First Predictive Mapping of Persistent Organic Pollutants in Mosses Across Germany, 2020

Abstract

Persistent organic pollutants (POPs) are globally distributed toxic contaminants. Since 1990, mosses have been used in the UNECE European Moss Survey as cost-effective biomonitors of atmospheric deposition. This study provides the first predictive maps of POP concentrations in mosses, revealing nationwide contamination patterns across Germany. As a case study within the Moss Survey, predictive models were built from POP concentrations measured at 21 sites in 2020 and combined with environmental and land-use data. Random Forest analyses explained more than 20% of the variance for seven of eleven POP groups, yielding robust spatial estimates, particularly for PAH, BDE 209, and DBDPE, despite moderate systematic differences. Explanatory power was limited for PCDD/F, PCDD/F TEQ values, DPTE, and HBBz, while HBCD, PBDE, DP, and PBT showed a moderate performance. A comparison with geostatistical reference maps indicated moderate to good concordance, though regional uncertainties persisted. Industrialized regions such as North Rhine–Westphalia, Rhine Neckar, Halle/Leipzig, and Saarland emerged as consistent hotspots, whereas rural and forested areas showed lower contamination. The findings highlight the value of moss surveys for spatial POP assessment and underscore the need for additional predictors, especially atmospheric deposition, and for integrating Random Forest models with geostatistical approaches such as regression kriging to enhance predictive accuracy.

1. Introduction

Persistent organic pollutants (POPs) are among the most significant global environmental contaminants due to their high chemical stability, toxicity, and long-range transport potential. They enter terrestrial and aquatic ecosystems via atmospheric processes, where they accumulate in soils, sediments, and biota [1]. Their mainly lipophilic properties promote bioaccumulation in organisms and biomagnification along food webs, with far-reaching consequences for biodiversity, ecosystem functioning, and human health [2,3].

Given their ecotoxicological relevance, a comprehensive assessment of the environmental impacts of POPs is required, along with—analogous to heavy metals (HMs) and nitrogen compounds (Ns)—a precise quantification of the atmospheric contribution to the contamination of terrestrial ecosystems. In this context, mosses (Bryophyta) have proven to be particularly effective biomonitors [4,5]. Owing to their poikilohydric physiology and the absence of a root system, the uptake of water and dissolved substances occurs almost exclusively through the surface directly from the atmosphere. These characteristics make mosses highly suitable for substance-specific environmental monitoring, especially for analysing spatial and temporal patterns of atmospheric deposition. Complementing technical deposition collectors and chemical transport models, mosses provide a spatially extensive and cost-effective method for assessing atmospheric depositions [5,6,7,8,9]. Their application enables the identification of regional pollution hotspots and the analysis of temporal trends, thereby supporting the development of air-quality measures and the protection of sensitive ecosystems in a manner complementary to other monitoring approaches.

A central instrument for implementing this monitoring approach is the United Nations Economic Commission for Europe (UNECE) European Moss Survey, conducted within the framework of the International Cooperative Programme on Effects of Air Pollution on Natural Vegetation and Crops [10]. Since 1990, moss samples have been collected every five years at up to approximately 7300 sites across as many as 32 European countries to determine concentrations of HMs, Ns, POPs, and, more recently, microplastics (MPs) [10]. Germany has participated in this Europe wide monitoring programme from the outset and has carried out six national campaigns to date: HMs have been monitored since 1990, Ns since 2005, POPs since 2015, and microplastics (MPs) since 2020 [11,12,13,14,15,16]. In the 2020 Moss Survey (MM2020), moss samples were collected at up to 21 sites in Germany to chemically analyse and statistically evaluate POP concentrations in mosses as indicators of atmospheric deposition [14]. The POP analyte spectrum comprised a wide range of environmentally relevant compounds, including polycyclic aromatic hydrocarbons (PAHs), polychlorinated dibenzo p dioxins and dibenzofurans (PCDDs/Fs), polychlorinated biphenyls (PCBs), brominated and chlorinated flame retardants such as polybrominated biphenyls (PBBs) and polybrominated diphenyl ethers (PBDE), additional novel brominated flame retardants (nBFRs), and per- and polyfluoroalkyl substances (PFASs) (Table S1).

In the moss surveys, the spatial generalisation of point-based concentration measurements plays a key role in analysing spatio-temporal changes in atmospheric deposition to terrestrial ecosystems [13]. For the first time, the 2020 Moss Survey spatially generalized measured POP concentrations across Germany using geostatistical interpolation between moss sampling sites [15]. In parallel, POP concentrations from up to 21 sites in Germany were analysed using Random Forest (RF) regression to assess their statistical relationships with potentially explanatory variables [14]. The multivariate analysis identified atmospheric deposition (based on EMEP/MSC-E modelling), the density of urban–industrial land-use and agricultural areas within a 100–300 km radius range, as well as local emissions sources, as key influencing factors. Distances to smaller built-up areas, arable land, and transport infrastructure also proved highly relevant. Vegetation structure showed strong effects as well, with leaf area index, distance to trees and shrubs, and their canopy cover emerging as additional major predictors [14]. A nationwide spatial generalisation using the RF models was not feasible, however, because the analysis relied on environmental parameters that, according to the Moss Manual [10], were collected only directly at the moss sampling sites. For these parameters—particularly vegetation structure metrics and distances to potential local emissions sources—no spatially comprehensive datasets were available.

The aim of this study is therefore to implement the most important predictors identified in the RF analysis using nationally available, spatially comprehensive data sources in order to establish, for the first time, a basis for nationwide predictive mapping of POP concentrations in mosses in Germany. To achieve this, we (1) compiled nationally available, spatially comprehensive datasets for potential predictors, (2) selected the key predictors identified in the RF analysis, (3) applied a fuzzy clustering approach to group highly correlated variables and examined whether this could reduce model complexity and improve the efficiency of the RF models, (4) built RF models using the resulting predictor sets, and (5) generated nationwide prediction maps. These maps were then (6) compared with geostatistical surface estimates from the 2020 Moss Monitoring [15] to evaluate model performance and (7) analysed in order to identify potential regional uncertainties, particularly in areas where the RF predictions and geostatistical surface estimates show substantial discrepancies.

2. Materials and Methods

2.1. Data

The present study is based on the concentrations of POPs in mosses collected during the 2020 Moss Survey in accordance with the recommendations of the Moss Manual [10]. Sample collection and the recording of sample- and site-specific field data were carried out at up to 21 sites in Germany (Figure 1). The moss samples included widespread species such as Pleurozium schreberi (BRID.) MITT., Hypnum cupressiforme HEDW. s.str., and Pseudoscleropodium purum (HEDW.) M. FLEISCH (syn. Scleropodium purum HEDW. LIMPR.). Samples for POP analysis were collected between August and September 2021. At each sampling site, two 2.5 L wide-neck amber glass bottles were filled with moss using nitrile gloves. The bottles had been pre-cleaned in the laboratory with methanol, acetone, and hexane and subsequently heated at 250 °C overnight.

The POP analysis was performed using validated and accredited [17] procedures for 120 target compounds and largely followed the methods of the 2015 moss survey [16]. Samples were transported and stored under cool and dark conditions; light-sensitive analytes (e.g., HFRs) were protected from light throughout the analytical process. Moss samples were dried at 40 °C and subsequently homogenised. To correct for analytical losses, more than 90 isotopically labelled standards were used. In addition, certified reference materials and a field blank were analysed. To ensure a sufficiently dense and statistically valid sampling network, and in view of the high number of measurements below the limit of quantification (LOQ), the statistical evaluation was restricted to substance-group-specific sum concentrations. Values below the LOQ were set to zero (Σ excl. LOQ with LOQ = 0) to retain information on very low concentrations in the analysis. The spectrum of analysed (persistent) organic contaminants comprised a wide range of environmentally relevant compounds. The following compound groups and individual compounds were considered (all acronyms are defined in the Abbreviations Section):

• Sum concentrations: 16 PAH, PCDD/F, PCDD/F toxic equivalent (TEQ) values, HBCD, 7 PBDE congeners (BDE-28, -47, -99, -100, -153, -154, -183), Dechlorane Plus (DP; sum of syn-DP and anti-DP);
• Individual compounds: BDE 209, DPTE, HBBz, PBT, and DBDPE.

