Surface-water sampling was conducted on two campaigns: October 2019 (n = 8) and June 2021 (n = 12), with sites uniformly distributed across the lake (Figure 1). An additional channel site was included to characterize inflow from the long-water region. Samples were collected at 0.30 m depth; at very shallow locations, surface grabs were taken. Immediately after collection, samples were placed in clean containers, stored on ice, transported under refrigeration, and held at 4 °C in the laboratory until analysis according to accredited holding-time procedures.

All determinations were performed at the Soil and Water Testing Laboratory, Department of Physical and Environmental Geography, University of Szeged (accredited laboratory NAH-1-1437/2018) under standard operating procedures. Methods were as follows: total suspended solids (TSS)—gravimetric determination after filtration through 0.45 µm pre-weighed membrane filters and drying to constant mass; total water hardness (reported as CaOeq)—EDTA titration with Eriochrome Black T, results expressed as mg L^-1 CaOeq; chemical oxygen demand (COD)—permanganate oxidation with 10 min boiling and oxalic-acid back-titration; ammonium (NH₄ ⁺)—flow-injection spectrophotometry (FOSS FIAstar 5000); arsenic (As)—ICP-OES after 0.45 µm filtration (PerkinElmer Optima 7000 DV).

All measurements were performed in triplicate with analytical precision ≤10% (laboratory records). Sampling, preservation, and analysis followed accredited SOPs, including refrigerated handling, analysis within prescribed holding times, and metals measured on filtered aliquots. Instrument performance was verified through the laboratory’s QA/QC program (calibration standards, blanks, and continuing checks). Detection limits are reported where relevant (e.g., NH₄ ⁺ < 0.25 mg L^-1 in 2021).

Google Earth Engine (GEE), specifically the “COPERNICUS/S2-SR-HARMONIZED” dataset, was utilized to identify Sentinel-2 images covering Vadkert Lake during the in situ measurements conducted in 2019 and 2021. The images were filtered to include only those with less than 20% cloud cover. The closest images to the in situ campaigns were acquired on 13 August 2019, and 9 May 2021, corresponding to the 34TCS tile and a relative orbit 36. Additionally, four images representing different seasons in 2024 were selected: January 29, May 13, July 17, and November 09. All scenes were used in Level-2A surface reflectance (SR) form produced by the ESA Sen2Cor atmospheric-correction processor. We used the Scene Classification Layer (SCL) to remove pixels flagged as cloud shadow, cloud (high and medium probability), cirrus, and snow/ice. We selected L2A/Sen2Cor—rather than per-scene empirical corrections such as QUAC—to ensure sensor-consistent, physically based SR across dates and to leverage the integrated SCL for masking. The modified normalized difference water index (MNDWI), developed by Xu (2006) (Equation 1), was applied with a threshold of 0 to mask water pixels, and the lake area was subsequently clipped. Predictor bands for retrieval were B2–B8a (10–20 m). Bands B1, B9, B10, and B12 were excluded due to lower spatial resolution and/or limited sensitivity to the targeted WQPs. Band B11 (SWIR, 1610 nm) was used only for MNDWI (Equation 1) and was not included in the regression models. M N D W I = R B 3 − R B 11 R B 3 + R B 11 (1) where MNDWI: modified normalized difference water index; R_B3​: surface reflectance in the green band (B3, ≈560 nm); R_B11​: surface reflectance in the shortwave-infrared band (B11, ≈1610 nm).

A window of 2 × 2 pixels was placed around the sampling sites (8 sites in 2019 and 12 sites in 2021), and the reflectance values of bands (B2 to B8a) were extracted and averaged. The correlation between band reflectance and measured WQPs (i.e., COD, CaOeq, TSS, As, and NH₄ ⁺H was assessed through the Pearson correlation coefficient. Regression models were subsequently developed for parameters with strong correlations, employing the stepwise multilinear regression technique. These models were applied to historical images from 2024 to investigate the spatial and temporal dynamics of the selected water quality parameters. The analysis was conducted using a combination of Sentinel Application Platform (SNAP, version 8.0), Python 3.0, and Quantum GIS (QGIS 3.34). Our choice of WQPs therefore spans parameters with direct optical coupling (TSS, COD) and indirect, management-relevant indicators (NH₄ ⁺, As, CaOeq), reflecting both the monitoring priorities for Hungarian small lakes and the environmental pressures pertinent to Lake Vadkert.

