Lough Feeagh and Sau Reservoir are located in western Ireland (53°56^′ N, 9°35^′ W) and northeastern Spain (41°58^′ N, 2°23^′ E), respectively (Fig. 1). The study sites have contrasting attributes. Feeagh (depth of 46.8 m and area of 3.95 km²) is a monomictic and oligotrophic lake surrounded by a relatively undisturbed landscape, while Sau (depth of 70 m and area of 5.8 km²) is an eutrophic system subjected to human activities and water abstraction. Feeagh has two primary inflows, the Black and the Glenamong rivers, while Sau has one, the Ter river. The catchment of Feeagh is relatively small (84 km²), with mid-range hills, and dominated by peatland. The catchment of Sau, in contrast, is larger (1525 km²), with a varying topography and land uses (Figs. 1, A1).

The dynamics of DOM in both study sites have been previously explored in which identified that natural dynamics related to soil temperature, river discharge and drought were important drivers in Feeagh, and in , which showed human activities (e.g., wastewater effluents and agricultural runoff) were important for Sau, in addition to its environmental dynamics. Catchment hydrology is key for carbon transport into both study sites, and contributes to a distinct seasonality related to climate. Feeagh has a wet temperate climate, with cooler air temperatures and higher rainfall levels that occur on more than 75 % of days in the year. The variability of carbon inputs reflects the sensitivity of a peatland-dominated landscape, which exacerbates climate-induced carbon release from the catchment into the lake. In contrast, Sau has a Mediterranean climate, characterised by hot, dry summers and mild, wet winters, dictating water availability, thermal stratification, and organic matter fluxes in the reservoir.

A five-step workflow was implemented to predict fDOM at each site (Fig. 2). First, all potential drivers for fDOM were collected as input data for each site. Second, an exploratory data analysis of fDOM was implemented and the ML models were trained using 85 % of the available time series data (1978 out of 2328 fDOM measurements for Feeagh, and 653 out of 769 for Sau), while the remaining 15 % (350 out of 2328 for Feeagh, and 116 out of 769 for Sau) future time series (hold-out period) was reserved for independent testing to evaluate performance and potential overfitting by comparing with the training period.

Third, a set of drivers was selected for each site based on the variable importance extracted from the ML models, retaining only those drivers that exceeded an importance threshold of 5 %. SHAP and partial dependence analyses were applied to these specific drivers to evaluate how fDOM predictions at each site varied as a function of individually changing the selected input variables. Fourth, we ran simulations using (i) the drivers with a higher variable importance (> 5 %), and (ii) using only drivers extracted from globally accessible reanalysis data, and assessed model.

The same workflow was applied to both sites using the same data sources, allowing for comparison. Following the FAIR principles, all data and workflow scripts are available and fully reproducible here: 10.5281/zenodo.19354955 . Large language models were used in this study to optimise the codes, improve the final plots and, for basic proofreading of the text.

Supervised ML models have advantages and limitations for time series prediction. In addition to capturing non-linear relationships typical in aquatic systems and water quality predictions , they provide flexibility to assess multiple drivers, temporal indicators, and variables external to the system . ML models do not require a fixed set of drivers to predict fDOM effectively, unlike process-based models, which typically rely on predefined inputs. Additionally, there is no need for parameter calibration but hyperparameter tuning, simplifying the modelling process. While some may argue that the lack of parameterization suggests a “black box” approach, supervised ML can provide insights into the potential drivers for predicting a target variable .

However, due to the intrinsic autocorrelation in time series, e.g., when predicting DOM, these models tend to overfit when using out-of-bag samples during training. To overcome this issue, robust validation and training are required. Here, we used a hold-out period for validation during testing at both study sites. Prior to selecting this method, we compared it with two alternative validation methods using random forest: (1) 5-fold cross-validation and (2) rolling window cross-validation with a two-year training period, a one-year testing period, and a window shift every 90 days (see Fig. A7).

