focal mean using base R is faster than using terra

focal-mean.R

library(terra)
library(Matrix)

# create a sample raster
r <- rast(ncols=10, nrows=10, ext(0, 10, 0, 10))
x <- 1:ncell(r)
values(r) <- x

# why is there NaN in here????
adj_raw <- adjacent(r, 1:ncell(r), "queen")

# create a sparse matrix in a ragged array
adj_list <- vector("list", nrow(adj_raw))
for (i in seq_len(nrow(adj_raw))) {
  adj_list[[i]] <- c(i, filter_row(adj_raw[i,]))
}

# prepare to create the sparse matrix
i <- rep.int(x, lengths(adj_list))
j <- unlist(adj_list)
w <- rep(1, length(i))

# create the sparse matrix
mr <- sparseMatrix(
  i, j, x = w, repr = "R",
) |> forceSymmetric()

mc <- sparseMatrix(
  i, j, x = w, repr = "C",
) |> forceSymmetric()


# calculate focal mean
bench::mark(
  base_r_row = (mr %*% x) / Matrix::rowSums(mr),
  base_r_col = (mc %*% x) / Matrix::rowSums(mc),
  terra = focal(r, fun = "mean", na.rm = TRUE),
  check = FALSE
)
#> # A tibble: 3 × 13
#> expression      min  median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
#> <bch:expr> <bch:tm> <bch:t>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
#> 1 base_r_row   21.4µs 24.27µs    36686.    9.03KB     3.67  9999     1      273ms
#> 2 base_r_col  18.66µs 19.76µs    48278.    3.31KB     0    10000     0      207ms
#> 3 terra        1.04ms  1.09ms      877.    4.05KB     2.17   404     1      460ms
#> # ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

what's filter_row() ?

Ope! @mdsumner it removes the NaN values I believe it was defined like so:

filter_row <- function(.row) .row[!is.nan(.row)]

Ah I see

Appreciate this example, I never understood use of Matrix but have toyed with my own versions of sparsity