Pooled accuracy does not transfer: validating kinetic and machine-learning models of wetland ammonium removal for design and compliance

Models of nitrogen removal guide the design and consenting of treatment wetlands, and machine-learning studies often report R 2 above 0.9. These usually come from random splits of pooled records, which test interpolation, not forecasting or transfer to new systems. At one treatment plant, we compared a first-order kinetic model and a Random Forest, given identical inputs, for effluent ammonium from three partially saturated vertical-flow wetland configurations (zeolite, gravel or woodchip upper layer; triplicate units; 16 dates in 2023). As dose volume rose from 2.75 to 11 mm and in-bed temperature fell from about 22 to 10°C, gravel and woodchip effluent ammonium rose 13- to 16-fold. In pooled random splits the Random Forest explained 63% of the variance when told each record's system, 16% when not. Forecasting later dates, the kinetic model with a dose term erred least of four model variants in every system, clearly for gravel (1.14 against 3.24 mg L −1 for the Random Forest), extrapolating to an unseen dose. Carrying the last observation forward matched or beat the Random Forest everywhere, and the kinetic model for woodchip. For a system left out of training, the kinetic model erred least of the four for gravel and woodchip; no model predicted the zeolite system better than a constant. Without dose, the fitted temperature coefficient (gravel and woodchip) rose from 0.99–1.02 to 1.07–1.08, crediting cold with an effect that came with the dose increase. Validation should match the decision a model supports, and operating variables belong in wetland models.

Authors

Institutions

Publication Details

Journal
Journal of Environmental Management
Published
2026-11-17
DOI
https://doi.org/10.1016/j.jenvman.2026.131392
Primary Topic
Constructed Wetlands for Wastewater Treatment
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pooled accuracy does not transfer: validating kinetic and machine-learning models of wetland ammonium removal for design and compliance

Rupert J. Craggs, Reza Moghaddam, Chris C. Tanner, Sukhjit Singh
Journal of Environmental Management
Constructed Wetlands for Wastewater Treatment
article

Pooled accuracy does not transfer: validating kinetic and machine-learning models of wetland ammonium removal for design and compliance

Rupert J. Craggs, Reza Moghaddam, Chris C. Tanner, Sukhjit Singh
article en

Abstract

Models of nitrogen removal guide the design and consenting of treatment wetlands, and machine-learning studies often report R 2 above 0.9. These usually come from random splits of pooled records, which test interpolation, not forecasting or transfer to new systems. At one treatment plant, we compared a first-order kinetic model and a Random Forest, given identical inputs, for effluent ammonium from three partially saturated vertical-flow wetland configurations (zeolite, gravel or woodchip upper layer; triplicate units; 16 dates in 2023). As dose volume rose from 2.75 to 11 mm and in-bed temperature fell from about 22 to 10°C, gravel and woodchip effluent ammonium rose 13- to 16-fold. In pooled random splits the Random Forest explained 63% of the variance when told each record's system, 16% when not. Forecasting later dates, the kinetic model with a dose term erred least of four model variants in every system, clearly for gravel (1.14 against 3.24 mg L −1 for the Random Forest), extrapolating to an unseen dose. Carrying the last observation forward matched or beat the Random Forest everywhere, and the kinetic model for woodchip. For a system left out of training, the kinetic model erred least of the four for gravel and woodchip; no model predicted the zeolite system better than a constant. Without dose, the fitted temperature coefficient (gravel and woodchip) rose from 0.99–1.02 to 1.07–1.08, crediting cold with an effect that came with the dose increase. Validation should match the decision a model supports, and operating variables belong in wetland models.

Journal of Environmental ManagementVol. 420
University of Waikato (NZ)
Ministry of Business, Innovation and Employment
Clean water and sanitation
Constructed Wetlands for Wastewater Treatment
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon