total of one year from February 2022 to January 2023, sampling one month per season (Fall- October, Winter- January and February, Spring- April and Summer- July). No lures were used in this study.

EcoAssist, an AI program, incorporates the model MegaDetector to detect wildlife presence in photos, and was used to identify and separate photos with an animal, human, or object from empty photos (van Lunteren 2023). Photos were then analyzed and tagged twice for confirmation from two independent volunteers when looking for species presence, human presence, and objects. If there was disagreement between tags, photos were sent to validation from which the primary author (AE) determined the correct tag or entered the correct tag if both were incorrect. Final taxonomic groups of interest were mammals and birds as camera traps were placed and angled to capture both groups using both ground and air. This project was exempt from Cal Poly Pomona Institutional Animal Care and Use Committee (IACUC) since there was no manipulation or handling of animals. IRB approval to capture human activity was not necessary since all landowners approved of camera locations and signs were placed in public locations indicating individual photos may be taken from the wildlife camera.

Land Cover Classification

A supervised land cover classification raster was created using ERDAS Imagine 16.6 (Hexagon, Stockholm, Sweden) of 3m resolution imagery taken on December 1, 2021 from the Planet-Scope satellite (Planet Labs PBC). Land cover types were categorized into 5 classes including woody cover, herbaceous cover, human-built, bare ground, and water. These classes were initially chosen from urban wildlife studies then narrowed down to common land cover types across the camera sites in San Gabriel Valley and based on species presence from the camera traps (Kie 2000; Livingston 2003; Cadenasso 2007; Liu 2014; Dennis 2018; Moll 2018). It was important to distinguish greenspace as both herbaceous and woody cover to account for mammals that use greenspace to traverse across the urban matrix and for bird species that utilize tree canopy (Blair 1996; Kie 2000; Livingston 2003; Cadenasso 2007; Lombardi 2018). Herbaceous cover was primarily composed of natural grasses in natural areas and manicured lawns, such as in recreation sites and yard spaces, while woody cover consisted of trees and shrubs. A 69% accuracy land cover classification of eastern Los Angeles County was achieved by assessing aerial images from Google Maps using 250 random points. An accuracy assessment was conducted as a pre-processing step in order to determine if the spatial model was truly representative of what is on the ground. Land cover percentages were extracted after the fact for analysis purposes. This accuracy is probably due to a larger survey area, eastern Los Angeles County (1,044 ) being used in the assessment compared to using the smaller, study transect (60 ). While 85% land cover accuracy attainment is often cited in literature, it is not widely applicable among all land cover analyses, not initially developed to support local, small scale land cover analyses, and might be unrealistically inflated (Foody 2008).

Buffers of 200m (Desert cottontail, adult female during breeding season 22.5 acres within woodlands of Michigan, USA) (Haugen 1942), 500m (Virginia opossum, adult female ~15-35 ha, adult male ~20- 70 ha, urban setting in Missouri) (Harmon et al. 2004, Wright 2012), 1km (Northern raccoon, adults 25- 50 ha in urban regions of Illinois) (Prange 2004), and 2km (Coyote, breeding adult male and female 3.5 - 4.5 km2 in metropolitan Illinois) (Gehrt 2009) in radii were created and clipped around each camera site (Examples shown in Fig. 3) to account for patterns that influence wildlife diversity at different scales based on circular home ranges of common urban species. A buffer of 500m, 1km, and 2km accounts for intermediate dispersal and foraging distances of common urban adapted birds such as the Northern mockingbird (Mimus polyglottos), mourning dove (Zenaida macroura), house finch (Carpodacus mexicanus), and American crow (Corvus brachyrhynchos) (Crooks et al. 2004; Matthies et al. 2017; Aberle et al. 2020). Buffers were created using the ‘Create Buffers’ tool around each camera site in ArcMap 10.7 (ESRI, Redlands, CA, USA) and the land cover classification raster was clipped at the buffer sizes using the Clip Raster tool for use in FRAGSTATS.

Class level landscape characteristics, hereon known as metrics, (Table 1), determined from literature review and the FRAGSTATS manual version 4 (McGarigal and Marks 1995, McGarigal 2015), were extracted from each of the clipped buffers from the land cover raster and analyzed using FRAGSTATS 4.2 software. These metrics were chosen based on the most commonly used metrics relating biodiversity to landscape ecology (Neel et al. 2004; Schindler et al. 2008; Perotto-Baldivieso et al. 2009; Tolessa et al. 2016; Mata et al. 2018; Lombardi et al. 2020). Patch area and distance from another patch were found to be the strongest predictors of bobcat, coyote, gray fox, raccoon, striped skunk, and Virginia opossum distribution (Crooks 2002). The influence of percent land cover of herbaceous, tree, urban, and water, oftentimes is species specific with some species like the bobcat and gray fox to prefer increased herbaceous cover (Rodriguez et al. 2021). Amongst the importance of studying landscape metrics at a local scale, is analyzing metrics like heterogeneity of habitat, fragmentation, and landcover at the class-level to determine create biodiversity assessments (Schindler et. al 2007).

