Research article Modelling and mapping of subsurface nitrate-attenuation index in agricultural landscapes Stephen B. Collins * , Ranvir Singh , Stuart R. Mead , David J. Horne School of Agriculture and Environment, Massey University, Private Bag 11 222, Palmerston North, 4442, New Zealand A R T I C L E I N F O Handling editor: Lixiao Zhang Keywords: Agriculture Water quality Nitrate attenuation Groundwater redox conditions Machine learning Extreme gradient boosting A B S T R A C T Environmental management of nutrient losses from agricultural lands is required to reduce their potential im- pacts on the quality of groundwater and eutrophication of surface waters in agricultural landscapes. However, accurate accounting and management of nitrogen losses relies on a robust modelling of nitrogen leaching and its potential attenuation – specifically, the reduction of nitrate to gaseous forms of nitrogen – in subsurface flow pathways. Subsurface denitrification is a key process in potential nitrate attenuation, but the spatial and tem- poral dynamics of where and when it occurs remain poorly understood, especially at catchment-scale. In this paper, a novel Landscape Subsurface Nitrate-Attenuation Index (LSNAI) is developed to map spatially variable subsurface nitrate attenuation potential of diverse landscape units across the Manawatū-Whanganui region of New Zealand. A large data set of groundwater quality across New Zealand was collated and analysed to assess spatial and temporal variability of groundwater redox status (based on dissolved oxygen, nitrate and dissolved manganese) across different hydrogeological settings. The Extreme Gradient Boosting algorithm was used to predict landscape unit subsurface redox status by integrating the nationwide groundwater redox status data set with various landscape characteristics. Applying the hierarchical clustering analysis and unsupervised classifi- cation techniques, the LSNAI was then developed to identify and map five landscape subsurface nitrate atten- uation classes, varying from very low to very high potential, based on the predicted groundwater redox status probabilities and identified soil drainage and rock type as key influencing landscape characteristics. Accuracy of the LSNAI mapping was further investigated and validated using a set of independent observations of ground- water quality and redox assessments in shallow groundwaters in the study area. This highlights the potential for further research in up-scaling mapping and modelling of landscape subsurface nitrate attenuation index to accurately account for spatial variability in subsurface nitrate attenuation potential in modelling and assessment of water quality management measures at catchment-scale in agricultural landscapes. 1. Introduction Around the world, the need to provide food and fibre to growing populations has led to the intensification of agricultural land (Alexander et al., 2015; Fróna et al., 2019; Godfray and Garnett, 2014; Tscharntke et al., 2012). This has subsequently led to an increase in the leaching and runoff of agriculturally sourced nutrients to receiving waters, causing eutrophication and related environmental and human-health issues (Keatley et al., 2011; Savage et al., 2010; Withers et al., 2014). Inorganic nitrogen is a readily available nutrient for algae/periphyton growth in streams and rivers, posing potential toxicity effects to aquatic life, and degrading drinking water quality supplies posing risks to human health. Nitrate is generally the main form of inorganic nitrogen leached to groundwaters and in streams and rivers (Dymond et al., 2013; Mcdowell et al., 2009). In New Zealand, the National Policy Statement for Fresh- water Management has been developed to reduce nutrient losses and meet water quality objectives on a regional basis (Ministry for the Environment, 2020). As a response, various good management practices and land use scenarios need to be modelled to help assess where and how nitrate losses from agricultural lands are to be reduced or managed at catchment-scale. However, due to varying rates of potential nitrate attenuation ‘reduction’ in subsurface environments, it is difficult to accurately model these scenarios (Elwan, 2018; Singh et al., 2017a; Snelder et al., 2020). To overcome these challenges, new knowledge and tools to map and model spatially variable subsurface nitrate attenuation capacities at landscape unit scale are required for the development of * Corresponding author. E-mail address: s.collins1@massey.ac.nz (S.B. Collins). Contents lists available at ScienceDirect Journal of Environmental Management journal homepage: www.elsevier.com/locate/jenvman https://doi.org/10.1016/j.jenvman.2025.125628 Received 1 October 2024; Received in revised form 13 March 2025; Accepted 29 April 2025 Journal of Environmental Management 384 (2025) 125628 Available online 5 May 2025 0301-4797/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). https://orcid.org/0000-0001-5064-9332 https://orcid.org/0000-0001-5064-9332 https://orcid.org/0000-0003-3135-439X https://orcid.org/0000-0003-3135-439X https://orcid.org/0000-0001-8550-7457 https://orcid.org/0000-0001-8550-7457 mailto:s.collins1@massey.ac.nz www.sciencedirect.com/science/journal/03014797 https://www.elsevier.com/locate/jenvman https://doi.org/10.1016/j.jenvman.2025.125628 https://doi.org/10.1016/j.jenvman.2025.125628 http://crossmark.crossref.org/dialog/?doi=10.1016/j.jenvman.2025.125628&domain=pdf http://creativecommons.org/licenses/by/4.0/ targeted and effective mitigation strategies for nitrate losses from agri- cultural lands. Nitrate’s transport and fate in the subsurface is strongly influenced by landscape characteristics such as soil type, underlying geology and groundwater chemistry (Elwan, 2018; Jahangir et al., 2012; McAleer et al., 2017; Orr et al., 2016; Rivas et al., 2017). In this study, these landscape characteristics are referred to as landscape factors, which can be mapped to assess their potential influence on nitrate attenuation in the subsurface environment. Denitrification, where nitrate is trans- formed to gaseous forms of nitrogen under favourable conditions (Korom, 1992; Rivett et al., 2008), is usually the primary pathway for nitrate attenuation in subsurface flow pathways (Clague et al., 2015a, 2019; Collins et al., 2017; Jahangir et al., 2013a; Rivas et al., 2017, 2020). However, field-based measurements of subsurface denitrification are complicated, conducted infrequently, and limited to only a few re- gions with a narrow range of soil and rock type combinations (Burbery, 2018). This lack of relevant data makes it difficult to correlate subsur- face denitrification potential with the natural variability of landscape characteristics, posing a practical challenge for predicting and mapping landscape subsurface nitrate attenuation potential. To gain the spatial reach that subsurface denitrification data lacks, groundwater redox conditions have been used as a proxy for nitrate attenuation potential because of observations of subsurface dentifica- tion in reduced groundwater conditions (Clague et al., 2019; Collins et al., 2017; Rivas et al., 2017, 2020; Stenger et al., 2018). With the availability of extensive data sets including those for groundwater quality, soil, and geology, variability of groundwater redox conditions have been mapped in New Zealand and elsewhere (Close et al., 2016; Knoll et al., 2020; Koch et al., 2019, 2024; Rosecrans et al., 2017; Sarris et al., 2019; Tesoriero et al., 2015, 2024; Wilson et al., 2018, 2020). While the international research effort has been focused on identifying the depth of the subsurface redox interface (Hansen et al., 2014; Koch et al., 2019; Tesoriero et al., 2015), the focus in New Zealand has been on assessment of spatial variability of groundwater redox states and the depth at which they occur (Close et al., 2016; Wilson et al., 2018, 2020). In particular, efforts have been focused on reducing selection bias in the groundwater data (Wilson et al., 2020) and reducing uncertainty in assessment of groundwater redox status across different landscapes (Sarris et al., 2024). However, a robust management of water quality measures require tools to estimate and reduce nitrate fluxes in different flow pathways from agricultural lands at catchment-scale. While methods have been developed and refined to estimate catchment-scale nitrogen loads (Fraser, 2021; Snelder et al., 2017, 2020), the gap that currently exists is a robust assessment and effective accounting of subsurface nitrate attenuation in New Zealand catchments (Singh et al., 2017b; Snelder et al., 2020). Accounting of subsurface nitrate attenuation can inform both farm- and catchment-scale land use decisions and implementation of targeted water quality measures. However, further research is required to develop practical tools for a robust assessment and ac- counting of subsurface nitrate attenuation potential at catchment-scale that can accurately represent estimation and reduction of nitrate loads to streams and rivers in agricultural landscapes. This paper aimed to review and further develop a Landscape Sub- surface Nitrate-Attenuation Index (LSNAI) to map subsurface nitrate attenuation potential of diverse hydrogeological settings across the Manawatū-Whanganui region, located in the lower North Island of New Zealand. It defines LSNAI and presents a novel methodology for ana- lysing and integrating landscape hydrogeological variability, ground- water redox assessments, and on-farm shallow groundwater observations to predict and validate a regional-scale LSNAI map of different landscape units across the study area. Whereas previous studies predicting New Zealand groundwater redox conditions have employed linear discriminant analysis (Close et al., 2016; Wilson et al., 2018), or tested a range of machine learning algorithms (Friedel et al., 2020), this study evaluates and applies Extreme Gradient Boosting (XGBoost) to predict groundwater redox status (Chen and Guestrin, 2016). XGBoost is a flexible algorithm which is increasingly being used in hydrological studies (Niazkar et al., 2024; Ransom et al., 2022). This study develops a novel modelling approach, integrating applications of machine learning (XGBoost) and unsupervised classification techniques, to classify and map spatially variable subsurface nitrate attenuation potential of different landscape units. A robust approach is taken to validating pre- dictions by using field-scale measurements of shallow groundwater denitrification and redox data to independently verify landscape unit subsurface redox classifications. Finally, the landscape unit subsurface redox predictions and key landscape factors learned from the model training process are used to classify and map landscape units into groups of relatively high to low subsurface nitrate attenuation potential classes. The study provides new insights into the spatial variability of landscape factors’ potential influences on subsurface nitrate attenuation and its transport pathways from agricultural lands to receiving waters in diverse hydrogeological landscapes. It generates new information and knowledge to further develop accurate catchment-scale water quality models to robustly quantify potential impacts of various land use sce- narios and management practices for targeted and effective water quality measures across agricultural landscapes. 