Implement RF-based backward feature selection via MDA for sits models

[ctotti]

Describe the new API function requested
While sits_reduce() handles temporal combinations, it would be valuable to also include a model-driven feature selection method that:

1. Complements existing reduction by selecting optimal bands/features for specific classification tasks
2. Uses Random Forest's variable importance (e.g., Mean Decrease in Accuracy and Gini importance) to:
   - Rank features by predictive importance
   - Iteratively remove least important features (backward elimination)
3. Preserves sits workflow by returning a modified sits tibble

Associated sits API function

sits_feature_selection(
  samples,               # sits tibble (time series)
  bands = NULL,          # Bands to evaluate (NULL = all)
  importance_metric = "mda",  # or "gini"
  n_iter = 20,           # Max iterations
  accuracy_loss = 0.02,  # Allowed accuracy drop (2%)
  rf_params = list(),    # Custom RF params (num_trees, etc)
  multicores = 1         # Parallel processing
)

The returned tibble can then be used to generate a reduced sampled cube, continuing the standard pipeline with sits_train(), sits_classify(), etc.

Additional context
This function would be especially useful for reducing high-dimensional inputs from multiple indices and texture metrics, improving performance and minimizing overfitting. Focusing on band-level optimization, complementing the temporal reduction approach of sits_reduce().

Note: I’m a user and not familiar with the internal feasibility of this integration.

References
[1] Mahmood et al. (2025) - Demonstrated 80% feature reduction with <2% accuracy loss

[M3nin0]

Hi @ctotti,

Thank you very much for your contribution and the detailed explanation. We will look into integrating this method as soon as possible and will provide you with feedback once we have more information.

Thanks,
Felipe Carlos