Generating Seasonal Abundance and Range Maps and Stats

Matt Strimas-Mackey, Tom Auer, Daniel Fink

2019-04-02

Outline

  1. Introduction
  2. Data Download
  3. Seasonal Abundance
  4. Abundance Maps
  5. Range Maps
  6. Regional Statistics

Introduction

This vignette describes how to recreate the seasonal map products found on the eBird Status and Trends pages. First, the vignette will cover how to average the abundance data seasonally, then it will proceed with examples of making abundance maps, aggregating and smoothing data for range maps, and, finally, provide examples of calculating regional summary statistic. Throughout this vignette we will use the simplified example dataset available through the ebirdst_download() function, which consists of the Yellow-bellied Sapsucker data spatially subset to the state of Michigan. However, these methods can be applied to the full datasets for any of the eBird Status and Trends species.

Let’s begin by loading all the packages we’ll need for this vignette.

library(ebirdst)
library(raster)
library(velox)
library(sf)
library(smoothr)
library(rnaturalearth)
library(dplyr)
library(tidyr)
library(stringr)
library(ggplot2)
# resolve namespace conflicts
select <- dplyr::select

Data Download

Before we do anything else, we’ll need to download and load the example abundance data for Yellow-bellied Sapsucker.

# download to a temp directory for the vigette
# in practice, change to permanent directory the status and trends downloads
sp_path <- ebirdst_download(species = "example_data")
# load the abundance data
# this automaticaaly labels layers with their dates
abd <- load_raster("abundance_umean", path = sp_path)

The abundance data now consist of a RasterStack object with 52 layers, each corresponding to the relative abundance for a single week.

We’ll also need some additional spatial data (state and country borders, lakes, etc.) to provide context for the maps we’ll make. Natural Earth is the best source for free map data and there’s an associated R package (rnaturalearth) for accessing the data. For all the maps in this vignette, we’ll use the Mollweide projection, an excellent option for continental scale maps.

mollweide <- "+proj=moll +lon_0=-90 +x_0=0 +y_0=0 +ellps=WGS84"
ne_scale <- 50

# land polygon
ne_land <- ne_countries(scale = ne_scale, returnclass = "sf") %>%
  filter(continent %in% c("North America", "South America")) %>%
  st_set_precision(1e6) %>%
  st_union() %>% 
  st_geometry()
# function to subset other features to those  within this land area
wh_subset <- function(x) {
  in_wh <- as.logical(st_intersects(x, ne_land, sparse = FALSE))
  st_transform(x[in_wh], crs = mollweide)
}
# country lines
ne_country_lines <- ne_download(scale = ne_scale, category = "cultural",
                                type = "admin_0_boundary_lines_land",
                                returnclass = "sf") %>%
  st_geometry() %>%
  wh_subset()
# state lines
ne_state_lines <- ne_download(scale = ne_scale, category = "cultural",
                              type = "admin_1_states_provinces_lines",
                              returnclass = "sf") %>%
  st_geometry() %>%
  wh_subset()
# rivers
ne_rivers <- ne_download(scale = ne_scale, category = "physical",
                         type = "rivers_lake_centerlines",
                         returnclass = "sf") %>%
  st_geometry() %>%
  wh_subset()
# lakes
ne_lakes <- ne_download(scale = ne_scale, category = "physical",
                        type = "lakes",
                        returnclass = "sf") %>%
  st_geometry() %>%
  wh_subset()
ne_land <- st_transform(ne_land, crs = mollweide)

Seasonal Abundance

Generally, the seasons for eBird Status and Trends products are defined on a species-specific basis through expert review. For information on the details of defining seasons, please see the seasons section of the FAQ. While it is certainly possible to define your own seasons when making seasonal abundance and range maps, if you want to recreate the products with the same seasons as the website, you’ll need to use the definitions included in the ebirdst_runs data frame included in this package.

Let’s start by getting the seasonal definitions for Yellow-bellied Sapsucker and transforming the data into a more usable format. For some species, expert review may have indicated that the models are poor in certain seasons. These problematic seasons are identified by missing dates in ebirdst_runs for the season in question. Although all seasons passed review for Yellow-bellied Sapsucker, for generality I’ll add an additional column (passed) indicating whether a seasonal model passed review.

