Neural Networks

Uses CNN/deep models to learn spatial patterns from gridded covariates, improving soil predictions with confidence layers.

This project applies deep learning (especially CNNs) to spatial prediction of soil properties from stacked gridded covariates such as terrain derivatives and remote-sensing layers. CNNs learn multiscale spatial patterns and neighborhood context directly from these layers, helping capture complex soil–landscape relationships that may be missed by hand-crafted features. Models are trained with regularization and evaluated with spatially structured validation to ensure realistic generalization to new areas. Deliverables include high-resolution soil prediction maps and, where feasible, confidence/uncertainty information to communicate reliability.

Main focus

  • Learn spatial patterns from gridded covariates using CNN-based models.
  • Improve mapping performance for complex, nonlinear soil relationships.
  • Provide reliable outputs through spatial validation and confidence layers.

Objectives

  1. Build covariate “image stacks” that preserve spatial context for deep learning.
  2. Train CNN/deep models with spatial validation to test real-world generalization.
  3. Produce soil maps (and confidence layers) that support applied soil decisions.

Graphical abstract

:::: {.row .justify-content-sm-center} ::: {.col-sm-10 .mt-3 .mt-md-0}

::: ::::

::: caption Graphical abstract summarizing the workflow and key results. :::

Papers (selected)

Paper 1 — Multi-Task CNN Soil Texture Mapping

DOI




Paper 2 — Bio-Inspired ANN Hybrid for Soil Texture Fraction Mapping

DOI




Paper 3 — Uncertainty Quantification in ANN-Based Mapping of Soils

DOI



References