Produce geoTIFF maps of predicted vegetation cover from fitted models

Console command to launch a prediction run, typically as a batch job on Unity. Dispatches to do_map for RF/AdaBoost models or do_unet_map for U-Net models, determined automatically from the fits database.

Usage

map(
  fit,
  site = NULL,
  clip = NULL,
  result = NULL,
  which = "all",
  write_probs = FALSE,
  use_distance_weights = TRUE,
  requirecuda = TRUE,
  resources = NULL,
  local = FALSE,
  trap = FALSE,
  comment = NULL
)

Arguments

fit: Fit id in the fits database, fit object, or path to a .RDS with a fit object. U-Net models must be specified by fit id.
site: Three letter site code. If fitting from a fit id that was built on a single site, you may omit site to map the same site (this is the most common situation). If you want to map sites other than the site the model was built on, or the model was built on multiple sites, site is required.
clip: Optional clip, vector of xmin, xmax, ymin, ymax
result: Optional result name. Default is map_<site>_<fit id>_[clip_<size>_ha]; if a result name is specified, the result will be map_<result>_<site>_<fit id>_[clip_<size>_ha], retaining the site and fit id, as omitting these breaks your ability to track maps back to the fits they're based on.
which: For U-Net models: which model(s) to use for prediction. One of 'all' (default, ensemble of all CV folds), 'full' (full retrained model), or an integer CV fold number. Ignored for RF/AdaBoost models.
write_probs: For U-Net models: if TRUE, write per-class probability layers alongside the classification. Ignored for RF/AdaBoost models.
use_distance_weights: For U-Net models: if TRUE (default), weight patch contributions by distance to the nearest patch edge when averaging overlapping predictions, reducing visible tile seams. Set FALSE for uniform averaging. Ignored for RF/AdaBoost models.
requirecuda: If TRUE (default), abort immediately if CUDA is not available rather than silently falling back to CPU. Set to FALSE only for testing without a GPU.
resources: Slurm launch resources. See launch. These take priority over the function's defaults.
local: If TRUE, run locally; otherwise, spawn a batch run on Unity
trap: If TRUE, trap errors in local mode; if FALSE, use normal R error handling. Use this for debugging. If you get unrecovered errors, the job won't be added to the jobs database. Has no effect if local = FALSE.
comment: Optional launch / slurmcollie comment

Details

Note: When making U-Net maps, you may want to run unet_prep_map first to build patches, as that step (automatic if they don't exist yet) doesn't require a GPU node.

Note: if you're running this in local mode, multiple runs in a row in the same R session may blow out memory, thanks to R/Python memory shenanigans. If you run out of memory in this situation, restart R between runs. This does not apply to batch runs on Unity.