Knowledge-guided modelling of soil available water capacity
Researchers: Dr Kanthu Joseph Mhango; Megan J. Lewis; Joe M. Roberts; W. Edwin Harris
Available water capacity is a deceptively compact target produced by interacting texture, organic constituents and soil structure. This project asks whether machine learning can capture those nonlinear interactions while retaining parameter behaviour that soil scientists can inspect and challenge.
Scientific challenge
Pedotransfer functions are useful because direct hydraulic measurement is expensive, but conventional regression can miss context-dependent relationships. Flexible ensembles often improve prediction, yet variable importance alone does not guarantee compliance with established soil physics. Soil fractions also form compositional data, creating collinearity and interpretation problems that can make apparently stable coefficients misleading.
Research architecture
The project develops a hybrid ensemble in which empirical linear estimates provide structured initial information and neural components learn residual nonlinear behaviour. Randomised feature partitions and repeated submodels expose the model to different covariate combinations rather than allowing one convenient subset to dominate. Because the complete architecture is differentiable, empirical priors and nonlinear refinements can be optimised within one trainable system.
Compositional transformations and feature engineering are used to respect the geometry of soil fractions. Evaluation separates predictive performance from scientific behaviour: observed soil datasets test external usefulness, while controlled synthetic systems test whether known relationships can be recovered when confounding and interaction are introduced deliberately. The aim is to understand when interpretability is genuine and when it is an artefact of data structure.
Intended contribution
This work contributes to a broader lab programme in auditable environmental AI—models whose flexibility is retained, but whose learned relationships can be evaluated against domain expectations rather than accepted on accuracy alone. Paper-level architectures, tuned settings and emerging results remain undisclosed while the manuscript is in preparation.