Supervised machine-learning models were applied to predict fDOM dynamics, including Random Forest (RF) , eXtreme Gradient Boosting (XGBoost) , Light Gradient Boosting (LGB) , CatBoost (CTB) , k-Nearest Neighbors (KNN) , Support Vector Regression (SVR) and linear model. The three gradient-boosting frameworks were included to assess the robustness of results across closely related implementations. Given their similarities, the main results shown are focused on the highest-performing model among the three. Neural networks were not included because they typically require more complex architecture design and calibration, which can increase methodological complexity and may reduce reproducibility of the workflow. In contrast, the selected models allow more reproducible implementation due to their comparatively simpler hyperparameter tuning.

To assess the influence of the most important drivers, we included SHapley Additive exPlanations (SHAP) plots, using the best boosting method. The SHAP plots measures how much a single driver (feature) value contributes to moving the prediction away from the average value, the y-axis has the input drivers ranked by overall importance (from top to bottom), x-axis has the SHAP value representing the impact on model output for a single prediction, each point is a single data instance and the color reflects the driver value (blue = low, red = high). In addition, in the Appendix, partial dependence plots were generated to support and illustrate the individual influence of each driver on fDOM predictions by varying the driver's values across its entire range while keeping all other drivers constant in their average value. These partial dependence plots were generated using the outputs from the Random Forest model. To implement SHAP analysis and partial dependence plots, the shap package in Python and the pdp package in R were used, respectively.

The most influential predictors of fDOM at each site were identified from all 24 potential drivers based on the 5 % threshold of the variable importance extracted from the CTB model. This resulted in seven influential drivers being identified for Feeagh and five for Sau (Fig. 3), four of which were common to both sites. This selection was consistent across multiple machine-learning models (Fig. A8).

The variables for the deepest soil layer were relevant for both study sites. Soil temperature and soil moisture at 100–255 cm, were the most influential drivers for Feeagh. Similarly, for Sau, soil moisture and temperature at similar depths were the most influential drivers. Another key driver that was shared between Feeagh and Sau was Julian day. Lake volume was only important for Sau, while solar radiation, the amount of carbon entering the water body (indicated by the DOC inflow concentration) and both soil moisture and temperature at 28–100 cm were only influential for Feeagh.

Figure 4 introduces SHAP beeswarm plots for the most influential drivers selected in Fig. 3, enabling the assessment of the individual effect of each driver on fDOM predictions.

Seasonal patterns in fDOM concentrations were observed in both Feeagh and Sau, with higher values in winter and lower in summer, as reflected in the influence of Julian day. However, the key predictors and their effects differed substantially, shaped by contrasting catchment and climate characteristics. In Feeagh, where precipitation is relatively high and sustained year-round, deep soil temperature (100–255 cm) was the dominant and potentially limiting predictor, with fDOM increasing with temperature up to a threshold, beyond which a drop in the water table may counteract the effect (see partial dependance plots; Figs. A9 and A10). In addition, Feeagh showed minimal influence of mid-depth soil temperature (28–100 cm) and solar radiation on fDOM. This temperature-fDOM relationship was less relevant in Sau, where deep soil temperature (100–255 cm) had less explanatory power. Instead, fDOM dynamics in Sau, were driven primarily by soil moisture at both 28–100 and 100–255 cm depths, potentially depicting a limiting condition by water stress.

Water availability also shaped the role of other predictors differently across the two sites. For instance, surface water temperature in Feeagh showed a clear threshold behaviour, with fDOM increasing relatively linearly beyond 6.5 °C and stabilizing around 7.5 °C, while in Sau, water volume acted as a surrogate for fDOM production (see partial dependance plots; Figs. A9 and A10). Lower volumes corresponded to reduced fDOM values, which increase and stabilize beyond 146 hm³. Interestingly, inflow DOC concentration influences Feeagh more than Sau, likely due to differences in hydroclimatological processes governing these relationships. These differences highlight how catchment water availability fundamentally alters the relative importance and behavior of fDOM drivers.

