Build statistical models of vegetation cover

Given one or more sites and a model specification, builds a model of vegetation cover and report model assessment.

Usage

fit(
  site = NULL,
  datafile = "data",
  name = "",
  method = "rf",
  fitargs = NULL,
  vars = "{*}",
  exclude_vars = "",
  exclude_classes = NULL,
  include_classes = NULL,
  exclude_years = NULL,
  include_years = NULL,
  min_class = 500,
  reclass = c(13, 2),
  max_samples = NULL,
  years = NULL,
  minscore = 0,
  maxmissing = 20,
  max_miss_train = 0.2,
  top_importance = 20,
  holdout = NULL,
  bypoly = "bypoly01",
  byyear = NULL,
  blocks = NULL,
  auc = FALSE,
  hyper = NULL,
  notune = FALSE,
  resources = NULL,
  local = FALSE,
  trap = TRUE,
  comment = NULL
)

Arguments

site: Three letter site code, or vector of site names if fitting multiple sites
datafile: Name of data file. It must be an .RDS file, but exclude the extension. If fitting multiple sites, either use a single datafile name shared among sites, or a vector matching site.
name: Optional model name
method: One of rf for Random Forest, boost for AdaBoost. Default = rf.
fitargs: A named list of additional arguments to pass to the model (ranger or boost)
vars: Vector of variables to restrict analysis to. Default = {*}, all variables. vars is processed by find_orthos, and may include file names, portable names, search names and regular expressions of file and portable names.
exclude_vars: An optional vector of variables to exclude. As with vars, variables are processed by find_orthos
exclude_classes: Numeric vector of subclasses to exclude. This overrides fit_exclude that may be included in sites.txt.
include_classes: Numeric vector of subclasses to include - all other classes are dropped. include_classes overrides fit_exclude (in sites.txt) and exclude_classes.
exclude_years: A vector of one or more years of ground truth data to exclude (requires a year column in source data)
include_years: A vector of one or more years of ground truth data to include - all other years are dropped (requires a year column in source data). Overrides exclude_years.
min_class: Minimum number of training samples to allow in a class. All classes with fewer samples in training set as well as all classes with zero cases in the validation set will be dropped from the model. Use min_class = NULL to prevent dropping any classes.
reclass: Matrix or vector of paired classes to reclassify. Pass either a two column matrix, such that values in the first column are reclassifed to the second column, or a vector with pairs, reclass = c(13, 2, 3, 4), which would reclassify all 13s to 2 and 3s to 4, lumping each pair of classes. Reclassifying is not iterative, thus you could swap 1s and 2s with reclass = c(1, 2, 2, 1), not that you'd want to.
max_samples: Maximum number of samples to use - subsample if necessary
years: Vector of years to restrict variables to
minscore: Minimum score for orthos. Files with a minimum score of less than this are excluded from results. Default is 0, but rejected orthos are always excluded.
maxmissing: Maximum percent missing in orthos. Files with percent missing greater than this are excluded.
max_miss_train: Maximum proportion of missing training points allowed before a variable is dropped
top_importance: Number of variables to keep for variable importance
holdout: Proportion of points to hold out. For Random Forest, this specifies the size of the single validation set, while for boosting, it is the size of each of the testing and validation sets.
bypoly: Polygons to treat as holdout data. Supply the name of a bypoly cross-validation sequence in the sampled data. gather creates bypoly01 through bypoly05, with sequences of 1:10 for each subclass. Poly groups 1 and 6 will be used as holdouts. To specify different groups, use blocks = list(block = 'bypoly01', classes = c(2, 7), for instance.
byyear: One or more years to treat as holdout data. If supplied, this superceeds bypoly.
blocks: An alternative to holding out random points. Specify a named list with block = <name of block column>, classes = <vector of block classes to hold out>. Set this up by creating a shapefile corresponding to ground truth data with a variable block that contains integer block classes, and placing it in the blocks/ directory for the site. gather and sample will collect and process block data for you to use here.
auc: If TRUE, calculate class probabilities so we can calculate AUC
hyper: Hyperparameters. To be defined.
notune: If TRUE, don't do hyperparameter tuning. This can cost you a few percent in CCR, but will speed the run up six-fold from the default.
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