For the substance groups PCB, PBB, and PFAS, the number of measurements above the respective LOQ was too low to allow for reliable analyses, even at the level of sum concentrations. All 120 chemically analysed compounds, the range of sample-specific LOQs, and the selection of compounds included in the sum calculations and in the statistical evaluation are presented in detail in Table S1.

The selection of potentially explanatory variables for POP concentrations in mosses—referring to environmental parameters that are theoretically or empirically expected to influence or account for the spatial distribution and level of POP concentrations—builds on the results of the preceding multivariate analysis of factors influencing POP and MP concentrations in the 2020 Moss Survey in Germany [14]. This analysis identified numerous environmental parameters with high explanatory power, including vegetation-structure characteristics and the distance to potential local emissions sources. However, for these parameters, only site-specific, point-based data from the 21 moss sampling sites shown in Figure 1 were available. In view of the objective of this study—predictive mapping of POP concentrations across Germany—an effort was therefore made to derive these relevant environmental characteristics from publicly accessible data sources with nationwide coverage. Combined with already established Germany-wide datasets on atmospheric deposition, meteorology, topography, population density, and land-use structure, which have demonstrated strong explanatory capacity [14], these datasets are to be complemented by additional nationwide data on vegetation-structure characteristics and distances to potential local emissions sources, thereby further improving spatial modelling at the national scale.

Figure 1. Spatial distribution of sampling sites in the 2020 Moss Survey.

Figure 1. Spatial distribution of sampling sites in the 2020 Moss Survey.

The parameters listed in Table 1 and described in the following sections were therefore derived both for the 21 sampling areas (each 50 m × 50 m in size) and nationwide for Germany using a 3 km × 3 km grid.

Table 1. Potential explanatory variables for POP concentrations in mosses (MM2020).

Category	Variable Names	Description	Unit
Atmospheric deposition 1
Annual total deposition of B[a]P for the years 2018–2020 (EMEP 0.1° × 0.1°)	bap_dep_2018, bap_dep_2019, bap_dep_2020, bap_dep_2019_2020, bap_dep_2018_2020	Modelled total deposition of benzo[a]pyrene (EMEP; annual values and multi-year means 2018–2020)	μg m−2 a−1
Annual total deposition of PCDD/F TEQ values for the years 2018–2020 (EMEP 0.1° × 0.1°)	pcdd_dep_2018, pcdd_dep_2019, pcdd_dep_2020, pcdd_dep_2019_2020, pcdd_dep_2018_2020	Modelled total deposition of polychlorinated dibenzo-p-dioxins and furans (EMEP; toxicity equivalents; annual values and multi-year means 2018–2020)	ng TEQ m−2 a−1
Meteorology 2
Precipitation	prec18, prec19, prec20, prec18_20	Mean annual precipitation (annual and multi-year means 2018–2020)	mm a−1
Topography 3
Relief	elev_eu_gk	Orographic elevation	m a.s.l.
Vegetation structure 4
Leaf area index	lai_eu	Leaf area index (Sentinel-2 data)	---
Population density 5
Population density within radii of 5, 50, 100 km around the sampling site	popdens5–100	Population density	inhabitants per km2
Land-use density 6
Spatial density of surrounding land uses within radii of 5, 50, 100, 300 km around the sampling site	agr5–300	Density of agricultural land use	[%]
	arab5–300	Density of arable land use
	for5–300	Density of forest land use
	urb5–300	Density of urban–industrial land use
	indu5–300	Density of industrial land use
	urbf5–300	Density of urban land use
	mine5–300	Density of mineral extraction sites, dump sites, and construction sites
Land-use density (combined clusters)	LU_UrbCore	Urban cores with industrial characteristics
	LU_AgUrb	Mixed agricultural and urban–industrial areas
	LU_AgMix	Agriculturally dominated areas
	LU_Mix	Densely populated urban mixed-use areas
	LU_SubInd	Suburban areas with industrial characteristics
	LU_Forest	Near-natural and forestry-dominated areas
Distance to potential emissions sources 7
Distance to local emissions sources	Dist_indu	Distance to industrial and commercial areas	[km]
	Dist_urbf	Distance to residential areas
	Dist_arab	Distance to arable land
	Dist_highway	Distance to motorways
	Dist_road	Distance to federal roads
	Dist_rail	Distance to railway lines
Distance to potential emissions sources (combined clusters)	Dist_Cent	Distance to urban–industrial centres
	Dist_Infra	Distance to infrastructure and transport corridors
	Dist_RemNat	Distance from built-up areas and transport infrastructure

¹ EMEP/MSC-E: https://msceast.org/pollution-assessment/emep-domain-menu/data-hm-pop-menu [accessed on 8 November 2022]; ² DWD Climate Data Center (CDC), annual total of monthly precipitation raster data for Germany: https://opendata.dwd.de/climate_environment/CDC/grids_germany/annual/ [accessed on 19 December 2025]; ³ http://www.jpl.nasa.gov/srtm/cbanddataproducts.html [accessed on 1 September 2015]; ⁴ https://land.copernicus.eu [accessed on 15 July 2019]; ⁵ https://www.earthdata.nasa.gov/data/projects/gpw [accessed on 15 July 2019]; ⁶ CORINE Land Cover: https://land.copernicus.eu/en/products/corine-land-cover/clc2018 [accessed on 20 December 2024]; ⁷ Digital Landscape Model 1:250.000 (DLM 250): https://gdz.bkg.bund.de [accessed on 16 August 2025].

Atmospheric deposition. Within the European Monitoring and Evaluation Programme (EMEP), spatial distributions of the deposition of POPs were calculated using the MSC-E-POP model developed by the EMEP Meteorological Synthesizing Centre–East [18]. The model integrates multiple environmental compartments, including the atmosphere, soil, oceans, and vegetation—as well as relevant processes, such as emissions, long-range transport, deposition, degradation, and gas exchange between the atmosphere and the Earth’s surface [19]. The spatial resolution of the annually modelled total deposition for the EMEP region, and thus also for Germany, is 0.1° × 0.1°. In line with the two- to three-year moss shoots used for chemical analysis [10], this study used the three-year mean (2018–2020), the two-year mean (2019–2020), and the annual values of the total deposition of benzo[a]pyrene (B[a]P) and polychlorinated dibenzo-p-dioxins and dibenzofurans expressed as toxicity equivalent values (PCDD/F TEQ values), as modelled by MSC-E, as potential predictor variables. The available modelling data for the deposition of hexachlorobenzene (HCB) and polychlorinated biphenyls (PCB-153) could not be considered, as PCB-153 was not detected above the LOQ in any sample and HCB was analysed in too few samples to allow a meaningful evaluation. Deposition data for the other POPs were not available.

Meteorology. Precipitation was considered as a potential influencing factor and as an indicator of wet atmospheric deposition. The analysis was based on data from the German Meteorological Service (Deutscher Wetterdienst, DWD, Offenbach am Main, Hesse), from which the arithmetic mean of the annual precipitation totals was calculated for the period 2018–2020 at a 1 km × 1 km spatial resolution.

Topography. To account for the elevation of the moss sampling sites as a predictor variable, the Digital Elevation Model of the Shuttle Radar Topography Mission (SRTM, 90 m × 90 m) was used.