2. Data and methods 2.1. Study area This study focused as a case study on the Manawatū-Whanganui region of New Zealand, also known as the Horizons region (Fig. 1). The region covers 22,215 km2 (2,221,500 ha) and is characterised by several large river catchments that discharge into the Tasman Sea on the west coast (Fig. 1a). Local geology ranges from Mesozoic-age basement greywacke exposed in the southern part of the region; an abundance of unconsolidated Plio-Pleistocene-age deposits throughout; and volcanic deposits in the north (Begg et al., 2005) (Fig. 1b). The Manawatū catchment is divided by two mountain ranges, known as the Ruahine and Tararua ranges, with the river flowing east to west through a narrow gorge, known as the Manawatū Gorge (Fig. 1a and b). These two mountain ranges are formed by greywacke bedrock, shown in Fig. 1b, with Ruahine in the north and Tararua in the south. Groundwater resources are limited to two distinct basins that have developed either side of the Ruahine and Tararua ranges (Fig. 1c). On the eastern side, unconsolidated gravel and sand within the main river valleys, which rest on Tertiary-age mudstone, forms the main ground- water resource across Tararua valleys in the Manawatū catchment (Zarour, 2008). The groundwater resources on the western side are part of the Whanganui Basin, a deep sedimentary basin characterised by marine and alluvial deposits (Carter and Naish, 1998). Groundwater recharge and discharge dynamics also differ between these basins. On the eastern side, rainfall recharge infiltrates quickly into the shallow aquifer and discharges into the local rivers making for relatively short groundwater mean residence times (Morgenstern et al., 2017). Groundwater recharge is thought to reach much deeper depths on the western side, with groundwater discharging at the coastal margin into shallower groundwater and springs due to an upwards vertical hydraulic gradient (Thomas, 2017). Groundwater mean residence times are also significantly longer on the western side (Morgenstern et al., 2017). However, in most cases, groundwater abstraction is associated with Late Quaternary deposits (Zarour, 2008). Land use ranges from native forest/conservation land and low in- tensity farming in steep and hill country areas, to intensive agriculture such as dairy farming and cropping on the river plains and coastal areas (Fig. 1d). Soils in the region are primarily considered well-drained (Newsome et al., 2008), though drainage characteristics are highly dependent on a soil’s landscape position (Cowie et al., 1967; Cowie and Rijkse, 1977). A unique feature of the region is the sand dune formations on the region’s west coast that have developed over the late Holocene. A S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 2 major component of these dune formations is the titanomagnetite-rich ironsands that have been deposited along the coast by onshore winds, the source of which is erosion of andesitic volcanic rocks of nearby Taranaki volcanoes, located north-west of the Horizons region (Brathwaite et al., 2017). 2.2. Groundwater redox assessments and landscape characteristics Collins et al. (2025) collated and analysed a comprehensive groundwater data set collected over the 30-year period from 1990 to 2019 across 15 regions of New Zealand. These data were collected pri- marily from domestic/drinking water wells, farm supply and irrigation wells, or monitoring wells. The groundwater samples were sourced from a range of depths, with most groundwater samples taken from depths of 50 m or less (Collins et al., 2025). The groundwater quality data were classified into groundwater redox status as per the groundwater redox classification by Close et al. (2016), which was modified from McMahon and Chapelle (2008) for New Zealand conditions. It classifies ground- water redox status based on the thresholds of dissolved oxygen (DO) < 1.0 mg/L; nitrate-nitrogen (NO3 − -N) < 0.5 mg/L; and dissolved manga- nese (Mn2+) > 0.05 mg/L (Close et al., 2016; Collins et al., 2025). The collated groundwater dataset included some missing Mn2+ values (an important redox variable), which were imputed using locally-developed random forest regression models between observations of Mn2+ and Ca2+, Mg2+, K+, Na+ and Cl− values in the groundwater dataset (Collins et al., 2025). Across the study area, shallow groundwater was classified predominantly oxidised, whereas the reduced groundwater conditions were assessed more frequently at greater depths (Collins et al., 2025). Refer to Collins et al. (2025) for further details on the assessment of the collated groundwater quality dataset and its application for assessing the spatial variability and temporal stability of groundwater redox conditions across New Zealand regions. The collated national-scale Fig. 1. Maps of a) the main river catchments, b) geological types (GNS Science, 2012), c) hydrogeological units (White et al., 2019, Revised 2023) and d) land uses (Herzig et al., 2020) across the Manawatū-Whanganui region of New Zealand, also known as the Horizons region. S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 3 groundwater data set resulted in 34,098 individual groundwater redox assessments from 3764 wells across New Zealand, of which 1242 groundwater redox assessments were located across the Horizons region from 227 wells. Table 1 summarises the basic statistics associated with the classification of the national-scale groundwater data set into either an oxidised, mixed or reduced category (Collins et al., 2025). Oxidised groundwater is the majority category and typically has higher mean and median DO and NO3 − -N concentrations, while reduced groundwater has higher mean and median dissolved Mn2+ and Fe2+ concentrations. This pattern conforms to the general understanding of subsurface redox conditions (Rivett et al., 2008). Within the Horizons region, groundwater redox data was sourced from a depth of 2 m–235 m (Table 2). Like the national data set, these data were collected from groundwater wells primarily used for domes- tic/drinking water, farm or industrial use. The groundwater data were collected from 1990, though particularly from 2000 when the frequency and coverage of groundwater monitoring increased in the study region. Oxidised groundwater was, on average, assessed at shallower depths, compared to more mixed or reduced groundwater assessed at relatively deeper groundwater depths (Table 2). A geographical dataset of landscape attributes and landscape unit boundaries was obtained from the New Zealand Land Resource In- ventory (LRI; Newsome et al., 2008). The LRI is an inventory of five physical factors, soil, rock type, slope, presence and severity of erosion, and vegetation, mapped at a scale of 1:50,000, that drive land use capability mapping in New Zealand (Lynn et al., 2009). In addition, the LRI is supplemented with 16 additional soil properties collectively known as the Fundamental Soil Layers (FSL; Barringer et al., 1998; Newsome et al., 2008). Together, these geographical data sets provide around 100,000 polygons (landscape units) across New Zealand featuring a range of landscape attributes. In this study, a total of fourteen attributes were selected from the LRI and the FSL to represent landscape characteristics, potentially related to redox processes in the subsurface environment (Collins et al., 2025). These include top rock, base rock, slope, carbon content, gravel content, soil drainage, permeability, profile available water (PAW), pH, cation exchange capacity (CEC), phosphorus retention, potential rooting depth, flood class, and macroporosity. Each groundwater observation was spatially linked to the landscape polygon (i.e. landscape unit) in which the well was located, and the values of the corresponding landscape attributes were assigned accordingly. A complete list of these variables and their categories can be found in the supplementary data file. In the LRI (Newsome et al., 2008), the top rock and base rock are derivations of rock type mapped across the landscape. Top rock is the first-named entire rock type, irrespective of any underlying stratigraphy and can have several lithology expressions depending on the dominance of the rock types in a single polygon unit. Base rock is the single value that identifies the principal basement lithology (Newsome et al., 2008). In this study, these have been merged as ‘top rock over base rock’ and reclassified as ‘Rock type’. Where top rock and base rock are the same, just the single attribute is described (e.g. ‘Greywacke over Greywacke’ becomes ‘Greywacke’) and where they are distinct, both attributes have been retained (e.g. ‘Loess over Alluvium’). In addition, for both top rock and base rock types, the various ash deposits have been reclassified as ‘volcanic ash’. Other kinds of volcanic deposits that are not ash have been reclassified as ‘volcanic deposits’. This helped reduce and simplify the number of variables for rock type. Slope, which has eight classes, refers to a physiographic area of relatively homogeneous average relief. It is derived from stereo aerial photograph interpretation and field verification and allows up to two slope expressions within a single polygon (Newsome et al., 2008). In this study, just the first/dominant named slope attribute class has been retained (e.g. ‘Moderately steep + Steep’ becomes ‘Moderately steep’). Permeability refers to the rate at which water moves through saturated soil and includes three classes, ‘Slow’, ‘Medium’ and ‘Rapid’, of which multiple combinations are possible in layered soils (Clayden and Webb, 1994). In this study, just the first named permeability attribute class has been retained (e.g. ‘Moderate/Rapid’ becomes ‘Moderate’). The remaining attributes of carbon content; gravel; soil drainage; PAW; pH; CEC; phosphorus retention; potential rooting depth; flood class; and macroporosity have been retained as they are described in Newsome et al. (2008). In all cases, unrelated categories within these variables e. g., lakes, estuaries, towns/cities, have been classified as ‘Not specified’. New Zealand’s groundwater redox data is unevenly distributed among some landscape variables (Collins et al., 2025). For example, groundwater redox data was reasonably well distributed among the different categories of carbon class, soil drainage class and permeability class. However, it was somewhat more unevenly distributed among some rock type categories. Alluvium, gravels and windblown sands were over-represented in the groundwater redox data set (relative to their New Zealand distribution), while sandstone and mudstone were under-represented. This tends to reflect the hydrogeological conditions where groundwater can be accessed, primarily those areas characterised by alluvium, gravels and windblown sands (Zarour, 2008). Table 1 A summary of New Zealand groundwater quality data as per their redox status. Redox Status n Parameter Mean Median Min Max Standard deviation Oxidised 23,504 DO (mg/L) 6.58 6.90 1.00 11.85 2.42 NO3 − -N (mg/L) 4.90 3.80 <0.01 25.00 4.47 Mn2+ (mg/L) 0.01 <0.01 <0.01 0.05 0.01 Fe2+ (mg/L) 0.05 0.01 <0.01 8.20 0.18 SO4 2− (mg/L) 12.13 9.20 0.01 100.00 11.00 Mixed 4171 DO (mg/L) 2.27 1.15 0.01 11.64 2.50 NO3 − -N (mg/L) 1.79 0.20 <0.01 25.00 3.49 Mn2+ (mg/L) 0.26 0.10 <0.01 1.70 0.35 Fe2+ (mg/L) 1.01 0.11 <0.01 15.00 2.20 SO4 2− (mg/L) 14.09 8.00 0.01 100.00 18.73 Reduced 6423 DO (mg/L) 0.21 0.15 0.01 0.99 0.20 NO3 − -N (mg/L) 0.04 0.01 <0.01 0.50 0.08 Mn2+ (mg/L) 0.44 0.27 0.05 1.70 0.41 Fe2+ (mg/L) 2.53 1.20 <0.01 15.00 3.16 SO4 2− (mg/L) 7.53 2.60 0.01 99.90 12.96 Table 2 A summary of depth of groundwater wells, in metres below ground level, from which groundwater redox data was sourced from within the Horizons region. p25 and p75 refer to the 25th and 75th percentile. Redox Status n Min (m) p25 (m) Median (m) Mean (m) p75 (m) Max (m) Oxidised 548 2.0 10.5 17.0 19.3 22.3 135 Mixed 134 2.0 10.6 27.0 45.9 68.6 134 Reduced 560 2.6 18.0 30.3 49.5 68.6 235 S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 4 2.3. Prediction of landscape unit subsurface redox status The collated groundwater redox status and landscape factors were modelled to