Mosses ML: Machine-Learning Enhanced Biomonitoring of Emerging Contaminants Using Hylocomium splendens: An Integrated Approach Linking Atmospheric Deposition, Trace Metals and Predictive Risk Assessment

Monitoring emerging contaminants in the atmosphere remains a central challenge for environmental toxicology, public health and evidence-based risk assessment. Rapid industrialization, urban expansion and energy production have intensified emissions of toxic trace elements such as cadmium (Cd), lead (Pb), nickel (Ni), chromium (Cr) and zinc (Zn), all of which pose ecological and human-health risks due to their persistence, bioavailability and capacity for long-range atmospheric transport [5,16]. Traditional instrumental approaches for assessing airborne contaminants—including air filters and precipitation collectors—provide high analytical precision but are costly, labor-intensive and often unable to capture fine-scale spatial variability [2]. These limitations have led to the widespread adoption of terrestrial mosses as efficient biomonitors of atmospheric deposition within international initiatives such as UNECE ICP Vegetation [22]. Bryophytes, particularly pleurocarpous mosses such as Hylocomium splendens and Pleurozium schreberi, lack a root system and obtain nutrients exclusively from atmospheric inputs, making them sensitive integrators of both wet and dry deposition [4,19]. Their high cation-exchange capacity, surface roughness and morphological complexity facilitate the interception and retention of airborne particulate matter, including poorly soluble metal-bearing particles [11]. Active biomonitoring using transplanted mosses further reduces environmental variability and allows controlled exposure periods, improving comparability across sites and time [1,9]. Recent studies from Central Europe—including Upper Silesia—have shown pronounced spatial gradients in moss chemistry and elevated concentrations of industrially derived trace elements, especially in areas influenced by mining, metallurgical and coal-based activities [13]. Despite these strengths, conventional statistical approaches used in moss biomonitoring (ANOVA, PCA, correlation matrices) lack predictive capability and often fail to capture nonlinear interactions among deposition pathways (bulk vs dry), emission sources and biological accumulation processes. As emerging contaminants increasingly exhibit complex environmental behaviour, new analytical frameworks are needed to improve detection sensitivity, interpret high-dimensional datasets and support high-resolution environmental decision-making. Machine learning (ML) offers clear advantages in this context: algorithms such as Random Forests and Gradient Boosting can uncover hidden structure within environmental datasets, quantify feature importance and predict contamination risk with substantially greater accuracy than classical methods [7,10]. However, ML has not yet been systematically integrated into moss biomonitoring, and no standardized framework exists for combining bioaccumulation data with atmospheric deposition metrics to support predictive risk assessment. To address this gap, we developed Mosses ML, a comprehensive ML-enhanced workflow that integrates (i) trace-element concentrations in Hylocomium splendens transplants, (ii) bulk and dry deposition measurements, (iii) relative accumulation factors (RAF) and (iv) site-level environmental metadata (industrial, urban, rural). This framework builds on validated experimental datasets from Upper Silesia—one of the most heavily industrialized regions in Central Europe—where dry deposition has repeatedly been shown to dominate particulate-bound metal fluxes. To our knowledge, this is the first study to integrate explainable machine-learning methods (including SHAP analysis) with moss biomonitoring to quantitatively disentangle the relative contributions of dry and bulk deposition to trace-element accumulation. By combining RAF metrics, deposition measurements, multivariate ordination and predictive modelling, the Mosses ML framework provides a mechanistic and data-driven evaluation of atmospheric contamination that extends beyond traditional descriptive biomonitoring approaches. This methodological integration offers a generalizable template for assessing emerging contaminants using bryophyte-based monitoring systems. By incorporating machine learning, Mosses ML aims to:

• predict trace-metal concentrations in mosses from local environmental parameters;

• identify the most influential drivers of contamination using feature-importance and SHAP interpretability;

• classify sampling sites into contamination-risk categories with improved accuracy;

• provide a scalable tool applicable to other bryophyte species and to emerging contaminant groups beyond metals.

This study presents the development, validation and environmental application of the Mosses ML framework. By uniting established biomonitoring methods with modern computational tools, our approach enhances the interpretative and predictive power of moss-based assessments and strengthens their role as early-warning indicators within atmospheric pollution monitoring and regulatory policy.

Dry deposition was quantified using glass plates (80 mm diameter) coated uniformly with pharmaceutical-grade white petrolatum. The plates were pre-weighed (±0.2 mg) after gentle heating (42 °C) to standardize the adhesive layer. Five plates were deployed per site for the full 90-day period. Particles adhering to the plates were removed with hexane, filtered onto ash-free cellulose filters, dried, combusted, and digested for elemental determination.

