Pollinators provide an important ecosystem function and service (Klein et al. 2007, Ollerton et al. 2011), but are declining worldwide (Potts et al. 2010, Zattara and Aizen 2021). Many pollinators rely on the availability of floral resources, both nectar and pollen, at broad scales across the landscape for survival (Hines and Hendrix 2005, Steffan-Dewenter and Westphal 2008, Du Clos et al. 2020). Furthermore, the abundance and diversity of floral resources in a landscape change over the growing season (Guezen and Forrest 2021). The provision of floral resources depends on the composition of the flowering plant community, which varies with habitat and land use (Mallinger et al. 2016). Some habitats, such as deciduous forests, may have a narrow period of high flower production, whereas other habitats may provide fewer resources over an extended period. Since most pollinators do not produce substantial quantities of honey to store floral resources, their population is limited by the time of year when floral resources are scarcest.

Human-modified land uses, such as agricultural and urban areas, may significantly alter the distribution of floral resources in space and time. For example, mass-flowering crops concentrate flowering to an intense, limited period, which has an effect on pollinator behaviour (Holzschuh et al. 2011, Holzschuh et al. 2013). Urban environments can provide important pollinator resources (Tew et al. 2021), though the prevalence of exotic species may also shift the pollinator community (Wilson and Jamieson 2019, Theodorou et al. 2020). Development also creates new floral habitats, such as roadside ditches.

Therefore, estimates of floral resources for pollinators must take into account the land use and land cover within a heterogeneous landscape in order to model variability over space and time (Lonsdorf et al. 2009). An important step in developing this understanding is characterising the landscape into land-cover classes that can be translated to potential pollinator communities (Koh et al. 2015). Further, fine-scale information may play an important role in understanding pollinator distributions in some landscapes (Lonsdorf et al. 2009). Here we describe the process we used to create a spatial dataset that classifies land cover into categories relevant to their flowering vegetation communities at a high resolution (1 m) within a region of Central New York State. This dataset can be combined with data on flowering area and flowering phenology of plant communities in each land-cover category to predict floral resources available to pollinators over the year (Iverson et al., in prep.).

The focal area of this dataset covers 12 counties in New York State, within the United States of America (USA): Cayuga, Chemung, Cortland, Monroe, Onondaga, Ontario, Schuyler, Seneca, Tioga, Tompkins, Wayne and Yates (Fig. 1). We produced dataset versions that include crop data for the years 2012-2019.

Figure 1.

Coverage of dataset includes 12 counties in New York State (USA): Cayuga, Chemung, Cortland, Monroe, Onondaga, Ontario, Schuyler, Seneca, Tioga, Tompkins, Wayne and Yates Counties.

We combined land-cover data relevant to estimating floral resources, including natural habitat types (including wetlands), crops, grasses (like pasture, hayfields, oldfields and urban lawns), roadside ditches and urban areas (see Table 2). This involved combining and reclassifying annual crop cover data from the Cropland Data Layer (CDL) (Boryan et al. 2011) and a natural habitat layer covering the northeast US and Atlantic Canada (Ferree and Anderson 2013) into classes relevant to predicting flowering plant communities. We then downscaled the land-cover classification information from the combined crop and habitat layer to a 1 m resolution lidar-based dataset that classifies, based on vegetation height and impervious cover (Chesapeake Conservancy 2020), for most (nine) of the counties within our study area. Counties not covered by this high resolution layer were still downscaled to 1-m resolution to match the rest of the data for further processing. To the downscaled data, we added wetland and waterbody delineations derived from the National Wetland Inventory (USFWS 2018) and our own delineations of roadside ditches, based on road vector data.

All input geographic datasets are publicly available from the sources listed in Table 1. We converted these layers to the same projected coordinate system, USA Contiguous Albers Equal Area Conic USGS (ESRI WKID: 102039), within the geographic coordinate system North American 1983 (EPSG: 4269). Geoprocessing was conducted using tools in ArcGIS ESRI (2020) available under the spatial analyst and data management licences and were scripted with the ArcGIS visual programming application “ModelBuilder”. The full modelling workflow is described in Suppl. material 1.

Crop and natural habitat land-cover information

High resolution landscape features

Combining crop and natural habitat information with high resolution landscape features

The order of data synthesis is outlined below. These are encoded as ArcGIS Modelbuilder tools that were developed for this project and are uploaded to the repository associated with this article. More details on the geoprocessing routines within each step are described in Suppl. material 1.