# subset to the yellow-bellied sapsucker season definitions
yebsap_dates <- filter(ebirdst_runs, species_code == "yebsap") %>% 
  # just keep the seasonal definition columns
  select(setdiff(matches("(start)|(end)"), matches("year_round"))) %>% 
  # transpose
  gather("label", "date") %>% 
  # spread data so start and end dates are in separate columns
  separate(label, c("season", "start_end"), "_(?=s|e)") %>% 
  spread(start_end, date) %>% 
  select(season, start_dt, end_dt)
# did the season pass review
yebsap_dates <- mutate(yebsap_dates, pass = !(is.na(start_dt) | is.na(end_dt)))
yebsap_dates

Now, we’ll use these season definitions to assign each of the weekly abundance layers to a season.

# dates for each abundance layer
weeks <- parse_raster_dates(abd)
# assign to seasons
weeks_season <- rep(NA_character_, length(weeks))
for (i in seq_len(nrow(yebsap_dates))) {
  s <- yebsap_dates[i, ]
  # skip seasona assignment if season failed
  if (!s$pass) {
    next()
  }
  # handle seasons cross jan 1 separately
  if (s$start_dt <= s$end_dt) {
    in_season <- weeks >= s$start_dt & weeks <= s$end_dt
  } else {
    in_season <- weeks >= s$start_dt | weeks <= s$end_dt
  }
  weeks_season[in_season] <- s$season
}
table(weeks_season)

Finally, let’s average all the weeks within each season to produce seasonal relative abundance rasters.

# drop weeks not assigned to season
week_pass <- !is.na(weeks_season)
abd <- abd[[which(week_pass)]]
weeks <- weeks[week_pass]
weeks_season <- weeks_season[week_pass]
# average over weeks in season
mean_season <- function(s) {
  calc(abd[[which(weeks_season == s)]], mean, na.rm = TRUE)
}
seasons <- unique(weeks_season)
abd_season <- lapply(seasons, mean_season) %>% 
  stack() %>% 
  setNames(seasons)
abd_season

Abundance Maps

Before we create maps of seasonal relative abundance we need to account for two addition considerations that the online maps make: whether to split pre- and post-breeding migration and whether to explicitly show the year-round distribution as a separate color. For species that use different paths for their two migrations (less than 40% overlap) we show pre-breeding migration (green) and post-breeding migration (yellow) separately.

migration_threshold <- 0.4
mig_seasons <- c("prebreeding_migration", "postbreeding_migration")
if (all(mig_seasons %in% names(abd_season))) {
  # identify areas with abundance in only one season
  abd_nz <- abd_season[[mig_seasons]] > 0
  just_pre <- mask(abd_nz[["prebreeding_migration"]],
                   abd_nz[["postbreeding_migration"]], 
                   maskvalue = 1)
  just_post <- mask(abd_nz[["postbreeding_migration"]],
                   abd_nz[["prebreeding_migration"]], 
                   maskvalue = 1)
  # count the number of cells with abundance in only one season
  n_just <- cellStats(stack(just_pre, just_post), sum)
  n_all <- cellStats(abd_nz, sum)
  # is the proportion of one season cells above the 40% threshold
  split_migration <- max(n_just / n_all, na.rm = TRUE) >= migration_threshold
} else {
  split_migration <- FALSE
}
n_just / n_all
split_migration

In this case, there is essentially complete overlap between pre- and post-breeding migration, so we won’t split them into separate colors on the map. Next, we’ll calculate the average annual abundance, which we’ll display as a separate color if at least 1% of the range is occupied in all four seasons.

threshold_yearround <- 0.01
# decide whether to show year-round layer
if (nlayers(abd_season) == 4) {
  # annual abundance
  abd_yr <- calc(abd, fun = mean, na.rm = TRUE)
  # mask out cells that aren't occupied year-round
  year_round <- calc(abd_season > 0, fun = sum, na.rm = TRUE) == 4
  abd_yr_mask <- mask(abd_yr, year_round, maskvalue = 0)
  # determine proportion of celss that are occupied year round
  n_yr <- cellStats(abd_yr_mask > 0, sum)
  n_an <- cellStats(abd_yr > 0, sum)
  # only show year round abundance if it's above 1% of range threshold
  show_yearround <- ((n_yr / n_an) >= threshold_yearround)
} else {
  show_yearround <- FALSE
}
show_yearround

Mapping relative abundance across the full-annual cycle requires a specialized set of color bins in order to ensure the full range of abundance values is effectively displayed. The function calc_bins() calculates the optimal break points of bins for Status and Trends abundance data.

bin_breaks <- calc_bins(abd_season)

Now that everything is in place, we can actually make the seasonal relative abundance map! Note that the Status and Trends maps distinguish between regions of zero abundance and regions with no prediction, which are displayed in different shades of gray, and regions with non-zero abundance, shown in color. To account for this, we’ll need to generate a raster layer delineating the region within which predictions were made.