Table 1 presents a comparative evaluation of the seven ML and statistical models employed in this study. In Lough Feeagh, CatBoost (CTB) demonstrated the best overall performance with the highest R2 (0.50) and KGE (0.68), and a relatively low RMSE (8.17 QSU). Similarly, in Sau Reservoir, CTB again has the highest R2 (0.54) and KGE (0.66), and one of the lowest RMSE (7.11 QSU). While some models like RF and LM showed moderate performance at Feeagh, others such as SVR and XGB performed poorly in Sau, with SVR even having a negative KGE (-0.82), suggesting significant model bias. Overall, CatBoost (CTB) consistently outperformed all other models across both sites, supporting its selection for fDOM prediction using the selected environmental drivers.

The results during the training phase (Table A2) confirm that most ML models, especially XGB, RF, and KNN, had a high performance during training. However, the performance dropped in some models (e.g., XGB at Sau Reservoir) during the testing phases, underscoring the importance of evaluating models on independent test data to assess generalisability and overfitting. CatBoost (CTB), again, presented a more stable performance, showing slightly lower training metrics (especially in Feeagh) compared with the other ML models and better generalisation when comparing with the metrics during testing. Feeagh exhibited more consistent model performance, with less overfitting between training and testing phases, across the different validation methods (Fig. A7), compared to Sau. This difference is likely attributable to the limited amount of available data at Sau. In addition, we assess whether this overfitting could be related to the moderate skewness of the fDOM data in Sau, however similar results and performance metrics were found when applying log transformation (to remove skewness) before applying the ML models (see Fig. A11 and Tables A3 and A4).

The most influential drivers for fDOM identified for each site suggest potential carbon-producing processes that drive DOM dynamics in each lake. Interestingly, the most influential drivers were related to temperature and moisture at the deepest soil layers, indicating a common relevance for carbon inputs from terrestrial processes in both catchments. The Irish site is dominated by peat, a highly organic soil known to be sensitive to temperature-induced carbon release, especially as temperature levels rise. This is supported by the relation of soil temperatures with fDOM at this site . In contrast, water availability constraints are much more pronounced in drier soils such as those of the Spanish site, as is validated by the partial dependence relationship of soil moisture with fDOM . A plausible explanation is that, in Sau, increased soil moisture could boost biological activity, enhancing fDOM production, while in Feeagh, soil moisture showed a bell-shaped relationship, suggesting a more nuanced interplay between oxygen availability and microbial processes (Figs. 4, A9 and A10). In addition, higher soil moisture can reflect stronger soil–lake hydrological connectivity, particularly during wet periods and flushing events, facilitating the transport of terrestrially derived DOM into the lake.

DOM dynamics are driven by physical and biogeochemical processes in the soil that are sensitive to changes in temperature and moisture, e.g., microbial processes that break down organic matter . The fact that the deepest layers of the soil were more important in the model for both sites than those shallower could be linked to potential carbon attenuation processes, such as soil organic matter decomposition and retention in the soil, including sorption processes . In any case, the production of carbon in the catchment that eventually ends up in the lake requires concurrent downstream transport, governed by rainfall events.

Climate and topography (see Fig. A1) dictate the flushing of accumulated DOM during rainfall events, but also can influence sustained baseflow DOM contributions. In Feeagh, carbon exports from the catchment have been observed regularly throughout the entire annual cycle, with a seasonal variability . In contrast, DOM in Sau accumulates primarily during the summer and is mainly flushed out via surface runoff during the wetter winter months . These patterns are supported by the relationship between inflow DOC concentration and fDOM at both sites (Figs. 4, A9 and A10), which shows a slight increase in predicted fDOM under lower carbon input conditions. Thus, inflow DOC concentration could reflect discharge pulses and dilution effects driven by precipitation , following the characteristic seasonality of each site.