1. Prepare high-resolution data layers: reclassify relevant vegetation height classes (where available), rasterise national wetland inventory and estimate ditches.
2. Combine high resolution layers prepared in step 1.
3. Reclassify natural habitat 30 m raster to represent vegetation categories relevant to floral resources and erase wetland classes in preparation for combination with high resolution wetland data.
4. Combine reclassified natural habitat layer prepared in step 3 with crop layer, reclassified to reflect relevant floral resources land-cover classes.
5. Downscale the combined natural habitat and crop layer to 1 m resolution and add the high-resolution vegetation, wetland and ditch features.
6. In three counties where the high resolution vegetation data are not available, use developed land-cover classes or percent permeable land cover to estimate urban lawn and tree coverage.

Since we downscaled the 30 m resolution input data to 1 m resolution, the final land-cover data layer may not always match the classification indicated by the originating land-cover layer at a given point. This is due to the inclusion of fine-scale landscape information from the high resolution layers (the Chesapeake Conservancy, NWI and ditches layers). The additional details provided by these layers may indicate mismatches in vegetation type (e.g. trees mixed within field) or finer scale landscape features (e.g. ditches or small waterbodies), which were not included in the coarser resolution layers. In order to check that the data processing steps downscaled the 30 m resolution land-cover information with adequate fidelity, we compared the final land-cover class to the classes of the originating data layers using contingency tables based on 10,000 randomly placed points that sampled the land-cover identity in the final and input layers. In Table 4, we calculate the percent of the sample points whose land cover in the final layer matches the land cover of the originating layer ("Fidelity"). Higher percent fidelity classes deviate less from the originating layers indicated in the "Data origin" column of Table 4.

In general, agricultural classes are preserved in the downscaled dataset, with fidelity values above 80% and, in many cases, above 90%. This reflects the homogeneity of agricultural areas, which makes it unlikely that the high resolution vegetation layer would indicate an unexpected vegetation type (e.g. trees in alfalfa cells). Exceptions to this could be along field edges bordering forest or other contrasting land-cover types or cases where the CDL was misclassified (Lark et al. 2021).

Vegetation in developed land-cover classes have 100% fidelity because cells with these two land-cover classes are only found outside of the coverage of the high resolution land-cover dataset and generally do not coincide with waterbodies or ditches that would change their identity in the final layer. Within the coverage of the high resolution land-cover dataset, low vegetation is reclassified as "lawn" and tree canopy is reclassified as "urban trees".

Natural areas have lower fidelity, likely because these land covers are more heterogeneous. Classifications at 30 m resolution represent the most predominant land cover, whereas the 1 m vegetation data can better reflect a mix of land-cover types. Our downscaling process approximated this by taking land-cover information from nearby areas with the appropriate vegetation type, but this would lead to more cases where the final land cover differed from the class of the originating layer. This is shown in more detail in Suppl. material 2, which provides the full contingency tables across all land-cover combinations. Shrubland, which has 42% fidelity in Table 4, also occurs prominently in areas classified as agricultural land in the TNC dataset (38% in Suppl. material 2). This could indicate woody vegetation in old fields undergoing succession that border shrublands. Likewise, wet emergent vegetation falling outside the original NWI delineations occur mainly within shrub wetland, likely representing low wetland vegetation that would not have been noted in the NWI dataset, but was mapped by the high resolution vegetation layer.

In addition to the full contingency tables associated with Table 4, Suppl. material 2 also contains contingency tables comparing how well the original land cover is preserved in the final land-cover data, i.e. the percentage of sample points whose original classification matches the final dataset. These tables give an idea of the composition of the coarser resolution cells, once downscaled. The two types of contingency tables are roughly analogous to "user's" and "producer's" accuracy typically used in remote sensing classification (Lillesand et al. 2015), except that we compare the final land-cover classifications to professionally-produced input land-cover datasets rather than field data. User's accuracy estimates how often the map class is present on the ground, while producer's accuracy estimates how often the habitat on the ground is mapped correctly. An exhaustive field validation of the final land-cover dataset is beyond the scope of this project, though extensive methods documentation and accuracy assessments are available for many of the input layers, for example, Boryan et al. (2011), Ferree and Anderson (2013), Lark et al. (2021). Likewise, the accuracy of flowering ditch locations were not validated to field conditions, but could be the subject of future resesarch.