# project the abundance data to mollweide
# use nearest neighbour resampling to preserve true zeros
abd_season_proj <- projectRaster(abd_season, crs = mollweide, method = "ngb")
# determine spatial extent for plotting
ext <- calc_full_extent(abd_season_proj)
# set the plotting order of the seasons
season_order <- c("postbreeding_migration", "prebreeding_migration", 
                  "nonbreeding", "breeding")

# prediction region, cells with predicted value in at least one week
pred_region <- calc(abd_season_proj, mean, na.rm = TRUE)
# mask to land area
ne_land_buffer <- st_buffer(ne_land, dist = max(res(pred_region)) / 2)
pred_region <- mask(pred_region, as_Spatial(ne_land_buffer))

# remove zeros from abundnace layers
abd_no_zero <- subs(abd_season_proj, data.frame(from = 0, to = NA), 
                    subsWithNA = FALSE)

# set up plot area
par(mar = c(0 , 0, 0, 0))
plot(ne_land, col = "#eeeeee", border = NA, 
     xlim = c(ext@xmin, ext@xmax),
     ylim = c(ext@ymin, ext@ymax))
# prediction region and explicit zeros
plot(pred_region, col = "#dddddd", maxpixels = raster::ncell(pred_region),
     legend = FALSE, add = TRUE)
# lakes
plot(ne_lakes, col = "#ffffff", border =  "#444444", lwd = 0.5, add = TRUE)
# land border
plot(ne_land, col = NA, border = "#444444", lwd = 0.5, add = TRUE)
# seasonal layer
plot_seasons <- intersect(season_order, names(abd_no_zero))
for (s in plot_seasons) {
  # handle splitting of migration seasons into different colors
  if (!split_migration && s %in% c("prebreeding_migration", 
                                   "postbreeding_migration")) {
    pal_season <- "migration"
    
  } else {
    pal_season <- s
  }
  pal <- abundance_palette(length(bin_breaks$bins) - 1, pal_season)
  plot(abd_no_zero[[s]], col = pal, breaks = bin_breaks$bins,
       maxpixels = ncell(abd_no_zero[[s]]),
       legend = FALSE, add = TRUE)
}
# year round
if (show_yearround) {
  year_round_proj <- projectRaster(year_round, crs = mollweide, method = "ngb")
  plot(year_round_proj, 
       col = abundance_palette(length(bin_breaks$bins) - 1, "year_round"), 
       breaks = bin_breaks$bins,
       maxpixels = ncell(year_round_proj),
       legend = FALSE, add = TRUE)
}
# linework
plot(ne_rivers, col = "#ffffff", lwd = 0.75, add = TRUE)
plot(ne_state_lines, col = "#ffffff", lwd = 1.5, add = TRUE)
plot(ne_country_lines, col = "#ffffff", lwd = 2, add = TRUE)

# legends
legend_seasons <- plot_seasons
if (split_migration) {
  legend_seasons[legend_seasons %in% c("prebreeding_migration", 
                                       "postbreeding_migration")] <- "migration"
  legend_seasons <- unique(legend_seasons)
}
if (show_yearround) {
  legend_seasons <- c(legend_seasons, "year_round")
}
# thin out labels
lbl_at <- bin_breaks$bins^bin_breaks$power
lbl_at <- c(min(lbl_at), median(lbl_at), max(lbl_at))
lbl <- lbl_at^(1 / bin_breaks$power)
lbl <- format(round(lbl, 2), nsmall = 2)
# plot legends
for (i in seq_along(legend_seasons)) {
  pal <- abundance_palette(length(bin_breaks$bins) - 1, legend_seasons[i])
  if (i == 1) {
    axis_args <- list(at = lbl_at, labels = lbl, line = -1,
                      cex.axis = 0.75, lwd = 0)
  } else {
    axis_args <- list(at = lbl_at, labels = rep("", 3),
                      cex.axis = 0.75, lwd = 0)
  }
  legend_title <- legend_seasons[i] %>% 
    str_replace_all("_", " ") %>% 
    str_to_title()
  fields::image.plot(zlim = range(bin_breaks$bins^bin_breaks$power), 
                   legend.only = TRUE, 
                   breaks = bin_breaks$bins^bin_breaks$power, col = pal,
                   smallplot = c(0.05, 0.35, 0.01 + 0.06 * i, 0.03 + 0.06 * i),
                   horizontal = TRUE,
                   axis.args = axis_args,
                   legend.args = list(text = legend_title, side = 3, 
                                      cex = 0.9, col = "black", line = 0.1))
}
title("Yellow-bellied Sapsucker Relative Abundance (birds per km/hr)", 
      line = -1, cex.main = 1)