We ran a non-metric multidimensional scaling (NMDS) analysis with Bray-Curtis dissimilarity using R (Vienna, Austria) 4.2.0 (vegan package) to detect distinctions of mammal and bird species communities within greenspaces of different land use types. The stress value for the NMDS data was 0 < x <0.2. Pearson’s correlation coefficient (r) was calculated and a multidimensional scaling (MDS) analysis was conducted in R 4.2.0 (magrittr, dplyr, ggpubr packages) for landscape metrics to eliminate those that may be highly correlated to each other (> 0.7, < -0.7). Focusing on species richness for mammals and birds at each site, we used SpadeR (Species-richness Prediction And Diversity Estimation in R) (Chao 2015) from the online R application Shiny (Chao 2015), to develop estimated species richness values factoring in the number of trap nights each camera was in use and species detection per trap night which included number of uniques (species detected in one trap night) and duplicates (species detected in two trap nights) (Chao 2019). Species detection was determined by noting presence/absence of a species per trap day along with total trap days per site to create incidence-frequency data. Chao2 estimator, developed as a lower bound, uses uniques and duplicates to estimate the number of undetected species and has been used in another study to estimate species richness of mammals and birds for camera trap data (Chengcheng et al. 2018). iChao2 estimator was selected to estimate species richness because it has an improved lower bound compared to Chao2, good accuracy, reduced bias compared to traditional estimators, and improves confidence interval coverage (Chiu et al. 2014; Chao 2019). Comparisons of richness values from camera trap data and iChao2 estimates (with 95% CI) per site and land use category are shown in Supporting Information.

To understand the relationship between estimated species richness and landscape metrics, we used a generalized linear mixed model (GLMM) with a negative-binomial distribution in R 4.2.0 (lme4 package) for each scale. Estimated species richness was the response variable and percent of land cover of woody (PLAND Woody), herbaceous cover (PLAND Herbaceous), woody patch density (PD Woody), herbaceous patch density (PD Herbaceous), patch area mean woody/herbaceous (AREA_MN Woody/Herbaceous), edge density of woody/herbaceous cover (ED_Woody/Herbaceous), patch cohesion index of woody and herbaceous cover (COHESION Woody/Herbaceous), and largest patch index (LPI Woody/Herbaceous) were proposed fixed effects. Study sites nested within land use type were used as a random effect. To further narrow metric selection and prevent overfitting we plotted linear relationships for response-fixed factor relationships. We established a threshold of three or more positive or negative correlations among the site types (Tredennick et al. 2021). Akaike Information Criterion (AIC) values were used to compare and favor models with lower AIC values and normal residual diagnostic plots to determine fixed effects for the final model. The final models for 200m, 500m, 1km, and 2km examining estimated species richness and landscape metrics contained percent landscape woody cover (PLAND_Woody), percent landscape herbaceous cover (PLAND_Herbaceous), patch density woody cover (PD_Woody), patch density herbaceous cover (PD_Herbaceous), edge density herbaceous (ED_Herbaceous) (Table 1) and all interactions were rescaled for standardization (scale between -1 to 1).

For visualization purposes we developed a constant at low (-1), medium (0), and high levels (1) of fixed effects to demonstrate the influence on richness for outcomes of the GLMMs and present an landscape buffer example from existing sites to visualize landscapes that represent the graphs as a whole. Two and three way interaction graphs factor in the dependence of each of the fixed effects to each other in order to predict multiple fixed effects on species richness. Two and three way interactions take priority over a single fixed effect. Post model diagnostic plots of residuals and Q-Q plots were used to visualize the fit of each model and if high dispersion or high variance were found those models were eliminated.

Concrete jungle to urban oasis: evaluating scale, vegetation cover, and aggregation of urban greenspaces on wildlife

Data files

Abstract

Description of the data and file structure

Files and variables

File: ChosenMetrics_HerbWoody_Class1km.xlsx

Variables

File: ChosenMetricsClass_1km.xlsx

Variables

File: ChosenMetrics_HerbWoody_Class2km.xlsx

Variables

File: ChosenMetrics_Class2km.xlsx

Variables

File: ChosenMetricsClass_200m.xlsx

Variables

File: ChosenMetricsClass_500m.xlsx

Variables

File: NMDS_SpeciesSpecific.xlsx

Variables

File: iChao2est_richness.xlsx

Variables

File: FinalRData_ModelsRichLand_txt.txt

File: FinalRData_ModelsRichLand.docx

Code/software

Land Cover Classification

Concrete jungle to urban oasis: evaluating scale, vegetation cover, and aggregation of urban greenspaces on wildlife

Data files

Abstract

README: Concrete jungle to urban oasis: evaluating scale, vegetation cover, and aggregation of urban greenspaces on wildlife

Description of the data and file structure

Files and variables

File: ChosenMetrics_HerbWoody_Class1km.xlsx

Variables

File: ChosenMetricsClass_1km.xlsx

Variables

File: ChosenMetrics_HerbWoody_Class2km.xlsx

Variables

File: ChosenMetrics_Class2km.xlsx

Variables

File: ChosenMetricsClass_200m.xlsx

Variables

File: ChosenMetricsClass_500m.xlsx

Variables

File: NMDS_SpeciesSpecific.xlsx

Variables

File: iChao2est_richness.xlsx

Variables

File: FinalRData_ModelsRichLand_txt.txt

File: FinalRData_ModelsRichLand.docx

Code/software

Methods

Land Cover Classification