predict and map the subsurface nitrate attenuation potential of landscape units across the study area. The model predictions do not consider any potential for nitrate attenuation in the soil profile, unsat- urated zone, riparian or in-stream environments of the landscape. However, the groundwater redox status is used as a proxy of the land- scape subsurface redox status and nitrate attenuation potential, as observed by field studies in New Zealand landscapes and elsewhere (Clague et al., 2019; Collins et al., 2017; Jahangir et al., 2013b; Kruisdijk et al., 2022; Rivas et al., 2017, 2020; Steiness et al., 2021; Yang et al., 2023). The groundwater redox classification by Close et al. (2016) included a ‘mixed’ status, however in this study only the probability of a land- scape unit being classified as either ‘oxidised’ or ‘reduced’ in its sub- surface environment was predicted for classification and mapping. The mixed groundwater redox status is characterised by relatively high values across all the redox-related hydrochemistry variables (Table 1), making its implication somewhat unclear for assessment of landscape subsurface nitrate attenuation potential of a landscape unit. The collated national-scale groundwater redox data set (Collins et al., 2025) and associated landscape attributes (Newsome et al., 2008; Supplementary Data) were used to train and test an XGBoost model that predicted a landscape unit’s subsurface redox status probability as either oxidised or reduced. The XGBoost model training data set consisted of point-level groundwater redox status observations distributed across New Zealand (classified as oxidised or reduced) (Collins et al., 2025), which were used as the response variable alongside landscape-level predictors such as rock type, soil drainage and carbon classes. While only areas with groundwater redox observations available were included in the training data set, the trained model was applied to predict the subsurface redox status of all ~11,000 landscape units within the Horizons region, including those without direct groundwater redox data. This approach uses the relationships learned by the XGBoost model in the training datasets to generate robust predictions of sub- surface redox status in areas lacking direct groundwater observations. The XGBoost algorithm was used to predict landscape unit redox status probabilities for the Horizons region, i.e. P(Oxidised) and P (Reduced). The XGBoost algorithm is an ensemble tree-based machine learning method that is widely used for supervised learning tasks such as regression and classification (Niazkar et al., 2024). It is based on gradient boosting, whereby subsequent trees are trained to correct the errors of their predecessors. However, XGBoost incorporates several additional features such as regularisation and tree pruning to mitigate overfitting (Niazkar et al., 2024). XGBoost has been used in a number of groundwater and water quality studies, from predicting river nitrate concentrations (Dorado-Guerra et al., 2022; Xie et al., 2024), ground- water nitrate and manganese concentrations (Bedi et al., 2020; Collins et al., 2025; DeSimone and Ransom, 2021; Ransom et al., 2022), to assessing the occurrence of edge of field runoff (Hu et al., 2021). Although boosted regression trees have been used to predict ground- water redox conditions previously (DeSimone et al., 2020), using XGBoost for such a task is a recent development (Koch et al., 2024). The national data set were first partitioned into training and testing data sets (75/25 split) with groundwater redox status stratified to ensure an even spread between the two data sets. A 10-fold cross-validation data set was created from the training data set for the model’s hyper- parameter tuning. Several data pre-processing steps were necessary prior to training. First, infrequent categories were collapsed for the Rock type variable, as there were dozens of individual attributes within this variable. Attributes that featured less than 0.5 % of the total number of landscape units were collapsed to ‘other’. The remaining landscape variables (e.g. Soil drainage class, CEC class) had relatively few indi- vidual categories that did not require collapsing (See Supplementary Data). Second, as XGBoost requires the creation of dummy variables (Kuhn and Silge, 2022), all categorical variables were converted into numerical data with one-hot encoding. Finally, predictor variables with zero variance were removed and the now-numeric data was also centred and scaled. The model’s performance was evaluated using sensitivity, speci- ficity, accuracy, kappa (Viera and Garrett, 2005) and area under the receiver operating characteristic curve (AUC-ROC; Fawcett, 2006). Sensitivity (true positive rate) measures the proportion of cases that were correctly identified as an oxidised land unit by the model. Speci- ficity (true negative rate) measures the proportion of cases that were correctly identified as a reduced land unit by the model. Accuracy measures the overall correctness of predictions made by the model across both oxidised and reduced land unit classes, while kappa accounts for chance agreement, offering a more robust metric for class imbalance (Viera and Garrett, 2005). The AUC-ROC evaluates the performance of a binary classification model across different threshold values. The ROC curve assesses sensitivity against 1-specificity at various threshold values. A higher value indicates better model performance. The hyperparameters of the XGBoost model were also tuned to optimise its performance. These are settings applied to the model but are not learned during the training process (Niazkar et al., 2024). To achieve a tuned model, Latin hypercube sampling, featuring a range of 30 possible hyperparameter configurations, was used. The cross-validation folds were used to evaluate the performance of different configurations generated by the Latin hypercube. Mean AUC-ROC (across the 10 cross-validation folds) was used to select the optimal combination of hyperparameters for the model. These were then used to train the model. All data analysis was carried out in R (R Core Team, 2022; v. 4.2.2) with predictive analysis undertaken with Tidymodels (Kuhn and Wick- ham, 2020; v. 1.1.1). 2.3.1. Ground-truthing landscape subsurface redox predictions A number of local studies have been recently undertaken in the Horizons region to better understand groundwater redox conditions and the potential for denitrification in the subsurface environment (Collins et al., 2017; Collins, 2015; Gonzalez Moreno, 2019; Rivas et al., 2014, 2017, 2020; Smith et al., 2017). A range of field measurement methods were employed in these studies, including the installation of shallow groundwater piezometers; sampling and analysis of shallow ground- water quality; and single-well push-pull tests. The push-pull test is an in-situ method for a site characterisation (Istok, 2013; Istok et al., 1997) and identifying subsurface biogeochemical processes including denitri- fication (Addy et al., 2002; Anderson et al., 2014; Jahangir et al., 2013a; Rivas et al., 2014). The observations and results of these existing local studies in the Horizons region were collated and used to provide inde- pendent validation of the XGBoost model’s predictions of landscape unit subsurface redox status across the study area. 2.4. Development of the landscape subsurface nitrate-attenuation index The LSNAI was developed by classifying the Horizons region’s ~11,000 landscape units into five distinct clusters using hierarchical clustering (Everitt et al., 2011). A landscape unit was defined as a spatial area, or unit of territory, that shares similar physical and geographic characteristics (Campos-Campos et al., 2018). The classification was achieved by integrating the XGBoost model’s landscape subsurface redox status predictions and associated key influencing landscape characteristics of rock type and soil drainage. Rock type and soil drainage were chosen due to their significant roles observed in nitrate attenuation in the subsurface environment (Elwan, 2018; Jahangir et al., 2012; McAleer et al., 2017; Orr et al., 2016; Rivas et al., 2017) and their high variable importance scores in the XGBoost model. A hierarchical (agglomerative) clustering analysis was applied for its interpretability and ability to provide meaningful clusters of similar landscape units. The number of LSNAI clusters was determined to ach- ieve a distinct range of mean subsurface redox status probabilities (P S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 5 (Oxidised) and P(Reduced)). This classification approach resulted in hi- erarchical clusters ranging from Very low nitrate attenuation potential, characterised by a high P(Oxidised) and low P(Reduced) redox status, to Very high nitrate attenuation potential, characterised by a high P (Reduced) and low P(Oxidised) redox status. These classes represent the relative potential of subsurface nitrate attenuation assessed for different landscape units in the Horizons region, forming the basis of the devel- opment of the LSNAI for classification and mapping of landscape units and their subsurface nitrate attenuation potential in agricultural landscapes. 3. Results and discussion 3.1. Prediction of landscape unit subsurface redox status The pre-processing steps outlined above resulted in a total of 73 features to train the XGBoost model. The 10 cross-validation folds were used to optimise the model’s hyperparameters, with a range of values tested during the tuning process (Table 3). The combination of values that yielded the highest mean AUC-ROC result across the 10 cross- validation folds was selected to train the final model. Model performance was evaluated using the held-out testing data set, which was not used during model training. The testing data set results showed a sensitivity (the true positive rate for oxidised conditions) of 0.96 and a specificity (the true negative rate for reduced conditions) of 0.80. This implies the model is more effective at identifying oxidised conditions than reduced. The relatively higher sensitivity compared to specificity is likely a result of class imbalance in the data set, where oxidised conditions were more prevalent. The overall rate of accuracy on the testing data was 0.93, with a kappa statistic of 0.78 and an AUC-ROC of 0.97. The lower kappa value, compared to the higher accuracy, is also consistent with class imbalance. As the model was trained on the national data set, but only applied to the Horizons region, these metrics have also been calculated for training and testing data within the study area (Table 4). The model’s perfor- mance for training and testing data within the study area was similar for the model trained using the national data set (Table 4), except for a slightly higher specificity and kappa, demonstrating a slight improve- ment on the model’s ability to predict reduced conditions in the study region, compared to nationally. The trained XGBoost model was used to predict the subsurface redox status probability for each landscape unit in the Horizons region (Fig. 2). The results demonstrate widely variable landscape subsurface redox conditions throughout the region, with some areas likely to be more strongly oxidising and other areas likely to be more strongly reducing. While no subsurface redox assessment has been completed specifically for the Horizons region, Wilson et al. (2020) produced a national-scale groundwater redox assessment for New Zealand, exempting the moun- tainous and steep areas of the country. Wilson et al.’s (2020) assessment predicted groundwater (subsurface) redox conditions at various depths (15 m, 50 m and 80 m) with a vertical redox gradient apparent in some areas over this range. There are broad similarities