Vegetation structure. To assess the potential canopy-drip effect of trees, the leaf area index (LAI) was used as an indicator of vegetation structure. The LAI data originate from the Copernicus service for the year 2018 and are based on analyses of Sentinel-2 satellite imagery. With a spatial resolution of approximately 2 km × 2 km, these data provide nationwide coverage of vegetation density in Germany.

Population density. Population density was examined as an indicator of the spatial concentration of potential POP emissions sources. The analysis was based on data from the Gridded Population of the World (GPW) for the year 2015, with a spatial resolution of approximately 1 km × 1 km. Aggregated population density values were derived as predictor variables by calculating mean values within radii of 5, 50, and 100 km around each sampling site.

Land-use density. To capture land-use density as an additional indicator of the potential density of emissions sources, land-use classes from Corine Land Cover 2018 [20] were used. The classification comprised seven groups [21]: agricultural areas (CLC categories: 211 (Non-irrigated arable land), 212 (Permanently irrigated land), 213 (Rice fields), 221 (Vineyards), 222 (Fruit trees and berry plantations), 223 (Olive groves), 231 (Pastures), 241 (Annual crops associated with permanent crops), 242 (Complex cultivation patterns), 243 (Land principally occupied by agriculture, with significant areas of natural vegetation), 244 (Agro-forestry areas)), arable land (CLC categories: 211 (Non-irrigated arable land), 212 (Permanently irrigated land), 213 (Rice fields)), forests (CLC categories: 311 (Broad-leaved forest), 312 (Coniferous forest), 313 (Mixed forest)), urban–industrial areas (CLC categories: 111 (Continuous urban fabric), 112 (Discontinuous urban fabric), 121 (Industrial or commercial units), 122 (Road and rail networks and associated land), 123 (Port areas), 124 (Airports), 131 (Mineral extraction sites), 132 (Dump sites), 133 (Construction sites))—subdivided into urban land uses (CLC categories: 111 (Continuous urban fabric), 112 (Discontinuous urban fabric)), industrial land uses (CLC categories: 121 (Industrial or commercial units), 122 (Road and rail networks and associated land), 123 (Port areas), 124 (Airports)), and extraction sites and dump sites (CLC categories: 131 (Mineral extraction sites), 132 (Dump sites), 133 (Construction sites)). The proportional area shares of these groups were calculated using neighbourhood analysis (ESRI ArcGIS 10.1 Focal Statistics) within radii of 5 km and 50 km (based on CLC data at 100 m × 100 m resolution) as well as 100 km and 300 km (resampled to 250 m × 250 m) around the moss sampling sites, providing nationwide coverage for Germany.

Distance to potential emissions sources. To assess the local influence of potential emissions sources of POPs, the Digital Landscape Model 1:250,000 (DLM250) of the Federal Agency for Cartography and Geodesy (BKG, Frankfurt am Main) was used, which is updated annually and was applied here in its 2024 version. Based on this topographic vector dataset, distances to relevant land-use categories were determined nationwide for Germany. These categories had been identified in previous studies [14] as influential factors affecting POP concentrations in moss samples. They include industrial and commercial areas, residential zones, intensively used agricultural land, and transport infrastructures such as motorways, federal roads, and railway lines. The positional accuracy of the objects contained in the DLM250 is ±100 m [22]. The calculated distances serve as spatial indicators of potential input pathways and were examined as possible predictors for the nationwide modelling of POP concentrations in mosses.

2.2. Grouping of Potential Predictor Variables

To classify and reduce the dimensionality of land-use density around the moss sampling sites, as well as their distance to potential emissions sources, a Fuzzy C Means (FCM) clustering was applied to the predictor variables. For each of the 21 moss sampling sites with measured concentrations of POPs, 28 variables describing land-use density and 6 distance metrics to relevant land-use categories were available. The clustering was not performed on the individual sites, but rather on the similarity patterns of the variables across all sites. The aim was to identify and group predictors that may affect comparable environmental influences on POP concentrations in mosses. In contrast to hard clustering approaches, the FCM method allows a fuzzy assignment of each variable to multiple clusters [23]. The resulting membership matrix, based on values determined by membership functions, assigns each predictor variable a continuous degree of membership (ranging from 0 to 1) to each of the identified clusters, thereby capturing functional overlaps—such as those between traffic, residential, and industrial areas—more realistically than classical methods like k means.

The number of clusters in this study was chosen based on thematic plausibility in order to group relevant spatial influencing factors in a clear and interpretable manner. The Fuzzy C Means analysis was performed using a fuzzification parameter of m = 2, which assigns each predictor a fuzzy, gradual membership to multiple clusters. This approach enables an integrative characterization of different features by assigning each predictor a degree of membership to several cluster centres.

Based on this approach, six cluster centres were identified for the land-use density parameters and three cluster centres for the distance metrics, enabling a grouping of predictors according to spatial density and proximity relationships. The resulting six land-use clusters can be thematically characterized as follows (Table 1):

• LU_UrbCore: This cluster represents highly urbanized areas with dense urban structures and industrial infrastructure at a broad spatial scale. Typical for central city districts and metropolitan regions (urb300, urbf300, indu300).
• LU_AgUrb: This cluster describes mixed areas with both agricultural and urban–industrial characteristics within large radii around the moss sampling sites (agr300, arab300, mine300).
• LU_AgMix: This cluster represents intensively used agricultural areas (agr5, agr50, agr100, arab5, arab50, arab100).
• LU_Mix: This cluster structure comprises urban and industrial land uses with high density of built-up areas at a local scale. The associated variables indicate inner-city neighborhoods, commercial zones, and urban infrastructure (urb5, urbf5, indu5, mine50).
• LU_SubInd: This cluster describes suburban areas characterized by residential and industrial, land uses as well as extraction sites (urb50, urb100, urbf50, indu50, indu100, mine100).
• LU_Forest: This cluster represents largely undisturbed or forestry-managed areas with low anthropogenic influence. The high density of forest-related land uses across all radii indicates forested and near-natural landscapes (for5, for50, for100, for300).
• For the distance metrics, three emissions-related clusters were identified:
• Dist_Cent: Characterized by short distances to anthropogenic land uses such as industrial areas and arable land (dist_arab, dist_indu).
• Dist_Infra: This cluster shows medium to short distances to roads and motorways (dist_road, dist_highway).
• Dist_RemNat: Characterized by large distances to urban and infrastructural land uses (dist_urbf, dist_rail).

To make these cluster structures usable as new predictors, an aggregated cluster value Cij was calculated for each moss sampling site and each grid cell of the nationwide dataset. This value results from the weighted sum of the original land-use variables at each site i with the membership degrees of the individual variables to the respective cluster j serving as weighting factors (Equation (1)).

$$\mathrm{A}\mathrm{g}\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{c}\mathrm{l}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{r} \mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}:C_{ij}=\sum _{k=1}^p x_{ik}\cdot w_{kj}$$

(1)

where x_ik represents the value of the k-th land-use variable at site i and w_kj denotes the membership degree of this variable to cluster j. These cluster features were subsequently used as predictor variables in the multivariate analysis.

2.3. Multivariate Analysis of Factors Influencing POP Concentrations in Mosses

To investigate potential factors influencing the (sum) concentrations of POPs in mosses, the Random Forest (RF; [24]) method was applied. RF is a non-parametric ensemble approach based on the aggregation of a large number of decision trees and is particularly well suited for modelling complex, non-linear relationships [25]. Compared to single decision-tree methods such as CART [26], RF is more robust to variations in the training dataset, outliers, and overfitting due to its random selection of predictors and data subsets [27,28]. The procedure for the multivariate analysis of the potential predictors listed in Table 1 follows the approach of [14], where it is documented in detail.

