Assessing the application of random forest (RF) to predict water-table (WT) in selected Irish peatlands

Abstract
Peatlands are important global terrestrial carbon (C) sink. Most of Irish peatlands have been 
influenced in past by anthropogenic management, primarily through drainage for forestry, 
agriculture, or energy and horticultural extraction. Given the recent Irish peatland restoration 
activities, it is essential to deepen our understanding of the key drivers of peatland C-dynamics 
and to improve methodologies for reporting and verifying terrestrial CO₂ removals/emissions 
from drained and restored peatlands. The dependency of CO₂ fluxes on water-table (WT) levels 
in peatland ecosystems, under different land-use (LU), has been recognised in existing literature 
[1], indicating on the importance of accounting for WT variable in predictive models. This study 
focuses on assessing the application of random forest (RF) to predict WT in total eight Irish 
peatland sites under different LU (natural, rewetted, forest, grassland), which were monitored - 
i.e. low-level Irish blanket-bog sites from Co. Mayo and raised-bog sites from Co. Offaly [2]. The 
RF was chosen due to its ability to effectively manage mixed-data (numerical and categorical) and 
to provide robust predictions without the need for extensive data-preprocessing. Used were the 
data from ca. 2017 to 2020 on-site measurements [2], as well as the selected geospatial data 
derived from E-OBS daily grided-meteorological dataset [4]. The RF was applied to a number of 
numerical and categorical variables, by splitting the data into training- and testing-datasets. 
Hyperparameter tuning was done using ‘caret’ R-package [5]. Model evaluation (using 
performance metrics) was conducted on WT-predictions from testing-dataset. While findings 
from this study on selected eight Irish peatland sites indicate a relatively good potential of RF to 
predict WT (R² = 0.78), the work highlights the importance of assessing the ‘variable importance’ 
to reduce the number of variables in the model for practical applicability purposes, as well as to 
include more sites.

Acknowledgements
The authors are grateful to the Irish Environmental Protection Agency (EPA) for funding projects 
CO2PEAT (2022-CE-1100) and AUGER (2015-CCRP-MS.30) [EPA Research Programmes 2021-
2030 and 2014–2020], and to University of Limerick funding.

References
[1] Tiemeyer, B., et al., 2020. A new methodology for organic soils in national greenhouse gas inventories: Data synthesis, derivation and application,
Ecological Indicators, Vol. 109, 105838,  https://doi.org/10.1016/j.ecolind.2019.105838.
[2] Renou-Wilson, F., et. al, 2022. Peatland Properties Influencing Greenhouse Gas  Emissions and Removal (AUGER Project) (2015-CCRP-MS.30), EPA Research Report, Irish Environmental Protection Agency (EPA) https://www.epa.ie/publications/research/climate-change/Research_Report_401.pdf.
[3] Premrov, A., et.al, 2023. Insights into the CO2PEAT project: Improving methodologies for reporting and verifying terrestrial CO₂ removals and emissions from Irish peatlands. IGRM2023, Belfast, UK. https://www.researchgate.net/publication/369061601_Insights_into_the_CO2PEAT_project_Im
proving_methodologies_for_reporting_and_verifying_terrestrial_CO2_removals_and_emissions
_from_Irish_peatlands.
[4] Copernicus Climate Change Service, Climate Data Store, (2020): E-OBS daily gridded meteorological data for Europe from 1950 to present derived from in-situ observations. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). https://doi.org/10.24381/cds.151d3ec6.
[5] Kuhn, M. 2008. Building Predictive Models in R Using the caret Package. Journal of Statistical Software, 28(5), 1–26. https://doi.org/10.18637/jss.v028.i05.