We estimated the error associated with predicting lawn and urban tree cover in counties without high resolution data by using the developed land-cover factors in Table 3 to calculate urban lawn and tree cover in the nine counties where this information is available. We calculated the root mean square error (RMSE) comparing between high resolution coverage estimates and category-based predictions for 100 points randomly placed in developed areas, for buffers ranging from 15 m to 1 km radius. The results in Table 5 show that the RMSE of both lawn and urban tree percent cover estimates decreases (i.e. prediction accuracy improves) with increasing buffer size. An alternative method of estimating lawn and urban tree cover using continuous permeable land-cover data had similar error values (Suppl. material 1, Step 6). However, we recommend estimating lawn and urban tree coverage using the simpler category-based method presented here, as the alternative method does not offer any improvement in prediction accuracy while adding more processing steps.

The output floral resources land-cover layers (Fig. 2) are stored as 1 m resolution rasters in the USA Contiguous Albers Equal Area Conic USGS projection (ESRI WKID: 102039). Data are available on Zenodo. The dataset covers 12 counties: Cayuga, Chemung, Cortland, Monroe, Onondaga, Ontario, Schuyler, Seneca, Tioga, Tompkins, Wayne and Yates. Of these, nine counties (Cayuga, Chemung, Cortland, Onondaga, Ontario, Schuyler, Tioga, Tompkins and Yates) have 1 m resolution delineations of low vegetation and tree canopies, which take on appropriate classifications based on the underlying 30 m data or nearby appropriate land covers, as described above. Outside of these counties, land covers follow the delineations of the 30 m layers, except for wetland, water and ditch features.

Figure 2.

Final land-cover layer (2016 shown). Area with high resolution vegetation data is outlined in black (majority of south-eastern region). The three counties without detailed vegetation data are available with urban vegetation represented by alternative methods shown in the two inset callouts in the upper left: either as a continuous percent permeability value (upper callout) or as low and medium developed intensity categories (middle callout). In the area with high resolution vegetation data that covers most of the region, urban vegetation consists of lawn or urban tree categories (example in lower callout).

There are two versions of the dataset, each consisting of eight rasters representing CDL crop data from years 2012-2019. The versions differ in their representation of the developed areas in the counties beyond the coverage of the high resolution vegetation layer (i.e. Monroe, Seneca and Wayne Counties). A simplified version classifies developed areas into "low" and "medium" development categories. An alternative version converts these areas to continuous values representing the percent permeable area. Either of these variables can be converted to an estimate of lawn and urban tree coverage, though we recommend the category-based version (see Usage Notes). Both versions are available online in the Zenodo repository.

Object name

Format names and versions

Creation dates

Dataset creators

Dataset contributors

License

Repository name

Repository location

Publication date

These layers are intended to be used to estimate available floral resources. Additional data collected within these habitats on flower phenology, abundance, size and community composition can be combined to understand landscape patterns in floral resources and associated bee abundance and richness over the growing season for this region (Kammerer et al. 2022, Iverson et al., in prep.). This information can also be the basis of floral availability scores that are the input of spatially-explicit pollinator abundance models, such as presented by Lonsdorf et al. (2009) or Zulian et al. (2013). The land-cover estimates, obtained from this dataset, are expected to be more accurate when aggregated over an area and should not be interpreted as a representation of on-the-ground conditions for a given location (i.e. pixel or point). Further, the land-cover categories in this dataset have combined multiple land-cover classes from the input layers based on similarities in floral resource characteristics, for example, flowering phenology, species composition and abundance. These groupings may not be appropriate outside of this original purpose.

Note that developed areas in Monroe, Seneca and Wayne Counties do not have high resolution spatial data of urban vegetation. Instead, the user must estimate urban vegetation in these counties by converting from either categorical development classes ("low" and "medium" categories) or a continuous percentage gradient of "permeable" land cover (the inverse of impervious cover). In order to estimate expected proportions of lawn and urban tree coverage within these areas, we present conversion factors, based on the relationships between the development categories and the proportional coverage of the two urban vegetation types (Table 3). We recommend using these categorical factors over converting percent permeable land cover (described in Step 6 of Suppl. material 1), because the former is simpler to use and has equivalent accuracy.

Suppl. material 1 gives a detailed description of geoprocessing and the associated ArcGIS Modelbuilder toolbox that was developed is available in a Zenodo repository. The data layers used to create this dataset are publicly available and described in Table 1.

Aaron L. Iverson conceived of the project, conducted fieldwork, procured funding, supervised and contributed to writing the manuscript. Kevin Li conducted geoprocessing and analysis and wrote the manuscript. Jon R.B. Fisher and Alison G. Power provided light edits to the manuscript and minor guidance to the project design.