Evaluation of model performance. The performance of the RF models was assessed using the so-called out-of-bag data (OOB; [29]). These are generated by randomly removing samples from the training dataset and serve as independent test data. The central metric used for evaluation is the mean squared error (MSE), calculated according to Equation (2).

$$\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n} \mathrm{S}\mathrm{q}\mathrm{u}\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{d} \mathrm{E}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{r}:{\mathrm{M}\mathrm{S}\mathrm{E}}_{\mathrm{O}\mathrm{O}\mathrm{B}}={\displaystyle \frac{1}{n}} \sum _{i=1}^n{\left\{y_i-y_i^{\mathrm{O}\mathrm{O}\mathrm{B}}\right\}}^2$$

(2)

An ideal MSE value is 0, which indicates a perfect concordance between measured and modelled data.

To assess model performance, the normalized root mean square error (NRMSE) was used. To ensure scale-independent comparability between models for different POP groups and concentration ranges, the RMSE was normalized using the range of the observed concentration values (Equation (3)).

$$\mathrm{Normalized} \mathrm{Root} \mathrm{Mean} \mathrm{Square} \mathrm{Error}: \mathrm{NRMSE}={\displaystyle \frac{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}{y_{\max }-y_{\min }}}={\displaystyle \frac{\sqrt{\frac{1}{n}\sum _{i=1}^n (y_i-{\widehat{y}}_i{)}^2}}{y_{\max }-y_{\min }}}$$

(3)

where y_i represents the observed concentration value, ŷ_i the value predicted by the model, n the number of sampling sites, and y_max–y_min the range of the concentration values. For interpretation, values below 0.1 were classified as indicating very high concordance, values between 0.1 and 0.2 as high concordance, and values between 0.2 and 0.3 as moderate concordance between measured and modelled concentrations.

The proportion of explained variance was quantified using a pseudo-R², calculated according to Equation (4).

$$\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{d} \mathrm{v}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e} [\%]:\mathrm{P}\mathrm{s}\mathrm{e}\mathrm{u}\mathrm{d}\mathrm{o}-{\mathrm{R}}^2=1-{\displaystyle \frac{{\mathrm{M}\mathrm{S}\mathrm{E}}_{\mathrm{O}\mathrm{O}\mathrm{B}}}{{\mathrm{S}\mathrm{D}}_y^2}}$$

(4)

SD denotes the standard deviation of the target variable, calculated as the square root of the uncorrected variance. Model performance was classified according to the proportion of explained variance as follows: weak (<30%), moderate (30–50%), good (50–70%), and very good (>70%) [30].

Evaluation of variable importance. The relative importance of the individual predictors was assessed based on the percentage increase in mean squared error (%IncMSE). To determine %IncMSE, the values of a given predictor are first randomly permuted while all other variables remain unchanged. This procedure is repeated for each predictor. Subsequently, for each RF model in which exactly one predictor has been permuted, the MSE is calculated according to Equation (3) and compared with the original model. A high %IncMSE value indicates—relative to the other predictors—a high importance of that predictor for explaining the target variable [31,32]. For interpretative purposes, predictors were qualitatively grouped according to their position within the model-specific distribution, with the upper quartile indicating comparatively high importance; this relative classification does not represent absolute effect sizes.

Model development and optimization. Predictors with low relevance were stepwise removed from the model through backward selection. The aim was to identify an optimized RF model with maximum explained variance (pseudo-R²).

All statistical analyses and modelling were performed using statistical packages of the R software environment [33]. Version 3.0.2 was used for the steps involving the rattle() package, while the remaining analyses were performed with version 4.1.2.

2.4. Spatial Prediction and Mapping of POP Concentrations in Mosses

To embed the sample-based measurements of POP concentrations in mosses into a broader spatial context, a nationwide spatial generalization was carried out for Germany. This was achieved through predictive mapping, in which the spatial distribution of the target variable is estimated using an empirical–statistical model [34]. In contrast to the commonly applied linear modelling approaches, the present study is based on non-parametric RF models, which allow for a more adequate representation of complex, non-linear relationships between predictor variables and concentration patterns in mosses. The statistical modelling was performed using full coverage predictor data at a spatial resolution of 3 km × 3 km, enabling high resolution predictive mapping of (sum) POP concentrations in mosses across the entire area of Germany.

To assess the spatial plausibility of the RF modelled concentration maps, a comparison with existing geostatistical surface estimates was conducted. The aim was to examine whether both map products exhibit comparable regional concentration hotspots, gradients, and spatial patterns, or whether contradictory or artefact-driven structures were produced. The geostatistical surface estimates are based on kriging interpolations of POP concentrations in mosses across Germany [15] and were generated independently of the RF results. Due to the limited nationwide sampling density of 21 sites, both geospatial datasets were aggregated to a common grid with a reduced resolution of 50 km × 50 km, with concentration values represented by the median within each grid cell.

The evaluation of spatial concordance between the modelling approaches was carried out both qualitatively and quantitatively. Visual map comparisons were used to identify spatial patterns, regional differences, and systematic deviations. This qualitative analysis supports the evaluation of spatial plausibility and provides indications of regional model uncertainties. In addition, scale-independent performance metrics were calculated to quantify spatial pattern similarity and systematic differences between the aggregated concentration maps.

Spatial pattern similarity was assessed using Spearman’s rank correlation coefficient (r_s). Values above 0.8 were interpreted as very strong similarity, values between 0.6 and 0.8 as strong, between 0.4 and 0.6 as moderate, between 0.2 and 0.4 as low, and values below 0.2 as very low similarity of spatial patterns in POP concentrations [35].

For a spatially differentiated analysis of systematic differences between the two modelling approaches, the logarithmic ratio of the POP-group-specific modelled concentrations in mosses was calculated for each grid cell. The resulting log ratio values were converted into percentage deviations (%log(A/B)) to determine the direction and magnitude of the cell-wise differences between the two model outputs (Equation (5)).

$$\mathrm{D}\mathrm{i}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e} \left[\%\right]: \%\log \left(A/B\right)=(e^(ln(R{F}_{\mathrm{i}}+\epsilon )-ln(Krigin{g}_{\mathrm{i}}+\epsilon ))-1)\times 100$$

(5)

where RF_i denotes the modelled concentration in grid cell i according to the Random Forest model, Kriging_i represents the corresponding concentration from the geostatistical surface estimation, and ε is a small constant added to avoid zero values. The central tendency across all grid cells was then determined using the median to summarize model deviations for each POP group. Positive deviations indicate that the RF model estimates higher concentrations than the geostatistical surface estimates, whereas negative deviations indicate lower RF estimates. Deviations below ±10% were classified as very small, ±10–30% as moderate, ±30–60% as substantial, and values above ±60% as strong systematic differences with potential uncertainty.

To assess the spatial concordance between the two map products, the upper 10% of concentration values were defined as hotspots and the lower 20% as coldspots based on their respective percentiles. The overlap of these grid cells was quantified using the Jaccard index [36], which describes the ratio of the intersection to the union of the grid cells identified as hotspots or coldspots by both models (Equation (6)).

$$\mathrm{J}\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{a}\mathrm{r}\mathrm{d} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}: \mathrm{J}=n_{11}/(n_{11}+n_{10}+n_{01})$$

(6)

In this context, n₁₁ represents the number of grid cells classified as hotspots (or coldspots) by both modelling approaches, n₁₀ those identified only by the RF model, and n₀₁ those identified only by the geostatistical surface estimation. Values above 0.6 were interpreted as high concordance, values between 0.3 and 0.6 as moderate concordance, and values below 0.3 as low concordance.

To capture systematic differences, the median of the differences (Median (A–B)) between the modelled concentrations was additionally calculated. Positive values indicate that the Random Forest model (A) tends to produce higher estimates, whereas negative values indicate lower estimates compared to the geostatistical surface estimation (B).

3. Results

3.1. Random Forest Models

Table 2 summarizes the model parameters and performance metrics of the Random Forest models used to relate POP concentrations in mosses to the environmental predictors listed in Table 3. By reporting sample sizes, model configurations, and predictive performance (MSE, NRMSE, pseudo-R²), the table provides the methodological basis for evaluating the reliability and stability of the models and thereby supports the interpretation of the subsequent results.