Descriptive statistics for metal concentrations in moss transplants are shown in Table 1. Metal concentrations in Hylocomium splendens transplants showed a clear spatial gradient consistent with the pollution regime of Upper Silesia, with industrial sites exhibiting the highest enrichment of Pb, Cd, Zn, Fe, Ni and Cr (Figure 2). Rural sites displayed the lowest levels, while urban sites showed intermediate contamination. These patterns reflect the well-known influence of mining, smelting and fossil-fuel combustion in the region [8,15]. The strong accumulation of Pb, Cd, Zn and Fe aligns with previous evidence that H. splendens efficiently captures particulate-bound metals due to its high surface area and cation-exchange properties [4,19]. Relative accumulation factors (RAF) calculated after 90 days confirmed this gradient. RAF values for all metals and site categories are summarized in Table 2. Industrial locations showed the highest enrichment for nearly all elements, particularly Pb, Cd, Zn and Fe—metals known to form strong bonds with moss exchange sites and therefore retained efficiently [6]. The consistently negative or negligible RAF values for Mn reproduced a well-documented pattern of leaching and active washout from moss tissues [3]. These RAF dynamics collectively indicate that enriched atmospheric deposition, especially from particulate sources, is the dominant mechanism driving metal uptake in transplanted mosses. The comparison between atmospheric deposition and moss chemistry revealed complementary roles of bulk and dry deposition pathways. Bulk deposition showed moderate correlations with moss concentrations (Table 3), particularly for more soluble metals such as Cd, Ni and Zn, supporting earlier observations that wet deposition can contribute to moss uptake under certain conditions [1,5]. However, dry deposition (Table 4) exerted a consistently stronger influence across all site categories. Dust collected from deposition plates contained substantially higher concentrations of Pb, Zn, Cr, Fe and Cu than bulk precipitation, confirming that particulate emissions dominate metal transport in this region [20]. Correspondingly, the strongest moss–deposition correlations were observed for dust, encompassing Cd, Cr, Cu, Mn, Ni, Pb and Zn at industrial sites and multiple metals at urban and rural sites. These results support the established view that H. splendens primarily accumulates metals through passive interception of dry and coarse particulate deposition [11,18,21]. Multivariate analyses further clarified these gradients. The PCA of standardized moss-metal concentrations showed strong separation among rural, urban and industrial clusters (Figure 3). Together, PC1 and PC2 explained ~86% of total variance, with Pb, Cd, Zn, Fe and Cr loading strongly on PC1. The biplot confirmed co-occurrence of metals associated with metallurgical and mining emissions, reinforcing the interpretation that spatial patterns in moss chemistry reflect regional industrial sources. Rural sites grouped tightly with low metal concentrations, while industrial sites aligned with high positive PC1 scores, indicating strong particulate exposure. Machine-learning models provided an additional layer of quantitative insight. Predictive performance metrics for all ML models are summarized in Table 5. Random Forest regression achieved high predictive performance for key metals—R² values reached ~0.90 for Pb, Cd and Zn—demonstrating that moss metal concentrations can be accurately reconstructed from deposition metrics and site characteristics (Figure 4). Feature ranking under the Random Forest model (Figure 5 and Table 6) indicates that dry deposition load and co-occurring metals were the dominant predictors.Predicted-versus-observed plots showed points clustering closely around the 1:1 line, and low RMSE values further confirmed robust model behaviour. The strong performance of ensemble algorithms is consistent with their effectiveness in capturing complex nonlinear relationships in environmental datasets [7,10]. Explainable ML analyses revealed mechanistic drivers underlying these predictions. SHAP global importance rankings identified dry deposition mass and the concentrations of co-occurring metals (notably Fe, Zn, Cu and Ni) as the most influential predictors of Pb accumulation, surpassing bulk deposition and categorical site descriptors (Figure 6). This aligns with chemical evidence that industrial particulates share mixed metal compositions and are deposited simultaneously onto moss surfaces. SHAP dependence plots for Fe-Pb and Zn-Pb further demonstrated nonlinear response curves with saturation zones, revealing interactions that would remain undetected in classical parametric models. The explainability suite therefore strengthened the mechanistic interpretation that particulate transport is the principal vector governing Pb and related metal uptake in H. splendens. Taken together, these results show that transplanted H. splendens reliably integrates signals of both coarse and fine particulate pollution across heterogeneous landscapes. Dry deposition emerged as the predominant determinant of moss chemistry, while bulk deposition played a secondary role for more soluble elements. The ML-enhanced analytical framework—combining deposition measurements, bioaccumulation data, multivariate analysis and model interpretability—offers a powerful extension to established ICP Vegetation protocols, allowing both predictive assessment and mechanistic insight. Such integrative approaches can improve spatial resolution in pollution monitoring, strengthen environmental risk assessments and be readily extended to other moss species, pollutant classes and multi-year biomonitoring programs. This study has several limitations. First, the relatively small number of sampling sites (n = 15) restricts the complexity of the ML models and may limit generalizability to other regions. Second, the dataset contains only short-term (90-day) exposure measurements; multi-season or multi-year campaigns would improve robustness of predictive modelling. Third, metal uptake by mosses reflects both depositional fluxes and species-specific physiological processes, meaning that environmental drivers unrelated to atmospheric inputs (e.g., humidity, microhabitat) could introduce noise. Finally, the ML models were trained on a single regional dataset without external validation, and future work should incorporate cross-regional or cross-species transferability tests to confirm the scalability of the Mosses ML framework.

This study demonstrates that integrating machine-learning techniques with traditional moss biomonitoring yields a significantly enhanced analytical framework for assessing atmospheric contamination. Hylocomium splendens transplants proved to be highly effective bioindicators of particulate-bound metals, consistently reflecting the strong industrial pollution gradient observed across the study area. The dominance of dry deposition as a predictor of metal uptake confirms that mosses act primarily as collectors of airborne particulates [18,21], a finding further supported by high RAF values, PCA separation, and SHAP-based mechanistic insights. The Mosses ML workflow achieved high predictive accuracy for key contaminants, surpassing the diagnostic capacity of classical statistical approaches. Its improved risk classification and transparent model interpretability address long-standing challenges in biomonitoring, particularly in complex industrial landscapes where pollutant sources co-occur. By combining biological indicators, deposition measurements and advanced computational tools, this framework enhances environmental surveillance and provides a scalable, transferable model for future biomonitoring initiatives, including multi-year monitoring, cross-regional comparisons, and assessments of emerging contaminants.