We retained Sentinel-2 Level-2A bands B2–B8a as predictors and used B11 only for water masking via MNDWI (Equation 1). Around each in-situ sample (2019: n = 8; 2021: n = 12), we extracted mean surface reflectance from a 2 × 2-pixel window and fitted stepwise multiple linear regressions for parameters that showed strong band–WQP relationships. The following lake-specific empirical models Equations 2–4 were used to generate maps:

For As (µg L^-1): As = 55.497 + 0.056   R B 5 − 0.116 R B 4 + 0.059 R B 3   R 2 = 0.99 (2)

For NH₄ ⁺ (mg L^-1): NH 4 + = − 0.219 + 0.001 R B 5 − 0.002 R B 4 + 0.001 R B 3   R 2 = 0.98 (3)

For COD (mg L^-1): COD = 5.21 + 0.010 R B 5 − 0.009 R B 4   R 2 = 0.91 (4)

We explored models for CaOeq and TSS; however, these did not generalize reliably (weak to moderate correlations, limited variability and range-coded values for TSS).

A comparative analysis of five water quality parameters—COD, CaOeq, TSS, As, and NH₄ ⁺—between 2019 and 2021 reveals notable temporal variations that reflect changes in the lake’s hydrological and biogeochemical conditions. Varying precipitation patterns, surface and subsurface inflows, and seasonal dynamics likely influence the observed differences. This section analyzes the specific trends and implications associated with each parameter, highlighting potential sources of pollution and shifts in water quality status over the 2 years (Table 1).

In 2019, the COD values ranged between 9.00 and 11.20 mg/L, indicating relatively elevated levels of organic matter. In contrast, the COD values in 2021 significantly decreased, ranging from 5.8 to 6.6 mg/L. This decline suggests an overall improvement in water quality and a reduction in organic pollution, potentially due to changes in land use, wastewater input, or hydrological conditions. Similarly, total hardness (CaOeq) values showed a notable reduction from 196 to 216 mg/L in 2019 to 129–171 mg/L in 2021, except for one outlier (290 mg/L in 2021). This reduction could be attributed to dilution effects from increased precipitation or surface inflow, supporting the hypothesis of wetter conditions in 2021.

The TSS values were generally lower in 2021 compared to 2019. While 2019 values ranged between 21.0 and 29.0 mg/L, most 2021 samples were below 21 mg/L, with only a few reaching 31 mg/L. This decrease indicates reduced erosion, sediment resuspension, or runoff, possibly due to calmer hydrodynamic conditions. NH₄ ⁺ concentrations further support this pattern: values in 2019 ranged from 0.66 to 0.78 mg/L, while all 2021 samples were below the detection limit (<0.25 mg/L). This dramatic reduction implies lower levels of nutrient pollution or improved nitrification, both of which are signs of better ecological status.

Concentrations showed a marked decline between the 2 years. In 2019, As ranged from 108.2 to 111.4 μg/L, but it dropped to 39.1–83.5 μg/L in 2021. The lower As levels suggest a potential change in geochemical dynamics, input sources, or mobilization processes, possibly influenced by hydrological conditions or seasonal shifts.

Overall, the comparison reveals a consistent trend of improving water quality from 2019 to 2021. This improvement may be driven by climatic factors (e.g., increased rainfall and dilution), reduced anthropogenic input, or internal lake processes stabilizing over time.

The correlation analysis demonstrated a very strong relationship between band reflectance and As, NH₄ ⁺, and COD, a moderate correlation with CaOeq, and a weak correlation with TSS (Table 2). Focusing on the correlation among the WQPs, As, NH₄, and COD were the most correlated parameters, showing the highest correlation magnitude (r = 0.962–0.994), while CaOeq and TSS were less correlated with these parameters, and they showed the highest correlation with NH₄ ⁺ (r = 0.585) and COD (r = −0.218), respectively. In terms of band reflectance, typically, B2 and B4 showed negative correlations with WQPs, while the other bands showed positive correlations. COD showed its highest correlation with B7, CaOeq with B2, TSS with B4, and both As and NH₄ ⁺ with B5.

