Skip to contents

MLSampling Overview

Source: vignettes/ml-sampling-overview.Rmd
ml-sampling-overview.Rmd

Introduction

MLSampling provides an end-to-end interface for optimizing soil sampling campaigns using advanced Machine Learning models including Bayesian Deep Learning (BDL), Random Forest (RF), Unified Deep Learning (UDL), and Unified Feature Network (UFN) models. The package bundles spatial data validation, benchmarking, and rich reporting so agronomic teams can iterate quickly on sampling strategies while maintaining reproducibility and auditability.

This vignette introduces the core components, how they interact, and the main entry points you will use in analysis workflows.

Package structure

The package is organized around an R6 class MLSampling that orchestrates configuration, validation, optimization, and reporting. Supporting modules in the R/ folder expose dedicated services for configuration management, spatial data validation, benchmarking, and data models.

Key collaborators include:

  • BayesianDeepLearning (BDL) module for uncertainty quantification.
  • RandomForestOptimization (RF) module for feature importance analysis.
  • MLEnsembleManager for combining multiple models.
  • VisualizationService and ReportingService for generating insights.
  • ConfigManager (from config-management.R) for layered configuration with validation.

Getting started

To begin, load the package and create an instance of the tool.

library(MLSampling)

# Initialize with default configuration
tool <- create_ml_sampling_tool()

Optimization workflows

Four primary optimization methods are available:

  1. BDL (Bayesian Deep Learning): Uses Monte Carlo Dropout to estimate uncertainty and guide sampling to areas of high epistemic uncertainty.
  2. RF (Random Forest): Uses feature importance and spatial autocorrelation to select locations that maximize information gain.
  3. UDL (Unified Deep Learning): Combines neural and heuristic techniques (Legacy support).
  4. UFN (Unified Feature Network): Balances neural models with statistical fallbacks (Legacy support).
# Run BDL optimization
bdl_result <- tool$run_bdl(
  field_data = field_data,
  existing_samples = existing_samples,
  n_new_samples = 30
)

# Run RF optimization
rf_result <- tool$run_rf_optimization(
  field_data = field_data,
  existing_samples = existing_samples,
  n_new_samples = 30
)

Comparing algorithms

The tool can orchestrate multi-algorithm comparisons, running repeated experiments and evaluating results under configurable confidence thresholds.

comparison <- tool$compare_models(
  field_data = field_data,
  existing_samples = existing_samples,
  n_new_samples = 50,
  algorithms = c("BDL", "RF", "udl"),
  n_iterations = 5
)

# Generate report
tool$generate_ml_report(comparison)

Next steps

  • See the Quickstart Workflow vignette for a runnable example.
  • Explore Advanced ML Optimization for detailed uncertainty and ensemble workflows.