Machado (2026) edaphos: Disruptive Algorithms for Digital Soil Mapping

Identification

• Journal: Open MIND
• Year: 2026
• Date: 2026-04-21
• Authors: Hugo Rodrigues Machado
• DOI: 10.5281/zenodo.19683709

Research Groups

• Hugo Rodrigues Machado (Independent Researcher/Author)

Short Summary

The edaphos R package implements a suite of frontier algorithms for Digital Soil Mapping (DSM) designed to move the field beyond traditional regression-tree models. It integrates Causal AI, physics-informed machine learning, and foundation models to improve soil property prediction and mapping.

Objective

• To provide a research-oriented software framework that implements disruptive algorithms for Digital Soil Mapping, specifically targeting the limitations of current state-of-the-art regression-tree approaches.

Study Configuration

• Spatial Scale: Regional (demonstrated via a bundled Cerrado covariate-and-response cube).
• Temporal Scale: 4D (incorporating temporal dimensions through ConvLSTM and Neural ODEs).

Methodology and Data

• Models used:
  • Causal AI: DAG-backed backdoor adjustment.
  • Physics-Informed ML: Parametric pedogenetic ODEs, differentiable Neural ODEs, and covariate-conditioned hierarchical Neural ODEs.
  • 4D Pedometry: Multi-layer stacked ConvLSTM with multi-step rollout and mass-balance physics loss.
  • Foundation Models: SimCLR-style contrastive pre-training.
  • Active Learning: Hybrid conditioned Latin Hypercube Sampling (cLHS) and Quantile Regression Forest (QRF) policy with physics-informed rejection gates.
• Data sources: Bundled reproducible Cerrado covariate-and-response cube and unlabelled raster patches for contrastive learning.

Main Results

• Development of a functional R package featuring six research pillars (Causal AI, Physics-Informed ML, 4D Pedometry, Foundation Models, Autonomous Active Learning, and a Quantum ML roadmap).
• Delivery of 109 unit and integration tests and six mathematically-derived vignettes with LaTeX formulations.
• Implementation of a cross-platform CI pipeline (macOS, Windows, Ubuntu).

Contributions

• Shifts the DSM paradigm from standard regression-tree models toward a hybrid approach combining deep learning, causal inference, and physical laws (ODE/mass-balance).
• Provides an open-source, reproducible toolkit for the pedometry community to implement complex "frontier" algorithms without building them from scratch.

Funding

• Not specified.

Citation

@article{Machado2026edaphos,
  author = {Machado, Hugo Rodrigues},
  title = {edaphos: Disruptive Algorithms for Digital Soil Mapping},
  journal = {Open MIND},
  year = {2026},
  doi = {10.5281/zenodo.19683709},
  url = {https://doi.org/10.5281/zenodo.19683709}
}