between the pre- dictions in this study (Fig. 2) and those of Wilson et al. (2020), but specific subsurface redox depth was not examined here, due to this study’s focus on predicting the probability of a landscape unit’s sub- surface redox status. The results of this study (Fig. 2) and those of Wilson et al. (2020), both indicate a greater likelihood of oxidising subsurface conditions in the northern part of the region and reducing subsurface conditions in the south-west coastal area. The northern part of the region features volcanic deposits in addition to various sedimentary marine deposits such as sandstone and mudstone (Fig. 1b). Significantly older basement greywacke also outcrops in the north-east of the region (Fig. 1b). In the south, the western coast is characterised by alluvial and marine sedimentary deposits including sands and gravel. Iron-rich windblown sands form a thin veneer across this area with sand dunes aligned with the predominant north-west wind direction. The axial ranges, dividing the south-east of the region with the south-west, were unmodelled by Wilson et al. (2020), but were predicted as being strongly oxidising landscapes in this analysis (Figs. 2 and 1). Fig. 3 plots the variable importance scores calculated from the trained XGBoost model. The most important variables in assessing landscape subsurface redox conditions were scored as poorly drained soils and alluvium rock type. In this context, soil drainage characteristics are based on the depth and duration of water tables inferred from soil colours and mottles; while alluvium refers to undifferentiated floodplain alluvium, colluvium and glacial drift (Newsome et al., 2008). Soil drainage and sediment type have previously been associated with denitrification and reducing conditions in groundwater (Clague et al., 2015b, 2019; Collins et al., 2017; Rivas et al., 2017; Wu et al., 2018). Wu et al. (2018) found that a low permeability silty clay layer within a Beijing aquifer attenuated nitrate at about twice the rate of an overlying medium grain sand layer. In New Zealand, Clague et al. (2019) observed higher denitrification potential under imperfectly/poorly drained soils compared to well-drained soils in the Reporoa Basin, Waikato; while Rivas et al. (2020) noted that sites underlain by alluvium had greater subsurface denitrification potential than those with loess over gravel. Similarly, Collins et al. (2017) observed denitrification occurring in shallow groundwaters at three sites characterised by poorly-drained Table 3 Hyperparameter values used to tune the XGBoost model, and the optimised values used to train the model. Hyperparameter name (Tidymodels alias) Descriptiona Hypercube range Optimal value Learning rate (learn_rate) The rate at which the boosting algorithm adapts from iteration to iteration. 1.35e-10 – 7.15e-02 0.07 Number of rounds (trees) The number of trees contained in the ensemble. 7–1796 296 Maximum tree depth (tree_depth) The maximum depth of the tree i.e. number of splits. 1–15 7 Minimum child weight (min_n) The minimum number of data points in a node that is required for the node to be split further. 2–40 4 Gamma (loss_reduction) The reduction in the loss function required to split further. 2.37e-10 – 18.17 5.74e-10 Subsample (sample_size) The proportion of data that is exposed to the fitting routine. 0.13–1.00 0.74 Column sample by node (mtry) The proportion of predictors that will be randomly sampled at each split when creating the tree models. 0.12–0.98 0.34 Early stopping (stop_iter) The number of iterations without improvement before stopping. 3–20 18 a Kuhn & Vaughan (n.d.) Boosted trees. Retrieved August 8, 2024, from htt ps://parsnip.tidymodels.org/reference/boost_tree.html. Table 4 The XGBoost model performance metrics calculated for the training and testing data for prediction of landscape subsurface (groundwater) redox status. The number of data points in each testing and training data set are in parentheses. Metric National data set Horizons data set Training (21,029) Testing (7011) Training (815) Testing (285) Sensitivity 0.97 0.96 0.96 0.98 Specificity 0.80 0.80 0.88 0.82 Accuracy 0.94 0.93 0.92 0.90 Kappa 0.79 0.78 0.84 0.81 ROC-AUC 0.97 0.97 0.96 0.96 S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 6 https://parsnip.tidymodels.org/reference/boost_tree.html https://parsnip.tidymodels.org/reference/boost_tree.html coastal soil types. 3.1.1. Ground-truthing of landscape subsurface redox predictions Ground-truthing the results of the XGBoost model involved comparing the predicted landscape subsurface redox status, as P(Oxi- dised) or P(Reduced), with observations of shallow groundwater redox status established though field techniques from previous studies within the Horizons region (Fig. 4 and Table 5). The redox status at each of these sites has been established either through shallow groundwater monitoring (Gonzalez Moreno, 2019; Smith et al., 2017) using McMa- hon and Chapelle’s (2008) classification (incorporating DO, NO3 − -N, Mn2+, Fe3+ and SO4 2− ), and/or through the single-well push-pull tests by observing the concentration of NO3 − -N against a conservative tracer (Collins et al., 2017; Gonzalez Moreno, 2019; Rivas et al., 2020). At each Fig. 2. Predicted landscape subsurface redox status determined by the XGBoost model: a) probability that landscape units are oxidised; b) probability that landscape unts are reduced. A higher probability value indicates a greater level of confidence in the prediction. Fig. 3. Top ten variable importance scores calculated from the trained XGBoost model to predict landscape redox status in the Horizons region. Fig. 4. Location of ground-truthing studies in the southern Horizons region showing both soil drainage class and rock type categories. Site locations do not necessarily reflect the exact locations of places they are named for. S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 7 location, the individual piezometers were generally within several me- tres of each other, though the Pahiatua sites were up to 1 km apart. No data from these studies were used in the training of the XGBoost model. The Santoft and Bulls sites (Table 5) were both located in the Ran- git̄ıkei catchment’s sand country on a sand dune complex classified geologically as windblown sand. Soils here are described in associations, where the landscape is made up of repeating phases of dune ridge, sand plain and peaty swamp (Cowie et al., 1967). Soil drainage characteristics can vary depending on the soil type and its position in the repeating phases. The studies identified in Table 5 used a combination of shallow groundwater monitoring and various push-pull tests carried out in shallow piezometers both of which suggest widespread subsurface (groundwater) reducing conditions (Collins et al., 2017; Gonzalez Moreno, 2019; Smith et al., 2017). The area is known to have strongly reducing groundwater conditions, possibly due to it being a regional discharge zone for older, more hydrochemically evolved groundwater (Morgenstern et al., 2017; Thomas, 2017), and groundwater redox conditions influenced by the high levels of iron in the sand dune deposits of volcanic origin (Brathwaite et al., 2017). The Sanson site was located further inland than the sand country, with surficial geology charac- terised by alluvium over gravels. The soil, Ohakea silt loam, is imperfectly-to-poorly-drained at the Sanson site. Across these three lo- cations (Santoft, Bulls and Sanson), the local studies have determined the shallow groundwater to be largely reduced (Collins et al., 2017; Gonzalez Moreno, 2019; Smith et al., 2017). This validates the XGBoost model’s prediction of a strong likelihood of reducing landscape sub- surface (groundwater) conditions in this area (Table 5). The Woodville and Pahiatua sites, which were located on the eastern side of the Manawatū catchment (Fig. 4), are more closely associated with alluvial and river deposits. The Woodville site was underlain by alluvium with soils described as silt loam and clay loam. At all three depths (5m, 6m, 7.5m below ground level) reducing groundwater con- ditions were measured in these field studies (Table 5), which validates Table 5 Shallow groundwater redox and denitrification studies, compared with the XGBoost predictions of landscape unit subsurface redox status in the Horizons region. Reference Piezo depth (m, below ground level) Redox status P (Oxidised) P (Reduced) Santoft sites Collins et al. (2017) 3.00 Reduced 0.05 0.95 Smith et al. (2017) 2.40 Reduced 0.02 0.98 Smith et al. (2017) 5.00 Reduced 0.02 0.98 Smith et al. (2017) 3.40 Reduced 0.01 0.99 Smith et al. (2017) 5.00 Reduced 0.01 0.99 Gonzalez Moreno (2019) 3.00 Reduced 0.01 0.99 Smith et al. (2017) 3.00 Reduced 0.01 0.99 Gonzalez Moreno (2019) 6.00 Reduced 0.01 0.99 Smith et al. (2017) 6.00 Reduced 0.01 0.99 Gonzalez Moreno (2019) 6.00 Mixed 0.01 0.99 Smith et al. (2017) 3.00 Reduced 0.01 0.99 Smith et al. (2017) 3.50 Reduced 0.01 0.99 Bulls sites Collins et al. (2017) 3.00 Reduced 0.05 0.95 Collins et al. (2017) 6.00 Reduced 0.05 0.95 Sanson sites Collins et al. (2017) 3.00 Reduced 0.01 0.99 Collins et al. (2017) 6.00 Reduced 0.01 0.99 Woodville sites Gonzalez Moreno (2019) 5.00 Reduced 0.37 0.63 Rivas et al. (2020) 5.00 Reduced 0.37 0.63 Gonzalez Moreno (2019) 6.00 Reduced 0.37 0.63 Rivas et al. (2020) 6.00 Reduced 0.37 0.63 Gonzalez Moreno (2019) 7.50 Reduced 0.37 0.63 Rivas et al. (2020) 7.50 Reduced 0.37 0.63 Pahiatua sites Gonzalez Moreno (2019) 3.60 Oxidised 0.70 0.30 Gonzalez Moreno (2019) 4.51 Oxidised 0.85 0.15 Gonzalez Moreno (2019) 6.40 Oxidised 0.85 0.15 Gonzalez Moreno (2019) 4.40 Oxidised 0.97 0.03 Rivas et al. (2020) 4.40 Oxidised 0.97 0.03 Table 5 (continued ) Reference Piezo depth (m, below ground level) Redox status P (Oxidised) P (Reduced) Gonzalez Moreno (2019) 5.40 Oxidised 0.97 0.03 Rivas et al. (2020) 5.40 Oxidised 0.97 0.03 Gonzalez Moreno (2019) 6.40 Oxidised 0.97 0.03 Rivas et al. (2020) 6.40 Oxidised 0.97 0.03 Palmerston North sites Gonzalez Moreno (2019) 5.50 Oxidised 0.20 0.80 Rivas et al. (2020) 7.50 Reduced 0.20 0.80 Gonzalez Moreno (2019) 7.50 Reduced 0.20 0.80 Gonzalez Moreno (2019) 8.70 Reduced 0.20 0.80 Dannevirke sites Gonzalez Moreno (2019) 4.50 Mixed 0.70 0.30 Rivas et al. (2020) 4.50 Oxidised 0.70 0.30 Gonzalez Moreno (2019) 7.50 Reduced 0.70 0.30 Rivas et al. (2020) 7.50 Reduced 0.70 0.30 S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 8 the XGBoost model’s prediction of 0.63 probability suggesting high likelihood of reducing subsurface conditions at the Woodville site. At Pahiatua sites, some piezometers were installed in alluvium, while others were installed in gravels with soils described at each site as a stony silt loam. Local studies here identified oxidised groundwater conditions at the various depths tested (3.6m–6.4m below ground level). Likewise, the XGBoost model also predicted a high likelihood of Pahiatua sites being oxidised (0.70–0.97) in their subsurface environment. The remaining two sites, at Palmerston North and Dannevirke (also in the Manawatū catchment), demonstrates the challenge of predicting landscape subsurface redox status where, potentially, a redox interface occurs at shallow depths (Table 5). In both cases, the shallowest pie- zometers were measured as oxidised (or mixed) groundwaters, and slightly deeper piezometers measured as reduced groundwaters (Table 5). However, the XGBoost model predicted Palmerston North as a higher probability (0.80) of reduced subsurface redox conditions (i.e. below the redox interface), while at the Dannevirke sites it predicted a greater probability (0.70) of oxidised subsurface redox (i.e. above the redox interface). While a redox interface has been observed in various subsurface environments (Böhlke et al., 2002; Hansen et al., 2014; Koch et al., 2019, 2024; Kolbe et al., 2019; Refsgaard et al., 2014), it may not be a widespread occurrence throughout the Horizons region with reducing conditions found at very shallow depths in places (Table 5). However, a high degree of correspondence between the field mea- surements and the predicted oxidised or reduced probabilities (Table 5) demonstrates independent validation of the trained XGBoost model’s performance in modelling and mapping of subsurface redox status probabilities of different landscape units across the Horizons region. 