Table 2. Model parameters and model performance metrics of the Random Forest models describing the relationships between POP concentrations in mosses and the predictors shown in Table 3.

	Sum 16PAH	Sum PCDD/F	Sum PCDD/F TEQ Values	Sum HBCD	Sum 7 PBDE	BDE-209	Syn-DP + Anti-DP	DBDPE	DPTE	PBT	HBBz
Model parameters
n	21	20	20	21	21	21	21	21	21	21	21
n < LOQ	0	3	3	0	0	9	0	7	7	5	9
ntree	500	200	400	200	200	400	400	500	500	300	250
mtry	2	2	1	3	2	2	2	2	2	2	2
Model performance metrics
MSE	4642	79.85	0.0052	0.0481	3867	96,459	18,803	2,714,564	13,562	447.39	25.46
NRMSE	0.19	0.24	0.20	0.18	0.24	0.19	0.22	0.22	0.31	0.26	0.25
Pseudo-R2	33.3	15.2	17.1	30.3	24.0	55.7	26.6	33.8	16.1	30.9	16.9

n = sample size; n < LOQ = number of samples below the limit of quantification; ntree = number of regression trees in the Random Forest model; mtry = number of variables randomly selected at each split; MSE = mean squared error; NRMSE = normalized root mean square error; RSQ = proportion of explained variance (pseudo-R²).

Table 3. Overview of the relative importance of predictors in the Random Forest models explaining variation in POP concentrations in mosses.

Variable Names	Sum 16 PAH	Sum PCDD/F	Sum PCDD/F TEQ Values	Sum HBCD	Sum 7 PBDE	BDE-209	Syn-DP + Anti-DP	DBDPE	DPTE	PBT	HBBz	Count
Atmospheric deposition
bap_dep_2019_2020	2.3			---	---	---	---	---	---	---	---	---
pcdd_dep_2018		3.3	4.5	---	---	---	---	---	---	---	---	---
Meteorology
prec19							2.8					1
prec18_20			5.3						7.5			2
Spatial density of surrounding land use
agr5											5.3	1
agr50	5.7											1
agr100	8.6									3.6		2
agr300	6.9											1
arab100								4.4	2.1			2
for100	2.0							4.6				2
urb100				3.0		7.9		5.0		3.8		4
urb300				1.6								1
urbf5									7.8			1
urbf300				4.6								1
indu100	3.0					6.7	1.8	5.4			1.8	5
indu300					0.5					2.6		2
mine5					5.8							1
mine50					3.2				2.2			2
mine100	8.0		6.5									2
mine300						7.8						1
LU_AgMix				2.6						2.2		2
LU_AgUrb				3.1		6.4			4.3	5.7		4
LU_UrbCore		2.6		3.6								2
Population density
popdens5					2.6							1
popdens100				2.8						4.5		2
Distances to potential local emissions sources
dist_highway				1.3			3.0					2
dist_indu		4.7			4.2		5.7			4.4		4
dist_rail				3.5		6.0	5.5	3.1				4
dist_urbf		5.2			2.4				2.0		3.1	4
Dist_Cent				2.5			7.3					2
Dist_SubAg				2.9								1
Dist_RemNat					3.9					4.6		2
Other predictors
lai_eu	2.5							5.0			0.7	3
elev_eur_gk											2.1	1

Variable name descriptions are provided in Table 1; the values represent the percentage increase in the mean squared error (IncreasedMSE); bold = comparatively high predictor importance within the respective model; --- = not determinable due to missing deposition data.

Of the eleven Random Forest models used for the spatial estimation of POPs, one showed good predictive performance (pseudo-R² > 50%), four models achieved moderate performance (pseudo-R² between 30% and 50%), and six models displayed low explained variance (pseudo -R² < 30%) (Table 2).

Table 3 presents the relative importance of all environmental and spatial predictors included in the Random Forest models. The table highlights which variables most strongly explain the variation in POP concentrations in mosses across the study area. Based on these results, the RF analyses yield the following predictor-specific insights:

Atmospheric deposition. Atmospheric deposition represents a key input pathway for POP in mosses. The modelled PCDD/F TEQ values (EMEP 2018) show a close relationship with the measured concentrations and explain several of the observed spatial patterns. The modelled B[a]P depositions (EMEP 2019–2020) also prove to be a relevant predictor for the measured Σ16 PAH concentrations.

Meteorology. Precipitation shows clear relationships with POP concentrations in mosses. Mean precipitation in 2019 constitutes an important predictor for the sum of syn-DP and anti-DP, while mean precipitation in 2018–2020 is particularly strongly associated with PCDD/F TEQ. As the driving process behind wet deposition, precipitation was identified as a relevant predictor for three of the eleven POP groups in the multivariate analysis.

Spatial density of surrounding land uses. Land-use parameters show the highest overall model relevance. The density of agricultural land explains roughly half of the observed spatial patterns. Agricultural land use within a 100 km radius is particularly closely associated with Σ 16 PAH concentrations as well as with certain brominated flame retardants (DBDPE, DPTE, PBT). The density of urban–industrial land uses contributes to explaining ten of the eleven spatial patterns, especially between 100 km (urban–industrial uses) and 50 km (extraction sites, dump sites) around the moss sampling locations. Among the land-use density clusters, large-scale mixed agricultural and urban-industrial uses (LU_AgUrb) stand out, being of relatively high importance for four of the eleven POP groups.

Population density. The population density parameters (popdens5, popdens100) show moderate model relevance. They are primarily associated with the concentrations of brominated flame retardants in mosses, including HBCD, the sum of the seven PBDEs, and PBT, and they contribute additionally to explaining the spatial patterns of other POP concentrations.

Distance to local emissions sources. Proximity to local emissions sources represents another important factor influencing POP concentrations in mosses. Distances to industrial facilities, railway lines, and urban areas show particularly high %IncMSE values and statistically explain nine of the eleven observed spatial patterns. The distance-based clusters additionally contribute to explaining four of the investigated POP groups.

Other factors. The leaf area index (LAI) and orographic elevation show only low relative variable importance. While they provide supplementary indications for individual POP groups, they contribute substantially less to explaining the spatial patterns compared to the other predictor groups.

The results show that POP concentrations in mosses are statistically closely linked to agricultural and urban–industrial land-use structures as well as to the proximity of emissions sources. Atmospheric deposition and precipitation act as large-scale processes driving the transport and deposition of POPs to ecosystems. Population density complements this explanation as an indicator of urban emissions, but plays a comparatively minor role.

3.2. Spatial Prediction and Mapping

The spatial generalisation of POP concentrations in mosses was carried out through predictive mapping, in which the distribution of a target variable is estimated by applying an empirical–statistical model to spatially distributed predictor data [34]. The basis for this approach consisted of all RF models developed for the eleven contaminants (or contaminant groups), regardless of their individual model performance. The full application was conducted with the aim of enabling a comprehensive comparison of the prediction maps with the geostatistical surface estimates of the 2020 Moss Survey [15]. Models with limited explanatory power—meaning models that explain only a small proportion of the observed variance, reflected by weak pseudo-R² values—were also included, as they—as shown below—provide valuable insights into spatial patterns, potential data deficiencies, and regional uncertainties that are relevant for evaluating model quality and the underlying data sources.

Table 4 compiles the statistical metrics of the RF-based predictive mapping and compares the modelled POP concentration surfaces with geostatistical surface estimates. The table provides percentile ranges of predicted concentrations as well as measures of spatial agreement (systematic differences, correlations, and hotspot/coldspot overlaps) with the geostatistical estimates, providing a quantitative basis for assessing the robustness and spatial consistency of the RF predictions. Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12 depict the spatial distribution patterns of the eleven analysed POP groups in mosses across Germany. With the exception of DP and DPTE, for which no geostatistical surface estimates are available, each figure compares the RF regression maps with the corresponding geostatistical estimates, enabling a visual assessment of spatial agreement, hotspot patterns, and regional differences in predicted concentrations.