Range Maps

The eBird Status and Trends range maps delineate the boundary of regions with non-zero relative abundance for a given species. We’ll start by aggregating the raster layers to a coarser resolution to speed up processing, then convert the boundaries of non-zero abundance regions to polygons. We’ll also convert the prediction areas to polygons so we can distinguish where the species is predicted to not occur from where it was not modeled.

# aggregate
abd_season_agg <- aggregate(abd_season_proj, fact = 3)
# raster to polygon, one season at a time
range <- list()
pred_area <- list()
for (s in names(abd_season_agg)) {
  # range
  range[[s]] <- rasterToPolygons(abd_season_agg[[s]], 
                                 fun = function(y) {y > 0}, 
                                 digits = 6) %>% 
    st_as_sfc() %>% 
    # combine polygon pieces into a single multipolygon
    st_set_precision(1e6) %>% 
    st_union() %>% 
    st_sf() %>% 
    # tag layers with season
    mutate(season = s, layer = "range")
  # prediction area
  pred_area[[s]] <- rasterToPolygons(abd_season_agg[[s]], 
                                     fun = function(y) {!is.na(y)}, 
                                     digits = 6) %>% 
    st_as_sfc() %>% 
    # combine polygon pieces into a single multipolygon
    st_set_precision(1e6) %>% 
    st_union() %>% 
    st_sf() %>% 
    # tag layers with season
    mutate(season = s, layer = "prediction_area")
}
# combine the sf objects for all seasons
range <- rbind(do.call(rbind, range), do.call(rbind, pred_area))
row.names(range) <- NULL
print(range)

Converting from raster to polygons frequently yields tiny fragments of polygons and jagged polygon edges, which result in maps that aren’t very aesthetically pleasing. To address this, we’ll apply some algorithms from the smoothr package to clean up the polygons.

# clean and smooth
cell_area <- (1.5 * prod(res(abd_season_agg)))
range_smooth <- range %>% 
  # drop fragment polygons smaller than 1.5 times the aggregated cell size
  drop_crumbs(threshold = cell_area) %>% 
  # drop holes in polygons smaller than 1.5 times the aggregated cell size
  fill_holes(threshold = cell_area) %>% 
  # smooth the polygon edges
  smooth(method = "ksmooth", smoothness = 2)
# clip zeros to land border, range to buffered land to handle coastal species
range_split <- split(range_smooth, range_smooth$layer)
range_smooth <- rbind(
  st_intersection(range_split$range, ne_land_buffer),
  st_intersection(range_split$prediction_area, ne_land))

Finally, we can produce a seasonal range map in a similar fashion to the abundance map above.

# range map color palette
range_palette <- c(nonbreeding = "#1d6996",
                   prebreeding_migration = "#73af48",
                   breeding = "#cc503e",
                   postbreeding_migration = "#edad08",
                   migration = "#edad08",
                   year_round = "#6f4070")

