Multi-scale model assessment

This vignette walks through how to use waywiser to assess model predictions at multiple spatial scales, using the ny_trees data in waywiser, adapted from that post.

First things first, we’ll set up our environment, loading a few packages and telling sf to download the coordinate reference system for our data, if needed:

library(sf)
#> Linking to GEOS 3.11.1, GDAL 3.6.4, PROJ 9.1.1; sf_use_s2() is TRUE
library(tidyr)
library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(waywiser)

invisible(sf_proj_network(TRUE))

The data we’re working with is extremely simple, reflecting the number of trees and amount of aboveground biomass (“AGB”, the total amount of aboveground woody bits) at a number of plots across New York State. We can plot it to see that there’s some obvious spatial dependence in this data – certain regions have clusters of much higher AGB values, while other areas (such as the area around New York City to the south) have clusters of much lower AGB.

library(ggplot2)

ny_trees %>%
  ggplot() +
  geom_sf(aes(color = agb), alpha = 0.4) +
  scale_color_distiller(palette = "Greens", direction = 1)

Because our focus here is on model assessment, not model fitting, we’re going to use an extremely simple linear regression to try and model AGB across the state. We’ll predict AGB as being a linear function of the number of trees at each plot, and then we’re going to use this model to predict expected AGB:

agb_lm <- lm(agb ~ n_trees, ny_trees)
ny_trees$predicted <- predict(agb_lm, ny_trees)

Now we’re ready to perform our multi-scale assessments. The ww_multi_scale() function supports two different methods for performing assessments: first, you can pass arguments to sf::st_make_grid() (via ...), specifying the sort of grids that you want to make. For instance, if we wanted to make grids with apothems (the distance from the middle of a grid cell to the middle of its sides) ranging from 10km to 100km long, we can call the function like this:

cell_sizes <- seq(10, 100, 10) * 1000
ny_multi_scale <- ww_multi_scale(
  ny_trees,
  agb,
  predicted,
  cellsize = cell_sizes
)

ny_multi_scale
#> # A tibble: 20 × 6
#>    .metric .estimator .estimate .grid_args       .grid            .notes  
#>    <chr>   <chr>          <dbl> <list>           <list>           <list>  
#>  1 rmse    standard       5077. <tibble [1 × 1]> <sf [3,350 × 5]> <tibble>
#>  2 mae     standard       3808. <tibble [1 × 1]> <sf [3,350 × 5]> <tibble>
#>  3 rmse    standard       3404. <tibble [1 × 1]> <sf [850 × 5]>   <tibble>
#>  4 mae     standard       2613. <tibble [1 × 1]> <sf [850 × 5]>   <tibble>
#>  5 rmse    standard       2893. <tibble [1 × 1]> <sf [391 × 5]>   <tibble>
#>  6 mae     standard       2197. <tibble [1 × 1]> <sf [391 × 5]>   <tibble>
#>  7 rmse    standard       2702. <tibble [1 × 1]> <sf [221 × 5]>   <tibble>
#>  8 mae     standard       2132. <tibble [1 × 1]> <sf [221 × 5]>   <tibble>
#>  9 rmse    standard       2541. <tibble [1 × 1]> <sf [140 × 5]>   <tibble>
#> 10 mae     standard       2032. <tibble [1 × 1]> <sf [140 × 5]>   <tibble>
#> 11 rmse    standard       2482. <tibble [1 × 1]> <sf [108 × 5]>   <tibble>
#> 12 mae     standard       1981. <tibble [1 × 1]> <sf [108 × 5]>   <tibble>
#> 13 rmse    standard       3517. <tibble [1 × 1]> <sf [80 × 5]>    <tibble>
#> 14 mae     standard       2294. <tibble [1 × 1]> <sf [80 × 5]>    <tibble>
#> 15 rmse    standard       2624. <tibble [1 × 1]> <sf [63 × 5]>    <tibble>
#> 16 mae     standard       2038. <tibble [1 × 1]> <sf [63 × 5]>    <tibble>
#> 17 rmse    standard       2101. <tibble [1 × 1]> <sf [48 × 5]>    <tibble>
#> 18 mae     standard       1664. <tibble [1 × 1]> <sf [48 × 5]>    <tibble>
#> 19 rmse    standard       2672. <tibble [1 × 1]> <sf [35 × 5]>    <tibble>
#> 20 mae     standard       1955. <tibble [1 × 1]> <sf [35 × 5]>    <tibble>

We’ve now got a tibble with estimates for our model’s RMSE and MAE at each scale of aggregation! We can use this information to better understand how our model does when predictions are being aggregated across larger units than a single plot; for instance, our model generally does better at larger scales of aggregation:

ny_multi_scale %>%
  unnest(.grid_args) %>%
  ggplot(aes(x = cellsize, y = .estimate, color = .metric)) +
  geom_line()

Note that we used the .grid_args column, which stores the arguments we used to make the grid, to associate our performance estimates with their corresponding cellsize.

In addition to our top-level performance estimates, our ny_multi_scale object also includes our true and estimated AGB, aggregated to each scale, in the .grid column. This lets us easily check what our predictions look like at each level of aggregation:

ny_multi_scale$.grid[[9]] %>%
  filter(!is.na(.estimate)) %>%
  ggplot(aes(fill = .estimate)) +
  geom_sf() +
  scale_fill_distiller(palette = "Greens", direction = 1)

In addition to specifying systematic grids via sf::st_make_grid(), ww_multi_scale() also allows you to provide your own aggregation units. For instance, we can use the tigris package to download census block group boundaries, as well as county and county subdivision boundaries, for the state of New York:

suppressPackageStartupMessages(library(tigris))

ny_block_groups <- block_groups("NY")
ny_county_subdivisions <- county_subdivisions("NY")
ny_counties <- counties("NY")

We can then provide those sf objects straight to ww_multi_scale:

ny_division_assessment <- ww_multi_scale(
  ny_trees,
  agb,
  predicted,
  grids = list(
    ny_block_groups,
    ny_county_subdivisions,
    ny_counties
  )
)

ny_division_assessment %>%
  mutate(
    division = rep(c("Block group", "County subdivision", "County"), each = 2)
  ) %>%
  ggplot(aes(x = division, y = .estimate, fill = .metric)) +
  geom_col(position = position_dodge())

By providing grids directly to ww_multi_scale(), we can see how well our model performs when we aggregate predictions to more semantically meaningful levels than the systematic grids.