Table 4. Statistical metrics of RF-based predictive mapping and comparison measures with geostatistical surface estimates for POP concentrations in mosses (Germany, 2020).

	Sum 16 PAH	Sum PCDD/F	Sum PCDD/F TEQ Values	Sum HBCD	Sum 7 PBDE	BDE-209	Syn-DP + Anti-DP	DBDPE	DPTE	PBT	HBBz
Random Forest modelling
P20	51.3	6.13	0.0238	0.1583	67.35	245.5	141.3	1094	126.1	32.24	3.93
P50	74.0	7.99	0.0362	0.2285	81.41	378.4	170.7	1351	165.7	41.21	5.32
P90	138.0	14.18	0.1192	0.4660	120.83	860.1	242.2	2256	231.2	51.30	7.62
P98	177.3	18.83	0.1999	0.6450	154.25	1206.2	364.8	5635	241.7	57.00	9.00
Unit	ng/g	pg/g	pg TEQ/g	ng/g	pg/g	pg/g	pg/g	pg/g	pg/g	pg/g	pg/g
Comparison with geostatistical surface estimates
Median (A–B)	−0.90	0.89	0.00	0.02	2.38	97.77	---	186.44	---	2.09	−0.67
Spearman (rs)	0.78	0.58	0.39	0.66	0.54	0.64	---	0.55	---	0.58	0.54
%log(A/B)	−1.0	14.6	2.6	12.9	2.7	36.3	---	14.0	---	4.9	−12.7
Jaccard Hot (Jhot)	0.43	0.48	0.43	0.25	0.43	0.54	---	0.67	---	0.11	0.33
Jaccard Cold (Jcold)	0.48	0.33	0.20	0.25	0.31	0.43	---	0.21	---	0.48	0.27

P20, P50, P90, P98 = percentiles; Median (A–B) = median directional difference between RF modelling (A) and geostatistical estimates (B), unit: same as A and B; r_s = Spearman’s rank correlation coefficient; %log(A/B) = logarithmically normalised deviation (%); J^hot/J^cold = Jaccard indices for hotspot and coldspot overlaps.

Polycyclic aromatic hydrocarbons (PAHs). The presented RF model for the sum of the 16 PAH shows moderate statistical performance (Table 2). The normalized RMSE (NRMSE) is 0.19, placing it between good and moderate predictive quality. The pseudo-R² of 33.3% indicates that roughly one third of the variability in PAH concentrations can be explained by the selected predictors. The most influential variables are atmospheric B[a]P deposition (2019–2020) and the density of agricultural and industrial land uses as well as mining and dump sites within a 100 km radius around the moss sampling sites (Table 3). In addition, vegetation structure (leaf area index) and the density of forestry land use within the same radius contribute to explaining the spatial distribution.

A spatial comparison with the geostatistical surface estimates from the 2020 survey confirms the robustness of the model (Figure 2; Figure S1). Correlations are high (Spearman: 0.78), the median difference is −0.9 ng/g, and the logarithmic deviation is −1.0%, indicating very small systematic differences (Table 4). Spatially, hotspots with PAH concentrations above the 90th percentile (138 ng/g) were identified in the northern Ruhr district and in the Halle–Leipzig region. These hotspots show moderate concordance with the geostatistical reference data (Jaccard index for hotspots: 0.43). In rural and forested regions, concentrations are lower as expected, and here too the concordance between the model and the geostatistical reference is moderate (Jaccard index for coldspots: 0.48).

In summary, the RF model for PAH provides a robust spatial estimate with plausible predictor relationships and strong concordance with the geostatistical surface estimation, although some uncertainty remains in hotspot identification.

Figure 2. Spatial distribution of total concentrations of 16 PAHs (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 2. Spatial distribution of total concentrations of 16 PAHs (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Polychlorinated dibenzo-p-dioxins and dibenzofurans (PCDD/F and PCDD/F TEQ values). Model performance is weak in both cases (Table 2). The model for the sum of PCDD/F concentrations (pseudo-R² of 15.2%) as well as the model for PCDD/F TEQ values (17.1%) each explain less than one third of the variance. The NRMSE values of 0.24 (PCDD/F) and 0.20 (TEQ values) indicate a moderate concordance between measured and modelled concentrations. The most influential predictors are atmospheric deposition, meteorology, land-use density, and distance to local emissions sources (Table 3). For both target variables, atmospheric background levels play a role, as well as the distances to industrial and residential areas (PCDD/F), the density of mining and dump sites within a 100 km radius (PCDD/F TEQ values), and precipitation amounts (PCDD/F TEQ values).

In comparison with the geostatistical surface estimates, PCDD/F shows a moderate similarity in spatial patterns (Spearman’s rs = 0.58), whereas PCDD/F TEQ values is characterized by a low to moderate similarity (r_s = 0.39) (Figure 3 and Figure 4; Figures S2 and S3). The median of the differences is +0.89 pg/g for PCDD/F, indicating slightly higher RF estimates compared to the geostatistical reference (Table 4). For PCDD/F TEQ values, the median is 0.00 pg TEQ/g, suggesting a balanced estimation. The logarithmic deviation is +14.5% for PCDD/F, representing a moderate systematic difference, whereas for PCDD/F TEQ values it is very small at +2.56%.

Figure 3. Spatial distribution of total concentrations of PCDD/F (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 3. Spatial distribution of total concentrations of PCDD/F (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

The RF-modelled concentrations of PCDD/F and PCDD/F TEQ values in mosses reveal spatial hotspots with values above 14.1 pg/g and 0.12 pg TEQ/g, respectively, primarily in North Rhine–Westphalia and Saxony. The Jaccard indices for the concordance of these hotspots with the geostatistical reference data are 0.48 for PCDD/F and 0.43 for TEQ, indicating moderate spatial correspondence in both cases. In Saxony, pronounced discrepancies occur, as RF predicts substantially higher concentrations than the geostatistical estimates. For coldspots, the Jaccard index is 0.33 for PCDD/F, again reflecting moderate agreement. For PCDD/F TEQ values, the value is 0.20, pointing to low concordance, particularly in southern Germany, where modelled and reference data diverge considerably.

Overall, the RF models for PCDD/F and PCDD/F TEQ values provide interpretable predictor relationships and plausible spatial estimates of concentrations. However, due to the low explained variance (pseudo-R² < 20%) and moderate NRMSE values, their robustness must be considered limited.

Figure 4. Spatial distribution of total concentrations of PCDD/F TEQ values (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 4. Spatial distribution of total concentrations of PCDD/F TEQ values (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Hexabromocyclododecane (HBCD). The RF model for estimating the summed HBCD concentration shows an overall moderate model performance (Table 2). With an NRMSE of 0.18, the model exhibits a high level of concordance with the measured values. The pseudo-R² is 30.3%, indicating that roughly one third of the spatial variability can be explained by the selected predictors. Key influencing variables include the density of urban–industrial land uses or population density within a 100 km radius around the moss sampling sites, as well as distance to transportation infrastructure (Table 3). The predictors grouped through fuzzy clustering also make a relevant contribution to explaining the summed HBCD concentrations in mosses.

In comparison with the geostatistical surface estimates, a moderate spatial agreement is observed (Figure 5 and Figure S4). The Spearman’s correlation of 0.66 highlights the clear similarity of the spatial patterns (Table 4). The median of the signed differences is +0.02 ng/g, indicating nearly balanced estimates. The logarithmic deviation of +12.9% falls within the range of moderate systematic differences. Spatially, hotspots with elevated HBCD concentrations above the 90th percentile (0.47 ng/g) can be identified particularly in North Rhine–Westphalia and Rhineland–Palatinate. However, concordance with the geostatistical reference is low, with a Jaccard index of 0.25, and similarly low for coldspots below the 20th percentile (0.16 ng/g). Overall, this results in a plausible spatial estimation of HBCD concentrations, characterised by moderate model quality and moderate agreement with the geostatistical reference map.