Regression models were developed for the most correlated WQPs with Sentinel-2 bands—namely, As, NH₄ ⁺, and COD. The models demonstrated promising accuracy, with R2 values ranging from 0.91 (COD) to 0.99 (As). The analysis revealed that specific Sentinel-2 bands, particularly B4 (red) and B5 (red edge), were the most influential contributors to the three WQPs. Besides, B3 (green) showed a notable contribution to the models for As and NH₄ ⁺.

The three models used for mapping (As, NH₄ ⁺, and COD) are reported in materials and methods (Equations 2–4) and are applied to the seasonal Sentinel-2 scenes described below.

The spatiotemporal distribution of the three WQPs was depicted across seasons by applying the developed models to Sentinel-2 images acquired on specific dates in 2024 (Figure 2). The maps in Figure 2 were generated using the retrieval equations listed in materials and methods (Equations 2–4). The seasonal variability of the parameters was evident, especially when comparing summer with the other seasons. The As concentration slightly increased from 64.7 ± 9.1 μg/L in winter to 66.3 ± 10.2 μg/L in spring, then surged by 15% (76.8 ± 20.7 μg/L) in summer, before recording the lowest concentration in autumn (63.9 ± 8.3 μg/L). Notably, the spatial distribution of As was almost uniform across seasons (Figures 2A,D,G,J). A similar temporal trend was noticed for NH₄ ⁺ since it slightly increased from 0.001 ± 0.0 mg/L in winter to 0.008 ± 0.0 mg/L, reached its peak in summer (0.2 ± 0.3 mg/L), and declined again in autumn to return to winter’s mean value of 0.001 ± 0.0 mg/L. However, its spatial distribution was non-uniform in winter and autumn, with relatively higher concentrations observed in the southern and eastern peripheries of the lake (Figures 2B,K). In contrast, a uniform distribution occurred in spring and summer (Figures 2E,H). The COD declined slightly from 5.6 ± 0.67 mg/L in winter to record its lowest concentration in spring (5.5 ± 0.48 mg/L). Then, it peaked in summer (7.3 ± 2.01 mg/L) and recorded a relatively high concentration in autumn (5.75 ± 0.79 mg/L). The spatial distribution of COD mirrored that of NH₄ ⁺, with higher concentrations observed in the southern and eastern lake peripheries (Figures 2C,L), and a more uniform distribution in spring and summer (Figures 2F,I).

Integrating Sentinel-2 MSI surface reflectance with in-situ WQPs via regression yielded spatially explicit, multi-date maps that enable continuous lake monitoring, with high predictive accuracy (R ² = 0.91–0.99) across the study period. The seasonal patterns align with established process controls: wind-driven resuspension and shoreline disturbance episodically elevate TSS, while organic loading and in-lake production/decay shape COD, particularly in warm months. Although As and NH₄ ⁺ are not directly optically active, they display strong associations with the red/red-edge bands (B5, B7), indicating indirect coupling through co-variation with optically active constituents and shoreline/inflow influences. These dynamics are consistent with observations from other systems (e.g., Mohsen et al., 2022; Mallick et al., 2014).

The spatiotemporal variability mapped for 2024 highlights localized impacts from both natural and anthropogenic activities. Elevated As, NH₄ ⁺, and COD in summer likely reflect increased evaporation, intensified microbial activity, and heightened agricultural/urban pressures around peak season aligned with Li et al. (2019). These maps enable managers to prioritize sampling, deploy early-warning screening after heatwaves or low-flow periods, and target mitigation at shoreline and inflow hotspots.

In our study, summer increases in NH₄ ⁺ and COD and the concentration of shoreline/inflow hotspots are consistent with climate-driven mechanisms that intensify water-quality stressors: warmer conditions and lake heatwaves strengthen stratification, extend residence time, and elevate oxygen demand, fostering ammonification and organic-matter processing that raise NH₄ ⁺ and COD, while low water levels heighten resuspension along shallow margins. These pathways align with evidence of widespread lake deoxygenation under recent warming (Jane et al., 2021) and with projections that lake heatwaves are becoming more frequent and severe, shifting biochemical regimes at sub-seasonal scales (Woolway et al., 2021a; Woolway, Anderson and Albergel, 2021b). Within this context, our multi-date Sentinel-2 workflow functions as a repeatable early-warning screen, capturing within-season shifts—including post-heatwave rises in NH₄ ⁺/COD—and guiding adaptive sampling toward emergent hotspots in small, shallow alkaline lakes like Vadkert.