# set up plot area
par(mar = c(0 , 0, 0, 0))
plot(ne_land, col = "#eeeeee", border = NA, 
     xlim = c(ext@xmin, ext@xmax),
     ylim = c(ext@ymin, ext@ymax))
# prediction region and explicit zeros
annual_pred_area <- filter(range_smooth, layer == "prediction_area") %>% 
  st_union()
plot(annual_pred_area, col = "#dddddd", border = NA, add = TRUE)
# lakes
plot(ne_lakes, col = "#ffffff", border =  "#444444", lwd = 0.5, add = TRUE)
# land border
plot(ne_land, col = NA, border = "#444444", lwd = 0.5, add = TRUE)
# seasonal layer
for (s in intersect(season_order, unique(range_smooth$season))) {
  # handle splitting of migration seasons into different colors
  if (!split_migration && s %in% c("prebreeding_migration", 
                                   "postbreeding_migration")) {
    col_season <- "migration"
  } else {
    col_season <- s
  }
  rng_season <- filter(range_smooth, season == s, layer == "range") %>% 
    st_geometry()
  plot(rng_season, col = range_palette[col_season], border = NA, add = TRUE)
}
# year round
if (show_yearround) {
  # find common area between all seasons
  range_combined <- filter(range_smooth, layer == "range")
  range_yearround <- range_combined[1, ]
  range_combined <- sf::st_geometry(range_combined)
  for (i in 2:length(range_combined)) {
    range_yearround <- sf::st_intersection(range_yearround, range_combined[i])
  }
  plot(st_geometry(range_yearround), 
       col = range_palette["year_round"], border = NA, 
       add = TRUE)
}
# linework
plot(ne_rivers, col = "#ffffff", lwd = 0.75, add = TRUE)
plot(ne_state_lines, col = "#ffffff", lwd = 1.5, add = TRUE)
plot(ne_country_lines, col = "#ffffff", lwd = 2, add = TRUE)

# legend
rng_legend <- rev(range_palette[legend_seasons])
names(rng_legend) <- names(rng_legend) %>% 
    str_replace_all("_", " ") %>% 
    str_to_title()
legend("bottomleft", legend = names(rng_legend), fill = rng_legend)
title("Yellow-bellied Sapsucker Seasonal Range Map", 
      line = -1, cex.main = 1)

Regional Statistics

In addition to maps and visualizations, the eBird Status and Trends website provides a set of statistics summarizing the spatial data over regions. In the US and Canada, statistics are provided for each state and province, and throughout North America they are provided for Bird Conservation Regions (BCRs). The five regional statistics are:

  1. Mean relative abundance: the average estimated relative abundance within the given region for a given season.
  2. Percentage of seasonal North American population: the sum of the estimated relative abundance within the selected region divided by the sum of the estimated relative abundance across North America.
  3. Percentage of region occupied: the percentage of the selected region within the range boundary of a species for a given season.
  4. Percentage of North American range in region: the fraction of a species’ total North American range that falls within the selected region.
  5. Days of occupation in region: the number of days that a species occupies the selected region, with occupation being defined as spatially covering the selected region by at least 5% based on estimated relative abundances averaged across the given season.

These statistics can be downloaded from the Status and Trends website for all regions and species; however, there may be situations where you want to calculate them over different regions than those provided. With that in mind, this vignette will cover how to calculate a couple of these statistics: mean relative abundance and days of occupation. The remaining 3 statistics can be calculated following the same approach with minor modifications.

Since the example data used in this vignette is restricted to Michigan, we’ll calculate the statistics over the counties in Michigan; however, this approach can easily be extended to any set of regions. Let’s start by downloading county boundaries for Michigan.

mi_counties <- getData("GADM", country = "USA", level = 2, path = tempdir()) %>% 
  st_as_sf() %>% 
  filter(NAME_1 == "Michigan") %>% 
  select(county = NAME_2, county_code = HASC_2) %>% 
  st_transform(crs = projection(abd)) %>% 
  # remove lakes which aren't true counties
  filter(county_code != "US.MI.WB")

We’ll also aggregate the data to coarser resolution to perform these calculations. This isn’t necessary for the example dataset, but when this is extended to the full dataset it will become critical for making these calculations efficiently.

abd_season_agg <- aggregate(abd_season, fact = 3, fun = mean, na.rm = TRUE)
abd_agg <- aggregate(abd, fact = 3, fun = mean, na.rm = TRUE)

Mean Relative Abundance

Mean relative abundance is the easiest statistics to calculate: we simply average all the raster cells within each region polygon. This could be done with the raster function extract(); however, we’ll extract() from the velox package instead, which is orders of magnitude faster.

Let’s make a quick map comparing the breeding and non-breeding mean relative abundance within counties in Michigan.

Days of Occupation

Days of occupation is the most challenging statistic to calculate because it involves several steps. First, for each week, we need to determine if a region (i.e. a county) is “occupied”. We define occupied as at least 5% of the cells within the region having a non-zero abundance.

Next, we tally up the number of weeks each region is occupied, then convert to days by multiplying by 7.

Finally, let’s produce a map comparing days occupation in the breeding and non-breeding season.