Figure 5. Spatial distribution of total concentrations of HBCD (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 5. Spatial distribution of total concentrations of HBCD (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Polybrominated diphenyl ethers (PBDEs). The RF model for BDE-209 achieved the highest goodness-of-fit among all investigated POPs (pseudo-R² = 55.7%, NRMSE = 0.19), indicating the strong explanatory power of the selected predictors (Table 2). In contrast, the model for the summed concentration of the seven PBDEs (excluding BDE-209) shows moderate model performance (pseudo-R² = 24.0%, NRMSE = 0.24). In both cases, the density of mining areas, dumping grounds, and urban–industrial land uses, as well as population density in the near (5 km) and wider surroundings (100–300 km) of the moss sampling sites, are among the key influencing variables (Table 3). Additionally, distances to urban–industrial land uses (seven PBDEs) and to railway lines (BDE-209) are relevant predictors. These findings suggest that both long-range atmospheric transport and local re-emissions substantially contribute to PBDE deposition to terrestrial ecosystems and influence PBDE concentrations in mosses.

Figure 6. Spatial distribution of BDE 209 concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 6. Spatial distribution of BDE 209 concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

The spatial concordance between the RF-based modelling and the geostatistical surface estimates is moderate to good for both PBDE models (Figure 6 and Figure 7; Figures S5 and S6). The Spearman’s correlation is 0.64 for BDE-209 and 0.54 for the seven PBDEs (Table 4). The logarithmically normalised deviation from the geostatistical reference map is substantially higher for BDE-209 (+36.3%) than for the seven PBDEs (+2.7%). The RF modelling identifies hotspots with BDE-209 concentrations above 860.1 pg/g primarily in North Rhine–Westphalia and Rhineland–Palatinate, while for the seven PBDEs the Halle–Leipzig region also emerges as a hotspot area (>120.8 pg/g). The hotspot overlap with the geostatistical reference is moderate, with a Jaccard index of 0.54 for BDE-209 and 0.43 for the 7 PBDE.

Overall, the results indicate a substantially better modelability of BDE-209, which is supported by the higher concordance of its spatial patterns with the geostatistical surface estimates.

Figure 7. Spatial distribution of total concentrations of 7 PBDEs (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 7. Spatial distribution of total concentrations of 7 PBDEs (excl. LOQ) in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Dechlorane Plus (DP). The performance metrics of the model for Dechlorane Plus (syn DP + anti DP; excl. LOQ) indicate a moderate explanatory power (Table 2). The normalized NRMSE is 0.22, placing it within the range of moderate concordance between measured and modelled concentrations. The pseudo-R² of 26.6% shows that roughly one quarter of the variability in DP concentrations can be explained by the selected predictors. The most important influencing variables include proximity to industrial land uses, railway lines, and motorways, as well as the spatial density of industrial land uses within a 100 km radius (Table 3). In addition, total precipitation contributes to explaining the spatial patterns.

The RF model identifies hotspots with DP concentrations above the 90th percentile (242 pg/g), particularly in highly industrialized and traffic intensive metropolitan regions such as North Rhine–Westphalia, Saarland, the Halle–Leipzig area, Hamburg, and the Rhine Neckar metropolitan region (Figure 8 and Figure S7). In contrast, lower concentrations are modelled in rural regions with limited industrial activity—such as Mecklenburg Western Pomerania, Brandenburg, and northern Saxony Anhalt—which is consistent with expectations.

Overall, the RF model provides an initial spatial estimate of DP concentrations with plausible predictor relationships and moderate model quality. A direct comparison with geostatistical surface estimates from the 2020 moss monitoring was not possible for DP, as no corresponding reference maps are available due to absent spatial autocorrelation [15].

Figure 8. Spatial distribution of total concentrations of Dechlorane Plus (syn-DP + anti-DP; excl. LOQ) in mosses in Germany (RF regression estimate).

Figure 8. Spatial distribution of total concentrations of Dechlorane Plus (syn-DP + anti-DP; excl. LOQ) in mosses in Germany (RF regression estimate).

Decabromodiphenylethane (DBDPE). The modelling results for DBDPE indicate a moderate statistical performance (Table 2). With a normalized RMSE of 0.22, the concordance between measured and estimated concentrations falls within the range of moderate model fit. The pseudo-R² of 33.8% shows that just over one third of the spatial variability can be explained by the predictors included in the model. Particularly relevant are the densities of arable land, forest areas, and urban–industrial land uses within a 100 km radius of the sampling sites, as well as distances to railway lines (Table 3). Additionally, the leaf area index contributes to explaining the spatial patterns.

A comparison with the geostatistical surface estimates from the 2020 Moss Survey shows a moderate level of concordance (Figure 9 and Figure S8). The Spearman’s correlation is 0.55, the median of the differences is +186.44 pg/g, and the logarithmic deviation is +14.0% (Table 4). This indicates a slight overestimation by the model relative to the geostatistical reference. Spatially, hotspots with DBDPE concentrations above the 90th percentile (2256 pg/g) are identified primarily in North Rhine Westphalia, Rhineland Palatinate, and the Rhine Neckar metropolitan region. Coldspots with substantially lower concentrations occur in Brandenburg and Bavaria. The hotspot overlap is high, with a Jaccard index of 0.67, whereas concordance for coldspots remains low (0.21). Overall, the RF model for DBDPE provides a plausible spatial estimate supported by reasonable predictor relationships and a moderate level of agreement with the geostatistical comparison data.

Figure 9. Spatial distribution of DBDPE concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 9. Spatial distribution of DBDPE concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

2,3 Dibromopropyl 2,4,6 tribromophenyl ether (DPTE). For DPTE, the RF modelling shows an overall low statistical explanatory power. The normalized RMSE of 0.31 indicates weak agreement between measured and predicted concentrations. With a pseudo-R² of 16.1%, less than one fifth of the spatial variability is explained. The most influential predictors are precipitation amounts, the density of urban–industrial land uses in the immediate surroundings, and agricultural land uses in the wider surroundings.

The model identifies elevated DPTE concentration above the 90th percentile (231 pg/g) primarily in parts of southern Germany, in North Rhine–Westphalia, and in Saarland (Figure 10 and Figure S9). In contrast, coldspots with substantially lower concentrations occur in Saxony Anhalt, northern Bavaria, and along the Baltic Sea coast.

Thus, the RF model provides an initial cartographic approximation of the spatial distribution of DPTE in Germany. A comparison with geostatistical surface estimates from the 2020 Moss Survey was not possible, as no reference maps are available for this contaminant due to absent spatial autocorrelation [15].

Figure 10. Spatial distribution of DPTE concentrations in mosses in Germany (RF regression estimate).

Figure 10. Spatial distribution of DPTE concentrations in mosses in Germany (RF regression estimate).

Pentabromotoluene (PBT). The RF modelling results for PBT indicate moderate statistical performance (Table 2). The normalized RMSE of 0.26 suggests a moderate level of concordance between measured and modelled concentrations. With a pseudo-R² of 30.9%, under one third of the spatial variability is explained by the selected predictors. Particularly relevant predictors include the density of agricultural and urban–industrial land uses, population density in the wider surroundings, and the fuzzy clusters “intensively used agricultural landscapes” and “agricultural–urban–industrial mixed-use areas” (Table 3).