Seasonality plays a crucial role in fDOM predictions, as evidenced by the relationship of the Julian day driver with DOM dynamics at both sites (Fig. 4). At the Irish site, DOM seasonality is primarily shaped by natural environmental processes, whereas in the Spanish site, human influence plays a much greater role. This distinction helps explain why surface lake water temperature and solar radiation, two variables typically linked to strong seasonal patterns, were important only for Feeagh, while reservoir volume was significant only for Sau (Figs. 3, 4, A9 and A10). Volume and soil variables produce a similar effect on fDOM as Julian day at Sau, given that higher volumes closer to the winter season can lead to higher fDOM values. Incorporating Julian day into the workflow offers a simple yet effective way to represent seasonality, potentially replacing seasonal variables (e.g., air temperature) (see the correlation matrix of all drivers in Fig. A12). This proves that the use of machine learning approaches opens up opportunities to assess diverse drivers under contrasting conditions.

Improving the accuracy of DOM predictions in lakes can enhance efforts to reduce further water quality deterioration and support lake management. This study demonstrated a feasible approach for simulating daily fDOM in two contrasting lakes, especially when using the Catboost model given its good generalisability. The performance metrics (Fig. 5) obtained at each site for the different model simulations lie in a similar or better range than comparable studies that modelled carbon dynamics in lakes e.g.,. It is important to be aware, however, that previous studies are based on different frameworks. These include variations in the machine learning algorithms used, the target variable for quantifying carbon dynamics, with most studies having focused on DOC, whereas fDOM is the target variable here, as well as differences in input driver data and site-specific conditions. While both fDOM and DOC are widely used indicators of dissolved organic matter in lakes, they differ in measurement principles and in the fractions of organic matter they represent. Both variables, however, remain ecologically relevant for understanding DOM sources, transport, and transformation processes, which aligns with the context of this work.

Our results suggest high potential for scalability, as predictive performance remained consistent across different driver sets even for two contrasting study sites, and performed good using only reanalysis data that is globally available and Julian day (Fig. 5). Importantly, this consistency remained even when specific highly-influential drivers were removed from the driver set. For instance, in Sau, where human intervention makes future reservoir outflows difficult to predict, avoiding reliance on water volume as a driver proved advantageous, as its removal from the set of drivers maintained model performance, despite being identified as an influential variable. It is likely that soil moisture at the deepest layer, a variable that showed a behavior similar to that of volume (Fig. 4), may have contributed to maintain the predictive performance when volume is removed from the driver set. In addition, for Feeagh, the predictive capacity was also maintained when using only meteorological and soil drivers. This demonstrates that a large driver dataset, such as the 24-variable set used in this study, would not be necessary to produce an accurate prediction when modelling fDOM using supervised machine learning.

Our approach offers the opportunity to validate and deploy a workflow capable of delivering daily DOM predictions in both undisturbed and anthropized sites, even when only limited data on input drivers are available, while at the same time providing insights into the dominant drivers. However, a site-specific model validation, including identification of appropriate training and testing periods, hyperparameter tuning for each specific study site and assessment of overfitting is essential. In terms of driver attribution, it is of note that the relationships identified using machine learning may not always be related at a process level . In our case, however, many of these same drivers had already been identified for river DOC levels in the Feeagh catchment e.g.,. While the workflow can be easily replicated, fDOM data or data for another proxy for DOM are required. It is of note that such proxies of DOM are increasingly being incorporated into water quality monitoring programmes, an aspect that is convenient for testing workflows such as the one described .

The workflow presented here is not recommended for climate change studies, as the drivers of DOM variability can significantly change under entirely new and unrecorded climatic conditions. Consequently, supervised machine learning may fail to capture the signal from the time series. Moreover, the method's reliance on historical patterns limits its ability to extrapolate beyond the range of observed environmental conditions . Future research could expand the application of this framework to a broader range of lakes, integrate additional drivers such as remote sensing-derived terrestrial and aquatic quantity and quality parameters .