3.2. Mapping of landscape subsurface nitrate-attenuation index A total of five clusters were selected for the LSNAI based on their interpretability (Table 6), with each group exhibiting distinct mean values of P(Oxidised) and P(Reduced) across different land units. These clusters were then categorised into LSNAI classes ranging from ‘Very low’ to ‘Very high’, resulting in distinct distributions of soil drainage class and rock type classes (Fig. 5), as well as a spatially distributed range of LSNAI categories across the major catchments of the Horizons region (Fig. 6). The ‘Very low’ LSNAI landscape class is composed of 737 land units, with a mean P(Oxidised) value of 0.89, i.e. highly oxidised subsurface redox conditions with very low nitrate attenuation potential, covering about 1802 km2 of the study area. This class is characterised entirely by greywacke rock type (100 %), but with a range of soil drainage classes, primarily well drained (50 %) and moderately well drained (43 %) (Fig. 5). The ‘Low’ LSNAI category, which is composed of 6033 land units, has a mean P(Oxidised) value of 0.78. There is a wider range of rock types but the soil drainage class is almost entirely well drained (99 %). Rock types primarily include mudstone (13 %) and sandstone (23 %) but also includes a range of volcanic deposits (Fig. 5). This category is the largest LSNAI class mapped in the region (Table 6). The ‘Medium’ LSNAI class is comprised of a balance between P (Oxidised) and P(Reduced) redox status (Table 6). Rock types in this class comprise a greater share of mudstone (25 %) and sandstone (40 %) with soil drainage class split between moderately well drained (74 %) and imperfectly drained (26 %) (Fig. 5). The ‘High’ LSNAI class, comprising 988 landscape units, has a mean P(Oxidised) value of 0.31 and a mean P(Reduced) value of 0.69 (Table 6). These landscape units are primarily composed of poorly drained soils (79 %), with loess (31 %) and alluvium (41 %) making up the dominant rock types (Fig. 5), which covers about 1841 km2 of the study area. Loess over gravels and loess over sandstone also make up minor portions of the ‘High’ LSNAI class. Finally, the ‘Very high’ LSNAI class features a range of soil drainage classes but is exclusively composed of windblown sands (100 %) (Fig. 5), located along the western coast of the region (Figs. 6 and 1). It has a mean P(Oxidised) of only 0.13, or mean P(Reduced) of 0.87, covering 774 km2 of the study area. While a few other assessments of nitrogen attenuation have been undertaken in the Horizons region (Elwan et al., 2015; Singh and Horne, 2019; Snelder et al., 2020), only a qualitative comparison can be made between their estimation of nitrogen attenuation rates and the current study because of the different methodologies and the scales at which nitrogen attenuation has been resolved, from landscape units (Singh and Horne, 2019) to sub-catchments (Elwan et al., 2015; Snelder et al., 2020). Elwan et al. (2015) quantified the soluble inorganic nitrogen (SIN) attenuation factors ranging from, 0.29 to 0.75, based a mass bal- ance approach using the differences between estimates of average annual land use nitrogen export coefficients (kg ha− 1 yr− 1) and calcu- lated instream SIN yields (kg ha− 1 yr− 1) across Tararua sub-catchments in the Manawatū catchment. Snelder et al. (2020) also took a mass balance approach to calculating the total nitrogen (TN) attenuation coefficients (ranging from 0.1 to 0.9) for the Horizons region derived from the difference between estimates of land use based export nitrogen coefficients (kg ha− 1 yr− 1) and calculated instream TN yields (kg ha− 1 yr− 1) across different sub-catchments in the Horizons region. Singh and Horne (2019) developed a model for the Rangit̄ıkei catchment to ac- count for the influence of different hydrogeological settings on SIN attenuation capacity across its five sub-catchments. They defined land units based on a combination of rock and soil types and calibrated ni- trate attenuation coefficients to produce a landscape nitrate attenuation index (LNAI) of different land units across the Tararua and Rangit̄ıkei districts of the Horizons region (Singh and Horne, 2019). However, none of these methods/approaches (Elwan et al., 2015; Singh and Horne, 2019; Snelder et al., 2020) directly included information from ground- water redox conditions and landscape attributes together in their assessment of either sub-catchment scale nitrogen attenuation factor or landscape nitrogen attenuation index of different land units. While all three models for the region (Fig. 6; Singh and Horne, 2019; Snelder et al., 2020) suggest that the magnitude of subsurface nitrate attenuation is spatially variable, there are both similarities and differ- ences in the spatial distribution of their assessments of subsurface nitrate attenuation potential across the study area. For example, in the upper Rangit̄ıkei catchment, relatively higher potential of nitrogen attenuation have been calculated by Snelder et al. (2020) and Singh and Horne (2019), yet a relatively low LSNAI was assessed in this study, likely due to prediction of high P(Oxidised) results in this part of the catchment (Fig. 2). Wilson et al. (2020) also predicted oxidised groundwater con- ditions in the northern Rangit̄ıkei catchment, in contrast to both Snelder et al.’s (2020) and Singh and Horne’s (2019) prediction of relatively higher nitrogen attenuation potential for upper Rangit̄ıkei catchment landscapes. However, a better agreement between the various models Table 6 Classification of Landscape Subsurface Nitrate-Attenuation Index (LSNAI) developed for the Horizons region. Shown are the total number of landscape units within each cluster (n); the mean probability of oxidised and reduced predictions for each cluster; the standard deviation; and the total area for each LSNAI class. Landscape Subsurface Nitrate- Attenuation Index, LSNAI categories Landscape units (n) Mean P (Oxidised) Mean P (Reduced) Standard deviation Area (km2) Very low 737 0.89 0.11 0.14 1802 Low 6033 0.78 0.22 0.20 12,422 Medium 3330 0.56 0.44 0.26 6168 High 988 0.31 0.69 0.25 1841 Very high 531 0.13 0.87 0.20 774 S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 9 (Fig. 6; Singh and Horne, 2019; Snelder et al., 2020) was observed across the eastern Manawatū catchment (also known as the Tararua catch- ment), where the nitrogen attenuation potential assessed in this study mostly varied between very low and medium LSNAI (Fig. 6), corre- sponding very well with low to medium nitrogen attenuation co- efficients quantified by Snelder et al. (2020) and between low and high attenuation in Singh and Horne (2019). Also, in the Whanganui catchment, both this study and that of Snelder et al. (2020) identified relatively low potential of nitrogen attenuation across the entire catchment. The independent validation of XGBoost’s redox predictions provides a useful assessment of the model’s performance, however the validation studies used above, while assessing diverse soil and geological condi- tions, do not cover the full range of landscape types found in the Hori- zons region. An important source of uncertainty in this model is the somewhat narrow range of landscape characteristics that groundwater redox measurements have been drawn from. Groundwater redox data used here is particularly well represented for alluvium, loess over allu- vium, volcanic deposits, windblown sands and gravel, but is underrep- resented for other rock types such as greywacke, metamorphic rocks, sandstone and mudstone (Collins et al., 2025). In addition, groundwater (subsurface) redox data is widely available in locations of flat relief but lacking in steeper areas such as hill country landscapes of the Horizons region. Few studies to date report on the variability of denitrification potential in hill country subsoils and seepage wetlands in the Horizons region (Chibuike et al., 2019, 2024; Sanwar, 2023; Sanwar et al., 2022). However, the XGBoost model trained and tested here could pose an uncertainty in the prediction and mapping of landscape subsurface ni- trate attenuation potential in steeper hill country landscapes, which makes up more than 60 % of the Horizons region. Therefore, further research is required to better characterise and understand subsurface redox conditions in these steeper landscapes, particularly hill country landscape units. This will help reduce overall uncertainty in the further development and validation of the LSNAI to comprehensively model and map spatially variable subsurface nitrate attenuation potentiation of different landscape units across agricultural catchments. 4. Conclusion Meeting the water quality standards prescribed by New Zealand’s National Policy Statement for Freshwater Management requires modelling various land use scenarios and water quality measures to reduce nitrogen losses from farm-scale to catchment-scale. This study has developed a landscape subsurface nitrate-attenuation index (LSNAI) to provide an estimate of subsurface nitrate attenuation potential of Fig. 5. Distribution of soil drainage class and rock type groupings of each Landscape Subsurface Nitrate-Attenuation Index (LSNAI) class predicted and mapped in the Horizons region. Fig. 6. Mapping of the Landscape Subsurface Nitrate-Attenuation Index (LSNAI) classes based on the hierarchical cluster analysis of groundwater redox probability predictions by the XGBoost model and soil drainage and rock type classes for the Horizons region. S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 10 different landscape units for the Horizons region of New Zealand. The index integrates classification of groundwater redox status with key influencing landscape characteristics into the assessment and mapping of potential nitrate attenuation in the subsurface environment of different landscape units. Using a national groundwater redox data set, the machine-learning model XGBoost was trained and tested to predict landscape unit subsurface redox status for the region’s ~11,000 land- scape units. The model predicted a probability of oxidised or reduced subsurface (groundwater) redox conditions. XGBoost was found to be a good predictor of landscape subsurface redox status, which was also demonstrated by ground-truthing the model’s predictions against independently gathered data on shallow groundwater redox conditions and nitrate attenuation measurements at various field sites located in different landscapes in the region. By combining the prediction of landscape unit subsurface redox status, with the key influencing land- scape variables of soil drainage and rock type characteristics, a hierar- chical clustering analysis and unsupervised classification identified and mapped five landscape subsurface nitrate attenuation potential classes. These LSNAI attenuation classes, mapped from Very low to Very high nitrate attenuation potential, reflect the spatially variable nature of subsurface (groundwater) redox conditions in the Horizons region, potentially allowing for subsurface nitrate attenuation accounting to be realised at catchment-scale. The developed LSNAI broadly identifies the lowest subsurface nitrate attenuation potential in areas where well drained soils were underlain by gravels and greywacke rock types, with increasing potential of subsurface nitrate attenuation on the plains and at the western coast with reduced groundwater conditions. However, further research and testing are required to understand the effectiveness of applying the LSNAI for catchment-scale modelling of nitrogen losses to waterways, especially in steeper hilly landscapes where the subsur- face (groundwater) redox data is currently limited. However, an accu- rate mapping and accounting of LSNAI is expected to improve simulation of various land use scenarios across agricultural catchments, based on potential for nitrate attenuation, and provide options for tar- geted and effective measures on reducing nitrogen losses at catchment- scale. CRediT authorship contribution statement Stephen B. Collins: Writing – original draft, Visualization, Meth- odology, Formal analysis, Data curation, Conceptualization. Ranvir Singh: Writing – review & editing, Supervision, Methodology, Conceptualization. Stuart R. Mead: Validation, Methodology, Super- vision. David J. Horne: Writing – review & editing, Conceptualization, Supervision. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This study was conducted as part of a collaborative research pro- gramme between Massey University School of Agriculture and Envi- ronment (SAE) and Horizons Regional Council (HRC), New Zealand. HRC partly co-funded this study and provided in-kind support for groundwater data collected and analysed in this study. HRC co-funding and in-kind support is greatly appreciated. Other New Zealand regional and district councils also kindly pro- vided groundwater quality data for this study. We also thank the anonymous reviewers whose contribution greatly improved the manuscript. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. org/10.1016/j.jenvman.2025.125628. Data availability Data will be made available on request. References Addy, K., Kellogg, D.Q., Gold, A.J., Groffman, P.M., Ferendo, G., Sawyer, C., 2002. In situ push–pull method to determine ground water denitrification in riparian zones. J. Environ. Qual. 31 (3), 1017–1024. https://doi.org/10.2134/jeq2002.1017. Alexander, P., Rounsevell, M.D.A., Dislich, C., Dodson, J.R., Engström, K., Moran, D., 2015. Drivers for global agricultural land use change: the nexus of diet, population, yield and bioenergy. Glob. Environ. Change 35, 138–147. https://doi.org/10.1016/j. gloenvcha.2015.08.011. Anderson, T.R., Groffman, P.M., Kaushal, S.S., Walter, M.T., 2014. Shallow groundwater denitrification in riparian zones of a headwater agricultural landscape. J. Environ. Qual. 43 (2), 732–744. https://doi.org/10.2134/jeq2013.07.0303. Barringer, J., Wilde, H., Willoughby, J., Burgham, S., Hewitt, A., Gibb, R., Newsome, P.J. F., Rijkse, W.C., 1998. Restructuring the New Zealand Land Resource Inventory to meet the changing needs for spatial information in environmental research and management. In: 10th Colloquium of the Spatial Information Research Centre. University of Otago, New Zealand, pp. 25–33. https://www.researchgate.net/pu blication/228892268. (Accessed 21 January 2025). Bedi, S., Samal, A., Ray, C., Snow, D., 2020. Comparative evaluation of machine learning models for groundwater quality assessment. Environ. Monit. Assess. 192 (12), 776. https://doi.org/10.1007/s10661-020-08695-3. Begg, J.G., Palmer, A., Gyopari, M., 2005. Geological Synopsis of the Manawatu- Horowhenua Area for a Review of the Region’s Hydrogeology (No. 2005/EXT/687). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentra l.org.nz/records. (Accessed 21 January 2025). Böhlke, J.K., Wanty, R., Tuttle, M., Delin, G., Landon, M., 2002. Denitrification in the recharge area and discharge area of a transient agricultural nitrate plume in a glacial outwash sand aquifer, Minnesota. Water Resour. Res. 38 (7), 10–26. https://doi.org/ 10.1029/2001wr000663. Brathwaite, R.L., Gazley, M.F., Christie, A.B., 2017. Provenance of titanomagnetite in ironsands on the west coast of the North Island, New Zealand. J. Geochem. Explor. 178, 23–34. https://doi.org/10.1016/j.gexplo.2017.03.013. Burbery, L., 2018. Nitrate reactivity in groundwater: a brief review of the science, practical methods of assessment, and collation of results from New Zealand field investigations. J. Hydrol. N. Z. 57 (2), 41–70. https://www.hydrologynz.org.nz/j ournal-of-hydrology. (Accessed 21 January 2025). Campos-Campos, O., Cruz-Cárdenas, G., Aquino, R.J.C., Moncayo-Estrada, R., Machuca, M.A.V., Meléndez, L.A.Á., 2018. Historical delineation of landscape units using physical geographic characteristics and land use/cover Change. Open Geosci. 10 (1), 45–57. https://doi.org/10.1515/geo-2018-0004. Carter, R.M., Naish, T.R., 1998. A review of Wanganui Basin, New Zealand: global reference section for shallow marine, Plio–Pleistocene (2.5–0 Ma) cyclostratigraphy. Sediment. Geol. 122 (1–4), 37–52. https://doi.org/10.1016/S0037-0738(98)00097- 9. Chen, T., Guestrin, C., 2016. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, San Francisco California USA, pp. 785–794. https://doi.org/10.1145/ 2939672.2939785. Chibuike, G., Burkitt, L., Bretherton, M., Camps-Arbestain, M., Singh, R., Bishop, P., Hedley, C., Roudier, P., 2019. Dissolved organic carbon concentration and denitrification capacity of a hill country sub-catchment as affected by soil type and slope. N. Z. J. Agric. Res. 62 (3), 354–368. https://doi.org/10.1080/ 00288233.2018.1508041. Chibuike, G., Singh, R., Burkitt, L., 2024. The effects of pastoral hill country natural landscape features and land management practices on nitrate losses and its potential attenuation for improved water quality. J. Sustain. Agric. Environ. 3 (1), e12096. https://doi.org/10.1002/sae2.12096. Clague, J., Stenger, R., Moorhead, B., 2015a. Investigating how soil drainage class affects the redox status of shallow groundwater. In: New Zealand Hydrological Society Annual Conference - from Data to Knowledge. New Zealand Hydrological Society, Hamilton, New Zealand. Clague, J.C., Stenger, R., Clough, T.J., 2015b. Denitrification in the shallow groundwater system of a lowland catchment: a laboratory study. Catena 131, 109–118. https:// doi.org/10.1016/j.catena.2015.03.012. (Accessed 21 January 2025). Clague, J.C., Stenger, R., Morgenstern, U., 2019. The influence of unsaturated zone drainage status on denitrification and the redox succession in shallow groundwater. Sci. Total Environ. 660, 1232–1244. https://doi.org/10.1016/j. scitotenv.2018.12.383. Clayden, B., Webb, T.H., 1994. Criteria for Defining the Soil Form: the Fourth Category of the New Zealand Soil Classification. Landcare Research, New Zealand, Lincoln, New Zealand. https://doi.org/10.7931/DL1-LRSS-3-1994. Close, M.E., Abraham, P., Humphries, B., Lilburne, L., Cuthill, T., Wilson, S., 2016. Predicting groundwater redox status on a regional scale using linear discriminant S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 11 https://doi.org/10.1016/j.jenvman.2025.125628 https://doi.org/10.1016/j.jenvman.2025.125628 https://doi.org/10.2134/jeq2002.1017 https://doi.org/10.1016/j.gloenvcha.2015.08.011 https://doi.org/10.1016/j.gloenvcha.2015.08.011 https://doi.org/10.2134/jeq2013.07.0303 https://www.researchgate.net/publication/228892268 https://www.researchgate.net/publication/228892268 https://doi.org/10.1007/s10661-020-08695-3 https://archivescentral.org.nz/records https://archivescentral.org.nz/records https://doi.org/10.1029/2001wr000663 https://doi.org/10.1029/2001wr000663 https://doi.org/10.1016/j.gexplo.2017.03.013 https://www.hydrologynz.org.nz/journal-of-hydrology https://www.hydrologynz.org.nz/journal-of-hydrology https://doi.org/10.1515/geo-2018-0004 https://doi.org/10.1016/S0037-0738(98)00097-9 https://doi.org/10.1016/S0037-0738(98)00097-9 https://doi.org/10.1145/2939672.2939785 https://doi.org/10.1145/2939672.2939785 https://doi.org/10.1080/00288233.2018.1508041 https://doi.org/10.1080/00288233.2018.1508041 https://doi.org/10.1002/sae2.12096 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref15 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref15 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref15 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref15 https://doi.org/10.1016/j.catena.2015.03.012 https://doi.org/10.1016/j.catena.2015.03.012 https://doi.org/10.1016/j.scitotenv.2018.12.383 https://doi.org/10.1016/j.scitotenv.2018.12.383 https://doi.org/10.7931/DL1-LRSS-3-1994 analysis. J. Contam. Hydrol. 191, 19–32. https://doi.org/10.1016/j. jconhyd.2016.04.006. Collins, S., Singh, R., Rivas, A., Palmer, A., Horne, D., Manderson, A., Roygard, J., Matthews, A., 2017. Transport and potential attenuation of nitrogen in shallow groundwaters in the lower Rangitikei catchment, New Zealand. J. Contam. Hydrol. 206, 55–66. https://doi.org/10.1016/j.jconhyd.2017.10.002. Collins, S.B., 2015. Investigating the Transport and Fate of Nitrogen from Farms to River in the Lower Rangitikei Catchment (Master of Science). Massey University, Palmerston North, New Zealand. https://mro.massey.ac.nz. (Accessed 21 January 2025). Collins, S.B., Singh, R., Mead, S.R., Horne, D.J., Roygard, J.K.F., 2025. Assessment of spatial variability and temporal stability of groundwater redox conditions in New Zealand. Environ. Monit. Assess. 197 (1), 58. https://doi.org/10.1007/s10661-024- 13427-y. Cowie, J.D., Fitzgerald, P., Owers, W., 1967. Soils of the manawatu-rangitikei sand country. NZ Soil Bureau Bulletin 29. https://doi.org/10.7931/dl1-sbb-029. Cowie, J.D., Rijkse, W.C., 1977. Soils of the Manawatu County, North Island. Department of Scientific and Industrial Research, New Zealand. https://doi.org/10.7931/DL1- SSR-30-1. Wellington, New Zealand. DeSimone, L.A., Pope, J.P., Ransom, K.M., 2020. Machine-learning models to map pH and redox conditions in groundwater in a layered aquifer system, Northern Atlantic Coastal Plain, eastern USA. J. Hydrol. Reg. Stud. 30, 100697. https://doi.org/ 10.1016/j.ejrh.2020.100697. DeSimone, L.A., Ransom, K.M., 2021. Manganese in the Northern Atlantic Coastal Plain aquifer system, eastern USA—modeling regional occurrence with pH, redox, and machine learning. J. Hydrol. Reg. Stud. 37, 100925. https://doi.org/10.1016/j. ejrh.2021.100925. Dorado-Guerra, D.Y., Corzo-Pérez, G., Paredes-Arquiola, J., Pérez-Martín, M.Á., 2022. Machine learning models to predict nitrate concentration in a river basin. Environ. Res. Commun. 4 (12), 125012. https://doi.org/10.1088/2515-7620/acabb7. Dymond, J.R., Ausseil, A.G.E., Parfitt, R.L., Herzig, A., McDowell, R.W., 2013. Nitrate and phosphorus leaching in New Zealand: a national perspective. N. Z. J. Agric. Res. 56 (1), 49–59. https://doi.org/10.1080/00288233.2012.747185. Elwan, A., 2018. Accounting of Nitrogen Attenuation in Agricultural Catchments (Doctor of Philosophy). Massey University, Palmerston North, New Zealand. https://mro. massey.ac.nz. (Accessed 21 January 2025). Elwan, A., Singh, R., Horne, D., Roygard, J., Clothier, B., 2015. Nitrogen attenuation factor: can it tell a story about nutrients in different subsurface environments? In: Currie, L.D., Burkitt, L.L. (Eds.), Moving Farm Systems to Improved Nutrient Attenuation, Occasional Report No. 28. Fertilizer and Lime Research Centre. Massey University, Palmerston North, New Zealand, p. 12. http://flrc.massey.ac.nz/publicat ions.html. (Accessed 21 January 2025). Everitt, B.S., Landau, S., Leese, M., Stahl, D., 2011. Cluster Analysis, first ed. Wiley Series in Probability and Statistics. Wiley. https://doi.org/10.1002/9780470977811. Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recognit. Lett. 27 (8), 861–874. https://doi.org/10.1016/j.patrec.2005.10.010. Fraser, C., 2021. Load Calculations for Rivers of the Manawatū-Whanganui Region to 31 December 2019 (No. 2021/EXT/1748). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentral.org.nz/records. (Accessed 21 January 2025). Friedel, M.J., Wilson, S.R., Close, M.E., Buscema, M., Abraham, P., Banasiak, L., 2020. Comparison of four learning-based methods for predicting groundwater redox status. J. Hydrol. 580, 124200. https://doi.org/10.1016/j.jhydrol.2019.124200. Fróna, D., Szenderák, J., Harangi-Rákos, M., 2019. The challenge of feeding the world. Sustainability 11 (20), 5816. https://doi.org/10.3390/su11205816. GNS Science, 2012. Geological map of New Zealand [digital data]. https://doi.org/ 10.21420/QF82-7D42. Godfray, H.C.J., Garnett, T., 2014. Food security and sustainable intensification. Philos. Trans. R. Soc. B Biol. Sci. 369 (1639), 20120273. https://doi.org/10.1098/ rstb.2012.0273. Gonzalez Moreno, M.A., 2019. Assessment of Biogeochemical Transformation of Nitrate in Shallow Groundwater in the Agricultural Landscape (Master of Science). Massey University, Palmerston North, New Zealand. https://mro.massey.ac.nz. (Accessed 21 January 2025). Hansen, A.L., Christensen, B.S.B., Ernstsen, V., He, X., Refsgaard, J.C., 2014. A concept for estimating depth of the redox interface for catchment-scale nitrate modelling in a till area in Denmark. Hydrogeol. J. 22 (7), 1639–1655. https://doi.org/10.1007/ s10040-014-1152-y. Herzig, A., Manderson, A., Jolly, B., Barnes, M., Baish, L., 2020. Opportunities and Constraints on Intensive Land-Use Expansion in the Horizons Region (No. 2020/ EXT/1796). Horizons Regional Council, Palmerston North, New Zealand. https://a rchivescentral.org.nz/records. (Accessed 21 January 2025). Hu, Y., Fitzpatrick, L., Fry, L.M., Mason, L., Read, L.K., Goering, D.C., 2021. Edge-of-field runoff prediction by a hybrid modeling approach using causal inference. Environ. Res. Commun. 3 (7), 075003. https://doi.org/10.1088/2515-7620/ac0d0a. Istok, J.D., 2013. Push-pull tests for site characterization. Lecture Notes in Earth System Sciences. Springer Berlin Heidelberg, Berlin, Heidelberg. https://doi.org/10.1007/ 978-3-642-13920-8. Istok, J.D., Humphrey, M.D., Schroth, M.H., Hyman, M.R., O’Reilly, K.T., 1997. Single- well, “push-pull” test for in situ determination of microbial activities. Groundwater 35 (4), 619–631. https://doi.org/10.1111/j.1745-6584.1997.tb00127.x. Jahangir, M.M.R., Johnston, P., Addy, K., Khalil, M.I., Groffman, P.M., Richards, K.G., 2013a. Quantification of in situ denitrification rates in groundwater below an arable and a grassland system. Water. Air. Soil Pollut. 224 (9). https://doi.org/10.1007/ s11270-013-1693-z. Jahangir, M.M.R., Johnston, P., Barrett, M., Khalil, M.I., Groffman, P.M., Boeckx, P., Fenton, O., Murphy, J., Richards, K.G., 2013b. Denitrification and indirect N2O emissions in groundwater: hydrologic and biogeochemical influences. J. Contam. Hydrol. 152, 70–81. https://doi.org/10.1016/j.jconhyd.2013.06.007. Jahangir, M.M.R., Johnston, P., Khalil, M.I., Richards, K.G., 2012. Linking hydrogeochemistry to nitrate abundance in groundwater in agricultural settings in Ireland. J. Hydrol. 448–449, 212–222. https://doi.org/10.1016/j. jhydrol.2012.04.054. Keatley, B.E., Bennett, E.M., MacDonald, G.K., Taranu, Z.E., Gregory-Eaves, I., 2011. Land-use legacies are important determinants of lake eutrophication in the anthropocene. PLoS One 6 (1), e15913. https://doi.org/10.1371/journal. pone.0015913. Knoll, L., Breuer, L., Bach, M., 2020. Nation-wide estimation of groundwater redox conditions and nitrate concentrations through machine learning. Environ. Res. Lett. 15 (6), 064004. https://doi.org/10.1088/1748-9326/ab7d5c. Koch, J., Kim, H., Tirado-Conde, J., Hansen, B., Møller, I., Thorling, L., Troldborg, L., Voutchkova, D., Højberg, A.L., 2024. Modeling groundwater redox conditions at national scale through integration of sediment color and water chemistry in a machine learning framework. Sci. Total Environ. 947, 174533. https://doi.org/ 10.1016/j.scitotenv.2024.174533. Koch, J., Stisen, S., Refsgaard, J.C., Ernstsen, V., Jakobsen, P.R., Højberg, A.L., 2019. Modeling depth of the redox interface at high resolution at national scale using random forest and residual Gaussian simulation. Water Resour. Res. 55 (2), 1451–1469. https://doi.org/10.1029/2018WR023939. Kolbe, T., De Dreuzy, J.-R., Abbott, B.W., Aquilina, L., Babey, T., Green, C.T., Fleckenstein, J.H., Labasque, T., Laverman, A.M., Marçais, J., Peiffer, S., Thomas, Z., Pinay, G., 2019. Stratification of reactivity determines nitrate removal in groundwater. Proc. Natl. Acad. Sci. 116 (7), 2494–2499. https://doi.org/10.1073/ pnas.1816892116. Korom, S.F., 1992. Natural denitrification in the saturated zone: a review. Water Resour. Res. 28 (6), 1657–1668. https://doi.org/10.1029/92WR00252. Kruisdijk, E., Eisfeld, C., Stuyfzand, P.J., Van Breukelen, B.M., 2022. Denitrification kinetics during aquifer storage and recovery of drainage water from agricultural land. Sci. Total Environ. 849, 157791. https://doi.org/10.1016/j. scitotenv.2022.157791. Kuhn, M., Silge, J., 2022. Tidy Modeling with R: a Framework for Modeling in the Tidyverse, first ed. O’Reilly, Sebastopol, CA, USA. Kuhn, M., Vaughan, D., n.d. Boosted trees https://parsnip.tidymodels.org/reference/ boost_tree.html (accessed 8 August 2024). Kuhn, M., Wickham, H., 2020. Tidymodels: a collection of packages for modeling and machine learning using tidyverse principles. https://www.tidymodels.org. (Accessed 21 January 2025). Lynn, I.H., Manderson, A.K., Page, M.J., Harmsworth, G.R., Eyles, G.O., Douglas, G.B., Mackay, A.D., Newsome, P.J.F., 2009. Land Use Capability Survey Handbook - a New Zealand Handbook for the Classification of Land, third ed. AgResearch, Landcare Research, GNS Science, Hamilton, Lincoln, Lower Hutt. https://doi.org/ 10.7931/DL1MG6. McAleer, E.B., Coxon, C.E., Richards, K.G., Jahangir, M.M.R., Grant, J., Mellander, P.E., 2017. Groundwater nitrate reduction versus dissolved gas production: a tale of two catchments. Sci. Total Environ. 586, 372–389. https://doi.org/10.1016/j. scitotenv.2016.11.083. Mcdowell, R.W., Larned, S.T., Houlbrooke, D.J., 2009. Nitrogen and phosphorus in New Zealand streams and rivers: control and impact of eutrophication and the influence of land management. N. Z. J. Mar. Freshw. Res. 43 (4), 985–995. https://doi.org/ 10.1080/00288330909510055. McMahon, P.B., Chapelle, F.H., 2008. Redox processes and water quality of selected principal aquifer systems. Ground Water 46 (2), 259–271. https://doi.org/10.1111/ j.1745-6584.2007.00385.x. Ministry for the Environment, 2020. National Policy Statement for Freshwater Management. Ministry for the Environment, Wellington, New Zealand. https://envi ronment.govt.nz/publications/. (Accessed 21 January 2025). Morgenstern, U., Martindale, H., Stewart, M., Trompetter, V., van der Raaij, R., Toews, M., Townsend, D., 2017. Groundwater dynamics source & hydrochemical processes as inferred from Horizons regional age tracer data (No. 2017/EXT/1525). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentra l.org.nz/records. (Accessed 21 January 2025). Newsome, P.J.F., Wilde, R.H., Willoughby, E.J., 2008. Land resource information system spatial data layers. Data Dictionary. Landcare Research NZ Ltd. https://lris.scinfo. org.nz/document/9162-lris-data-dictionary-v3/. (Accessed 8 August 2024). Niazkar, M., Menapace, A., Brentan, B., Piraei, R., Jimenez, D., Dhawan, P., Righetti, M., 2024. Applications of XGBoost in water resources engineering: a systematic literature review (Dec 2018–May 2023). Environ. Model. Softw. 174, 105971. https://doi.org/10.1016/j.envsoft.2024.105971. Orr, A., Nitsche, J., Archbold, M., Deakin, J., Ofterdinger, U., Flynn, R., 2016. The influence of bedrock hydrogeology on catchment-scale nitrate fate and transport in fractured aquifers. Sci. Total Environ. 569–570, 1040–1052. https://doi.org/ 10.1016/j.scitotenv.2016.06.148. R Core Team, 2022. R: a language and environment for statistical computing. https:// www.r-project.org/. (Accessed 21 January 2025). Ransom, K.M., Nolan, B.T., Stackelberg, P.E., Belitz, K., Fram, M.S., 2022. Machine learning predictions of nitrate in groundwater used for drinking supply in the conterminous United States. Sci. Total Environ. 807, 151065. https://doi.org/ 10.1016/j.scitotenv.2021.151065. Refsgaard, J.C., Auken, E., Bamberg, C.A., Christensen, B.S.B., Clausen, T., Dalgaard, E., Effersø, F., Ernstsen, V., Gertz, F., Hansen, A.L., He, X., Jacobsen, B.H., Jensen, K.H., Jørgensen, F., Jørgensen, L.F., Koch, J., Nilsson, B., Petersen, C., De Schepper, G., S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 12 https://doi.org/10.1016/j.jconhyd.2016.04.006 https://doi.org/10.1016/j.jconhyd.2016.04.006 https://doi.org/10.1016/j.jconhyd.2017.10.002 https://mro.massey.ac.nz https://doi.org/10.1007/s10661-024-13427-y https://doi.org/10.1007/s10661-024-13427-y https://doi.org/10.7931/dl1-sbb-029 https://doi.org/10.7931/DL1-SSR-30-1 https://doi.org/10.7931/DL1-SSR-30-1 https://doi.org/10.1016/j.ejrh.2020.100697 https://doi.org/10.1016/j.ejrh.2020.100697 https://doi.org/10.1016/j.ejrh.2021.100925 https://doi.org/10.1016/j.ejrh.2021.100925 https://doi.org/10.1088/2515-7620/acabb7 https://doi.org/10.1080/00288233.2012.747185 https://mro.massey.ac.nz https://mro.massey.ac.nz http://flrc.massey.ac.nz/publications.html http://flrc.massey.ac.nz/publications.html https://doi.org/10.1002/9780470977811 https://doi.org/10.1016/j.patrec.2005.10.010 https://archivescentral.org.nz/records https://doi.org/10.1016/j.jhydrol.2019.124200 https://doi.org/10.3390/su11205816 https://doi.org/10.21420/QF82-7D42 https://doi.org/10.21420/QF82-7D42 https://doi.org/10.1098/rstb.2012.0273 https://doi.org/10.1098/rstb.2012.0273 https://mro.massey.ac.nz https://doi.org/10.1007/s10040-014-1152-y https://doi.org/10.1007/s10040-014-1152-y https://archivescentral.org.nz/records https://archivescentral.org.nz/records https://doi.org/10.1088/2515-7620/ac0d0a https://doi.org/10.1007/978-3-642-13920-8 https://doi.org/10.1007/978-3-642-13920-8 https://doi.org/10.1111/j.1745-6584.1997.tb00127.x https://doi.org/10.1007/s11270-013-1693-z https://doi.org/10.1007/s11270-013-1693-z https://doi.org/10.1016/j.jconhyd.2013.06.007 https://doi.org/10.1016/j.jhydrol.2012.04.054 https://doi.org/10.1016/j.jhydrol.2012.04.054 https://doi.org/10.1371/journal.pone.0015913 https://doi.org/10.1371/journal.pone.0015913 https://doi.org/10.1088/1748-9326/ab7d5c https://doi.org/10.1016/j.scitotenv.2024.174533 https://doi.org/10.1016/j.scitotenv.2024.174533 https://doi.org/10.1029/2018WR023939 https://doi.org/10.1073/pnas.1816892116 https://doi.org/10.1073/pnas.1816892116 https://doi.org/10.1029/92WR00252 https://doi.org/10.1016/j.scitotenv.2022.157791 https://doi.org/10.1016/j.scitotenv.2022.157791 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref54 http://refhub.elsevier.com/S0301-4797(25)01604-4/sref54 https://parsnip.tidymodels.org/reference/boost_tree.html https://parsnip.tidymodels.org/reference/boost_tree.html https://www.tidymodels.org https://doi.org/10.7931/DL1MG6 https://doi.org/10.7931/DL1MG6 https://doi.org/10.1016/j.scitotenv.2016.11.083 https://doi.org/10.1016/j.scitotenv.2016.11.083 https://doi.org/10.1080/00288330909510055 https://doi.org/10.1080/00288330909510055 https://doi.org/10.1111/j.1745-6584.2007.00385.x https://doi.org/10.1111/j.1745-6584.2007.00385.x https://environment.govt.nz/publications/ https://environment.govt.nz/publications/ https://archivescentral.org.nz/records https://archivescentral.org.nz/records https://lris.scinfo.org.nz/document/9162-lris-data-dictionary-v3/ https://lris.scinfo.org.nz/document/9162-lris-data-dictionary-v3/ https://doi.org/10.1016/j.envsoft.2024.105971 https://doi.org/10.1016/j.scitotenv.2016.06.148 https://doi.org/10.1016/j.scitotenv.2016.06.148 https://www.r-project.org/ https://www.r-project.org/ https://doi.org/10.1016/j.scitotenv.2021.151065 https://doi.org/10.1016/j.scitotenv.2021.151065 Schamper, C., Sørensen, K.I., Therrien, R., Thirup, C., Viezzoli, A., 2014. Nitrate reduction in geologically heterogeneous catchments - a framework for assessing the scale of predictive capability of hydrological models. Sci. Total Environ. 