The spatial analysis identifies high PBT concentrations above the 90th percentile (51.3 pg/g) particularly in northern Germany, Bavaria, Saarland, and the Berlin metropolitan area. The comparison with the geostatistical surface estimates from the 2020 Moss Survey shows an overall moderate level of spatial concordance (Figure 11 and Figure S10). The Spearman’s correlation is 0.55, the median of the differences is +2.1 pg/g, and the logarithmic deviation is +4.9%, indicating a slight overestimation relative to the geostatistical reference (Table 4). While the overlap of hotspot regions is low, with a Jaccard index of 0.11, a moderate agreement is observed for the coldspots (Jaccard index: 0.48).

Overall, the RF model for PBT provides a methodologically sound, though still uncertain, spatial approximation, supported by plausible predictor relationships and a moderately strong model performance. The interpretability of the results is reinforced by moderate concordance with the geostatistical comparison data.

Figure 11. Spatial distribution of PBT concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 11. Spatial distribution of PBT concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Hexabromobenzene (HBBz). For HBBz, the RF modelling shows only a low statistical goodness of fit (Table 2). The NRMSE of 0.25 indicates limited concordance between measured and modelled concentrations, while the pseudo-R² of 16.9% suggests that only a small fraction of the spatial variability can be explained by the selected predictors. Key predictors are the density of agricultural land within a 5 km radius and proximity to built-up areas (Table 3). Additionally, industrial land uses within a 100 km radius, the leaf area index, and orographic elevation contribute to the model.

The spatial distribution shows hotspots with HBBz concentrations above the 90th percentile (7.62 pg/g) exclusively in Bavaria. In contrast, coldspots with substantially lower concentrations occur in North Rhine–Westphalia and Rhineland–Palatinate. The comparison with the geostatistical surface estimates from the 2020 moss monitoring indicates a moderate level of concordance (Figure 12 and Figure S11). The Spearman’s correlation is 0.54, the median of the differences is −0.67 pg/g, and the logarithmic deviation is −12.7%, suggesting a moderate underestimation relative to the geostatistical reference (Table 4). The overlap of hotspot regions is moderate, with a Jaccard index of 0.33, whereas concordance for coldspots is low (Jaccard index: 0.27).

Overall, the regression estimate provides an initial predictive map of HBBz concentrations in mosses with plausible influencing factors. The moderate agreement with the geostatistical comparison data supports the plausibility of the modelling results.

Figure 12. Spatial distribution of HBBz concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

Figure 12. Spatial distribution of HBBz concentrations in mosses in Germany: RF regression estimate vs. geostatistical surface estimate.

4. Discussion

The modelling of POP concentrations in mosses conducted for the first time within the European Moss Survey reveals a heterogeneous picture regarding model fit and spatial concordance; PAH, BDE 209, and DBDPE provide robust spatial estimates with moderate-to-high model performance and a clear agreement with geostatistical reference maps. In contrast, explanatory power remains limited for PCDD/F, TEQ values, DPTE, and HBBz, while HBCD, PBDE, DP, and PBT fall into an intermediate range with moderate model quality and plausible predictor relationships. This highlights that model performance based on broad coverage spatial data is substantially lower than when using corresponding site-specific information from the 2020 Moss Monitoring, where ten out of eleven RF models for POPs achieved a pseudo-R² of at least 20%. Where explanatory power is particularly low, this is also reflected in the regression maps, which show pronounced regional discrepancies compared with the geostatistical surface estimates, especially for PCDD/F TEQ and HBBz.

The analysis of predictor variables shows that atmospheric deposition, the density of different land-use types, population density, and proximity to industrial areas and transport infrastructure represent relatively important factors influencing POP concentrations in mosses. In addition, precipitation and vegetation structure contribute to the explanatory power of the models. This confirms—consistent with earlier studies—that identified atmospheric deposition and the density of urban–industrial and agricultural land uses within a range of 100–300 km as key long-range predictors, and population density as well as the density of extraction and dump sites within <5 km, together with distances to residential areas and transport infrastructure, as key short-range predictors [14]—that both large scale processes and structures as well as local emissions sources play a decisive role in shaping the spatial distribution of observed concentrations. The application of Fuzzy C Means clustering to the predictors improved model performance only in a few cases. An exception is the cluster representing agricultural and urban–industrial land uses within large radii around the moss sampling sites, suggesting that broad scale groupings of land-use densities are more influential than fine scale patterns. A key limitation is that deposition data were unavailable for roughly three quarters of the investigated substance groups. Since atmospheric deposition constitutes a major input pathway for POPs, its predictive relevance could only be assessed to a limited extent. Consequently, the interpretive power of the models remains restricted, particularly regarding the distinction between long-range transport and local sources.

The spatial generalization confirms industrialized regions such as North Rhine–Westphalia, the Rhine Neckar area, Halle/Leipzig, and Saarland as cross compound hotspots, whereas rural and forest rich regions predominantly exhibit lower concentrations in mosses. The predictive mapping should thereby not be interpreted as an exact representation of the concentration distribution but rather as an indicative tool primarily suited for analysing spatial patterns and relative differences. The aim was to identify broad scale trends and hotspots rather than to determine precise POP concentrations in mosses.

With the exception of DP, and DPTE, a direct comparison with geostatistical reference maps was possible. Although this comparison does not constitute a validation in the strict sense, it allows—through a spatial pattern comparison—an assessment of the spatial concordance between the RF modelling and an independent methodological approach, thereby contributing to the evaluation of its reliability and plausibility. Compared with the geostatistical references, the models with available comparison data for PAH, PCDD/F, TEQ values, HBCD, seven PBDEs, BDE 209, DBDPE, PBT, and HBBz show moderate-to-good concordance. However, the partly low overlap of hotspots and, in particular, coldspots indicates regional uncertainties. This means that while areas of higher contamination are captured more plausibly, the prediction of regions with low POP concentrations in moss is less reliable. Regional deviations—such as the systematic overestimation of PCDD/F TEQ values in Saxony—further illustrate the limitations of the RF modelling and the need for cautious interpretation of the predictive maps.

Beyond these aspects, the RF-based spatial predictions are subject to several sources of uncertainty. First, the relatively small number of sampling sites leads to data sparsity, which restricts the model’s ability to represent the full environmental variability across Germany. Second, the spatial predictions inevitably involve extrapolation into regions with environmental conditions not well represented in the training data, which may contribute to regional deviations in areas insufficiently represented in the training data. Third, correlations among predictor variables may reduce the stability of variable importance estimates and complicate the attribution of effects to individual predictors. These sources of uncertainty influence both the reliability of local predictions and the applicability of the maps for fine-scale assessments. Consequently, the spatial outputs should be interpreted as broad spatial tendencies rather than precise concentration estimates.

5. Conclusions

Overall, this study demonstrates that mosses, in combination with statistical modelling, represent a suitable tool for capturing spatial concentration patterns of persistent organic pollutants. The results show that robust hotspot regions can be identified, although model performance and agreement with geostatistical reference data vary across POP groups. To further improve model performance, the inclusion of additional predictors—particularly deposition data—is recommended. Moreover, the integration of RF models with geostatistical methods should be explored, as this is expected to increase the accuracy of concentration estimates. A well-established approach for this purpose is regression kriging, which combines the explanatory power of RF models with variogram based analysis of spatial autocorrelation and subsequent regionalization of residuals [37,38].

Future work should particularly focus on increasing sample sizes and validating the results with independent measurement data in order to represent spatial contamination patterns of persistent organic pollutants more reliably. Calculated substance group specific minimum sample sizes for POPs, defined to ensure a specified error tolerance in the calculation of arithmetic means, provide a useful benchmark for this purpose [14,39]. Finally, a stronger linkage between the modelled spatial patterns and results from other monitoring programmes appears advisable. The integrative inclusion of data from the Environmental Specimen Bank (ESB, [40]) or the Forest Soil Condition Survey (Bodenzustandserhebung im Wald–BZE II, [41]) could further improve validation and enable broader comparability.