Full title: Representative sample size for estimating saturated hydraulic conductivity via machine learning: A proof-of-concept study
Authors: Amin Ahmadisharaf, Reza Nematirad, Sadra Sabouri, Yakov Pachepsky, Behzad Ghanbarian Venue: Water Resources Research (AGU), 2024
Summary
Hydrology adopted machine learning widely but rarely examined how data heterogeneity and sample size affect accuracy. This proof-of-concept study asks the question directly for one target: saturated hydraulic conductivity (Ks).
The study draws on roughly 18,000 soil samples from the USKSAT database and measures how training size changes estimation accuracy. XGBoost models are trained over repeated random subsets of the database, so accuracy can be plotted as a function of how much data the model was given rather than reported at a single dataset size.
The result is a negative one, and that is the point: the learning and validation curves never flattened out, even at ~18,000 samples — accuracy had not stopped improving with added data, so the field’s working assumption about what counts as “enough data” for Ks estimation does not hold even at that scale.
Links
Related
ReaxFF Parameter Set for Boron Clusters and Icosahedral Boron Crystals — same first author, Amin Ahmadisharaf, and an adjacent computational-science collaboration.