468–469, 1278–1288. https://doi.org/10.1016/j.scitotenv.2013.07.042. Rivas, A., Singh, R., Bishop, P., Horne, D., Roygard, J., Hedley, M., 2014. Measuring denitrification in the subsurface environment of the Manawatu River catchment. In: Currie, L.D., Christensen, C.L. (Eds.), Nutrient Management for the Farm, Catchment and Community, Occasional Report No. 27. Fertilizer and Lime Research Centre. Massey University, Palmerston North, New Zealand, p. 13. http://flrc.massey.ac. nz/publications.html. (Accessed 21 January 2025). Rivas, A., Singh, R., Horne, D., Roygard, J., Matthews, A., Hedley, M.J., 2017. Denitrification potential in the subsurface environment in the Manawatu River catchment, New Zealand: indications from oxidation-reduction conditions, hydrogeological factors, and implications for nutrient management. J. Environ. Manage. 197, 476–489. https://doi.org/10.1016/j.jenvman.2017.04.015. Rivas, A., Singh, R., Horne, D.J., Roygard, J., Matthews, A., Hedley, M.J., 2020. Contrasting subsurface denitrification characteristics under temperate pasture lands and its implications for nutrient management in agricultural catchments. J. Environ. Manage. 272, 111067. https://doi.org/10.1016/j.jenvman.2020.111067. Rivett, M.O., Buss, S.R., Morgan, P., Smith, J.W.N., Bemment, C.D., 2008. Nitrate attenuation in groundwater: a review of biogeochemical controlling processes. Water Res. 42 (16), 4215–4232. https://doi.org/10.1016/j.watres.2008.07.020. Rosecrans, C.Z., Nolan, B.T., Gronberg, J.A.M., 2017. Prediction and visualization of redox conditions in the groundwater of Central Valley, California. J. Hydrol. 546, 341–356. https://doi.org/10.1016/j.jhydrol.2017.01.014. Sanwar, S., 2023. Characterisation and potential optimisation of seepage wetlands for nitrate mitigation. In: New Zealand Hill Country (Doctor of Philosophy). Massey University, Palmerston North, New Zealand. https://mro.massey.ac.nz. (Accessed 12 March 2025). Sanwar, S., Burkitt, L., Singh, R., Bretherton, M., Kereszturi, G., 2022. Hill country wetland hydrological characterisation for nitrate attenuation. In: Christensen, C.L., Singh, R., Horne, D.J. (Eds.), Adaptive Strategies for Future Farming, Occasional Report No. 34. Farmed Landscapes Research Centre. Massey University, Palmerston North, New Zealand, p. 11. http://flrc.massey.ac.nz/publications.html. (Accessed 21 January 2025). Sarris, T.S., Close, M.E., Moore, C., 2019. Uncertainty assessment of nitrate reduction in heterogeneous aquifers under uncertain redox conditions. Stoch. Environ. Res. Risk Assess. 33 (8–9), 1609–1627. https://doi.org/10.1007/s00477-019-01715-w. Sarris, T.S., Wilson, S.R., Close, M.E., Abraham, P., Kenny, A., 2024. Reducing uncertainty of groundwater redox condition predictions at national scale, for decision making and policy. Environ. Manage. https://doi.org/10.1007/s00267- 024-02098-7. Savage, C., Leavitt, P.R., Elmgren, R., 2010. Effects of land use, urbanization, and climate variability on coastal eutrophication in the Baltic Sea. Limnol. Oceanogr. 55 (3), 1033–1046. https://doi.org/10.4319/lo.2010.55.3.1033. Singh, R., Elwan, A., Horne, D., Manderson, A., Patterson, M., 2017a. Predicting land- based nitrogen loads and attenuation in the Rangitikei River catchment - the model development. In: Currie, L.D., Hedley, M.J. (Eds.), Science and Policy: Nutrient Management Challenges for the Next Generation, Occasional Report No. 30. Fertilizer and Lime Research Centre. Massey University, Palmerston North, New Zealand, p. 13. http://flrc.massey.ac.nz/publications.html. (Accessed 21 January 2025). Singh, R., Elwan, A., Horne, D., Matthews, A., 2017b. Nitrogen Accounting in the Rangitikei Catchment (No. 2017/EXT/1529). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentral.org.nz/records. (Accessed 21 January 2025). Singh, R., Horne, D.J., 2019. Water-quality issues facing dairy farming: potential natural and built attenuation of nitrate losses in sensitive agricultural catchments. Anim. Prod. Sci. 60 (1), 67. https://doi.org/10.1071/AN19142. Smith, G., Singh, R., Matthews, A., 2017. Assessing farm-scale nutrient flow pathways and nitrate attenuation in rangitikei sand country. In: Currie, L.D., Hedley, M.J. (Eds.), Science and Policy: Nutrient Management Challenges for the Next Generation, Occasional Report No. 30. Fertilizer and Lime Research Centre. Massey University, Palmerston North, New Zealand, p. 12. http://flrc.massey.ac.nz/publicat ions.html. (Accessed 21 January 2025). Snelder, T., Cox, T., Kerr, T., Fraser, C., Collins, S., 2020. Manawatū-Whanganui Region Catchment Nitrogen Models Supporting Regional Plan Change 2 (No. 2020/EXT/ 1697). Horizons Regional Council, Palmerston North, New Zealand. https://arch ivescentral.org.nz/records. (Accessed 21 January 2025). Snelder, T.H., McDowell, R.W., Fraser, C.E., 2017. Estimation of catchment nutrient loads in New Zealand using monthly water quality monitoring data. J. Am. Water Resour. Assoc. 53 (1), 158–178. https://doi.org/10.1111/1752-1688.12492. Steiness, M., Jessen, S., Van’T Veen, S.G.M., Kofod, T., Højberg, A.L., Engesgaard, P., 2021. Nitrogen-loads to streams: importance of bypass flow and nitrate removal processes. J. Geophys. Res. Biogeosciences 126 (5), e2020JG006111. https://doi. org/10.1029/2020JG006111. Stenger, R., Clague, J.C., Morgenstern, U., Clough, T.J., 2018. Vertical stratification of redox conditions, denitrification and recharge in shallow groundwater on a volcanic hillslope containing relict organic matter. Sci. Total Environ. 639, 1205–1219. https://doi.org/10.1016/j.scitotenv.2018.05.122. Tesoriero, A.J., Terziotti, S., Abrams, D.B., 2015. Predicting redox conditions in groundwater at a regional scale. Environ. Sci. Technol. 49 (16), 9657–9664. https:// doi.org/10.1021/acs.est.5b01869. Tesoriero, A.J., Wherry, S.A., Dupuy, D.I., Johnson, T.D., 2024. Predicting redox conditions in groundwater at a national scale using random forest classification. Environ. Sci. Technol. 58 (11), 5079–5092. https://doi.org/10.1021/acs. est.3c07576. Thomas, N., 2017. Santoft Groundwater Model Report (No. 2017/EXT/1556). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentral.org. nz/records. (Accessed 21 January 2025). Tscharntke, T., Clough, Y., Wanger, T.C., Jackson, L., Motzke, I., Perfecto, I., Vandermeer, J., Whitbread, A., 2012. Global food security, biodiversity conservation and the future of agricultural intensification. Biol. Conserv. 151 (1), 53–59. https:// doi.org/10.1016/j.biocon.2012.01.068. Viera, A.J., Garrett, J.M., 2005. Understanding interobserver agreement: the kappa statistic. Fam. Med. 37 (5), 360–363. https://pubmed.ncbi.nlm.nih.gov/15883903/. (Accessed 21 January 2025). White, P.A., Moreau, M., Mourot, F., Rawlinson, Z., 2019. New Zealand Groundwater Atlas: Hydrogeological-Unit Map of New Zealand (Consultancy Report No. 2019/ 144). Institute of Geological and Nuclear Sciences, Lower Hutt, New Zealand. Revised 2023. https://www.gns.cri.nz/data-and-resources/new-zealand-hydrogeol ogical-unit-map/. (Accessed 21 January 2025). Wilson, S.R., Close, M.E., Abraham, P., 2018. Applying linear discriminant analysis to predict groundwater redox conditions conducive to denitrification. J. Hydrol. 556, 611–624. https://doi.org/10.1016/j.jhydrol.2017.11.045. Wilson, S.R., Close, M.E., Abraham, P., Sarris, T.S., Banasiak, L., Stenger, R., Hadfield, J., 2020. Achieving unbiased predictions of national-scale groundwater redox conditions via data oversampling and statistical learning. Sci. Total Environ. 705, 135877. https://doi.org/10.1016/j.scitotenv.2019.135877. Withers, P.J.A., Neal, C., Jarvie, H.P., Doody, D.G., 2014. Agriculture and eutrophication: where do we go from here? Sustain. Switz. 6 (9), 5853–5875. https://doi.org/10.3390/su6095853. Wu, Y., Xu, L., Wang, S., Wang, Z., Shang, J., Li, X., Zheng, C., 2018. Nitrate attenuation in low-permeability sediments based on isotopic and microbial analyses. Sci. Total Environ. 618, 15–25. https://doi.org/10.1016/j.scitotenv.2017.11.039. Xie, F., Cai, G., Li, G., Li, H., Chen, X., Liu, Y., Zhang, W., Zhang, J., Zhao, X., Tang, Z., 2024. Basin-wide tracking of nitrate cycling in Yangtze River through dual isotope and machine learning. Sci. Total Environ. 912, 169656. https://doi.org/10.1016/j. scitotenv.2023.169656. Yang, M., Lee, J., Jang, S., Annable, M.D., Jawitz, J.W., 2023. Nitrate attenuation potential in karst conduits and aquifer matrix. J. Hydrol. 624, 129896. https://doi. org/10.1016/j.jhydrol.2023.129896. Zarour, H., 2008. Groundwater Resources in the Manawatu-Wanganui Region (No. 2008/EXT/948). Horizons Regional Council, Palmerston North, New Zealand. https://archivescentral.org.nz/records. (Accessed 21 January 2025). S.B. Collins et al. Journal of Environmental Management 384 (2025) 125628 13 https://doi.org/10.1016/j.scitotenv.2013.07.042 http://flrc.massey.ac.nz/publications.html http://flrc.massey.ac.nz/publications.html https://doi.org/10.1016/j.jenvman.2017.04.015 https://doi.org/10.1016/j.jenvman.2020.111067 https://doi.org/10.1016/j.watres.2008.07.020 https://doi.org/10.1016/j.jhydrol.2017.01.014 https://mro.massey.ac.nz http://flrc.massey.ac.nz/publications.html https://doi.org/10.1007/s00477-019-01715-w https://doi.org/10.1007/s00267-024-02098-7 https://doi.org/10.1007/s00267-024-02098-7 https://doi.org/10.4319/lo.2010.55.3.1033 http://flrc.massey.ac.nz/publications.html https://archivescentral.org.nz/records https://doi.org/10.1071/AN19142 http://flrc.massey.ac.nz/publications.html http://flrc.massey.ac.nz/publications.html https://archivescentral.org.nz/records https://archivescentral.org.nz/records https://doi.org/10.1111/1752-1688.12492 https://doi.org/10.1029/2020JG006111 https://doi.org/10.1029/2020JG006111 https://doi.org/10.1016/j.scitotenv.2018.05.122 https://doi.org/10.1021/acs.est.5b01869 https://doi.org/10.1021/acs.est.5b01869 https://doi.org/10.1021/acs.est.3c07576 https://doi.org/10.1021/acs.est.3c07576 https://archivescentral.org.nz/records https://archivescentral.org.nz/records https://doi.org/10.1016/j.biocon.2012.01.068 https://doi.org/10.1016/j.biocon.2012.01.068 https://pubmed.ncbi.nlm.nih.gov/15883903/ https://www.gns.cri.nz/data-and-resources/new-zealand-hydrogeological-unit-map/ https://www.gns.cri.nz/data-and-resources/new-zealand-hydrogeological-unit-map/ https://doi.org/10.1016/j.jhydrol.2017.11.045 https://doi.org/10.1016/j.scitotenv.2019.135877 https://doi.org/10.3390/su6095853 https://doi.org/10.1016/j.scitotenv.2017.11.039 https://doi.org/10.1016/j.scitotenv.2023.169656 https://doi.org/10.1016/j.scitotenv.2023.169656 https://doi.org/10.1016/j.jhydrol.2023.129896 https://doi.org/10.1016/j.jhydrol.2023.129896 https://archivescentral.org.nz/records Modelling and mapping of subsurface nitrate-attenuation index in agricultural landscapes 1 Introduction 2 Data and methods 2.1 Study area 2.2 Groundwater redox assessments and landscape characteristics 2.3 Prediction of landscape unit subsurface redox status 2.3.1 Ground-truthing landscape subsurface redox predictions 2.4 Development of the landscape subsurface nitrate-attenuation index 3 Results and discussion 3.1 Prediction of landscape unit subsurface redox status 3.1.1 Ground-truthing of landscape subsurface redox predictions 3.2 Mapping of landscape subsurface nitrate-attenuation index 4 Conclusion CRediT authorship contribution statement Declaration of competing interest Acknowledgements Appendix A Supplementary data Data availability References