fit_xgb_occupancy_models {surveyvoi}

Fit boosted regression tree models to predict occupancy

Description

Estimate probability of occupancy for a set of features in a set of planning units. Models are fitted using gradient boosted trees (via xgboost::xgb.train()).

Usage

fit_xgb_occupancy_models(
  site_data,
  feature_data,
  site_detection_columns,
  site_n_surveys_columns,
  site_env_vars_columns,
  feature_survey_sensitivity_column,
  feature_survey_specificity_column,
  xgb_tuning_parameters,
  xgb_early_stopping_rounds = rep(20, length(site_detection_columns)),
  xgb_n_rounds = rep(100, length(site_detection_columns)),
  n_folds = rep(5, length(site_detection_columns)),
  n_threads = 1,
  seed = 500,
  verbose = FALSE
)

Arguments

site_data

sf::sf() object with site data.

feature_data

base::data.frame() object with feature data.

site_detection_columns

character names of numeric columns in the argument to site_data that contain the proportion of surveys conducted within each site that detected each feature. Each column should correspond to a different feature, and contain a proportion value (between zero and one). If a site has not previously been surveyed, a value of zero should be used.

site_n_surveys_columns

character names of numeric columns in the argument to site_data that contain the total number of surveys conducted for each each feature within each site. Each column should correspond to a different feature, and contain a non-negative integer number (e.g. 0, 1, 2, 3). If a site has not previously been surveyed, a value of zero should be used.

site_env_vars_columns

character names of columns in the argument to site_data that contain environmental information for fitting updated occupancy models based on possible survey outcomes. Each column should correspond to a different environmental variable, and contain numeric, factor, or character data. No missing (NA) values are permitted in these columns.

feature_survey_sensitivity_column

character name of the column in the argument to feature_data that contains probability of future surveys correctly detecting a presence of each feature in a given site (i.e. the sensitivity of the survey methodology). This column should have numeric values that are between zero and one. No missing (NA) values are permitted in this column.

feature_survey_specificity_column

character name of the column in the argument to feature_data that contains probability of future surveys correctly detecting an absence of each feature in a given site (i.e. the specificity of the survey methodology). This column should have numeric values that are between zero and one. No missing (NA) values are permitted in this column.

xgb_tuning_parameters

list object containing the candidate parameter values for fitting models. Valid parameters include: "max_depth", "eta", "lambda", "min_child_weight", "subsample", "colsample_by_tree", "objective". See documentation for the params argument in xgboost::xgb.train() for more information.

xgb_early_stopping_rounds

numeric model rounds for parameter tuning. See xgboost::xgboost() for more information. Defaults to 10 for each feature.

xgb_n_rounds

numeric model rounds for model fitting See xgboost::xgboost() for more information. Defaults to 100 for each feature.

n_folds

numeric number of folds to split the training data into when fitting models for each feature. Defaults to 5 for each feature.

n_threads

integer number of threads to use for parameter tuning. Defaults to 1.

seed

integer initial random number generator state for model fitting. Defaults to 500.

verbose

logical indicating if information should be printed during computations. Defaults to FALSE.

Details

This function (i) prepares the data for model fitting, (ii) calibrates the tuning parameters for model fitting (see xgboost::xgb.train() for details on tuning parameters), (iii) generate predictions using the best found tuning parameters, and (iv) assess the performance of the best supported models. These analyses are performed separately for each feature. For a given feature:

1. The data are prepared for model fitting by partitioning the data using k-fold cross-validation (set via argument to n_folds). The training and evaluation folds are constructed in such a manner as to ensure that each training and evaluation fold contains at least one presence and one absence observation.

2. A grid search method is used to tune the model parameters. The candidate values for each parameter (specified via parameters) are used to generate a full set of parameter combinations, and these parameter combinations are subsequently used for tuning the models. To account for unbalanced datasets, the scale_pos_weight xgboost::xgboost() parameter is calculated as the mean value across each of the training folds (i.e. number of absence divided by number of presences per feature). For a given parameter combination, models are fit using k-fold cross- validation (via xgboost::xgb.cv()) – using the previously mentioned training and evaluation folds – and the True Skill Statistic (TSS) calculated using the data held out from each fold is used to quantify the performance (i.e. "test_tss_mean" column in output). These models are also fitted using the early_stopping_rounds parameter to reduce time-spent tuning models. If relevant, they are also fitted using the supplied weights (per by the argument to site_weights_data). After exploring the full set of parameter combinations, the best parameter combination is identified, and the associated parameter values and models are stored for later use.

3. The cross-validation models associated with the best parameter combination are used to generate predict the average probability that the feature occupies each site. These predictions include sites that have been surveyed before, and also sites that have not been surveyed before.

4. The performance of the cross-validation models is evaluated. Specifically, the TSS, sensitivity, and specificity statistics are calculated (if relevant, weighted by the argument to site_weights_data). These performance values are calculated using the models' training and evaluation folds.

Value

A list object containing:

parameters

list of list objects containing the best tuning parameters for each feature.

predictions

tibble::tibble() object containing predictions for each feature.

performance

tibble::tibble() object containing the performance of the best models for each feature. It contains the following columns:

feature

name of the feature.

train_tss_mean

mean TSS statistic for models calculated using training data in cross-validation.

train_tss_std

standard deviation in TSS statistics for models calculated using training data in cross-validation.

train_sensitivity_mean

mean sensitivity statistic for models calculated using training data in cross-validation.

train_sensitivity_std

standard deviation in sensitivity statistics for models calculated using training data in cross-validation.

train_specificity_mean

mean specificity statistic for models calculated using training data in cross-validation.

train_specificity_std

standard deviation in specificity statistics for models calculated using training data in cross-validation.

test_tss_mean

mean TSS statistic for models calculated using test data in cross-validation.

test_tss_std

standard deviation in TSS statistics for models calculated using test data in cross-validation.

test_sensitivity_mean

mean sensitivity statistic for models calculated using test data in cross-validation.

test_sensitivity_std

standard deviation in sensitivity statistics for models calculated using test data in cross-validation.

test_specificity_mean

mean specificity statistic for models calculated using test data in cross-validation.

test_specificity_std

standard deviation in specificity statistics for models calculated using test data in cross-validation.

Examples

## Not run: 
# set seeds for reproducibility
set.seed(123)

# simulate data for 30 sites, 2 features, and 3 environmental variables
site_data <- simulate_site_data(
  n_sites = 30, n_features = 2, n_env_vars = 3, prop = 0.1)
feature_data <- simulate_feature_data(n_features = 2, prop = 1)

# create list of possible tuning parameters for modeling
parameters <- list(eta = seq(0.1, 0.5, length.out = 3),
                   lambda = 10 ^ seq(-1.0, 0.0, length.out = 3),
                   objective = "binary:logistic")

# fit models
# note that we use 10 random search iterations here so that the example
# finishes quickly, you would probably want something like 1000+
results <- fit_xgb_occupancy_models(
   site_data, feature_data,
   c("f1", "f2"), c("n1", "n2"), c("e1", "e2", "e3"),
   "survey_sensitivity", "survey_specificity",
   n_folds = rep(5, 2), xgb_early_stopping_rounds = rep(100, 2),
   xgb_tuning_parameters = parameters, n_threads = 1)

# print best found model parameters
print(results$parameters)

# print model predictions
print(results$predictions)

# print model performance
print(results$performance, width = Inf)

## End(Not run)