For management context, we benchmarked our observations against typical EU Water Framework Directive (WFD) “good status” ranges: As ≤ 10 μg L^-1, NH₄ ⁺ ≈ ≤ 0.5 mg L^-1 (type-specific), COD ≈ ≤ 25 mg L^-1 (noting method alignment, COD_Mn vs. COD_Cr, and TSS typically < 25–50 mg L^-1; CaOeq is not directly regulated under WFD and is used contextually as alkalinity/hardness (commonly ∼20–200 mg L^-1 as CaCO₃ equivalents) (ERA-COMM EU Water Law training; Austrian Federal Ministry—BMLUK, WFD guidance).

Our lake-specific stepwise multilinear regressions on Sentinel-2 L2A reflectance achieved high predictive skill for As, NH₄ ⁺, and COD in a small, shallow alkaline lake—matching or exceeding accuracies often reported for band-ratio or machine-learning schemes in comparable settings (Ciężkowski et al., 2022; Kokal et al., 2024). The linear framework remains transparent and reproducible, with performance confirmed by k-fold cross-validation and by-year (withheld-year) tests. Because As and NH₄ ⁺ are inferred indirectly via co-variation with optically active constituents, model validity is context-dependent and can weaken under domain shift (e.g., atypical hydrology or loading). As with MSI data generally, multicollinearity among red/red-edge bands and limited sample size can inflate apparent fit; our diagnostics mitigate but cannot eliminate this risk. Potential saturation at high concentrations and adjacency/atmospheric residuals near shorelines also warrant caution; using L2A/Sen2Cor and explicit QA/QC reduces, but does not remove, these effects (cf. Jaelani and Ratnaningsih, 2018). Overall, the results indicate that well-calibrated linear baselines can deliver strong, interpretable performance for management use, while future work should broaden seasonal/hydrologic coverage, test non-linear learners against these baselines under rigorous external validation, and incorporate pH/temperature to contextualize ammonia risk. Future work should quantify high-stage groundwater–lake exchange using Hungary-proven geophysical levee-diagnostic techniques to test seepage-driven inputs to Lake Vadkert (Sheishah et al., 2022; 2023; Abdelsamei et al., 2024).

Our findings align with recent demonstrations of Sentinel-2 capability for inland-water WQPs. Studies such as Toming et al. (2016) and Xu et al. (2019) reported accurate retrievals of chlorophyll-a and turbidity, and our R ² = 0.91–0.99 for As, NH₄ ⁺, COD indicate that—when locally calibrated—even indirect indicators can be mapped credibly. In line with Kowe et al. (2023), who achieved strong regressions for optically active parameters (OAPs) in Lake Manyame (TSM R ² = 0.90; turbidity R ² = 0.95) and a moderate fit for the non-optically active TN (R ² = 0.63), our Sentinel-2–based, locally calibrated models for a small, shallow alkaline lake similarly deliver high performance (As, NH₄ ⁺, COD: R ² = 0.91–0.99) despite these targets being largely non-optically active. This contrast underscores two points: (i) Sentinel-2 consistently retrieves OAPs well across systems, and (ii) with careful local calibration and validation, indirect red/red-edge relationships can also support credible mapping of selected non-OAPs. Where our TSS correlations were weak—likely due to range-coded field values—Kowe’s strong TSM/turbidity results highlight the importance of rigorous in-situ characterization for particulate metrics. Together, the studies indicate complementary strengths across reservoir (Manyame) and small-lake (Vadkert) settings, while emphasizing the need for site-specific models and quality-controlled field data. We extend OAP-focused work by showing that locally calibrated red/red-edge proxies can map selected non-OAPs (As, NH₄ ⁺, COD) in a small alkaline lake. We also note the practical advantage of Sentinel-2 over coarser platforms such as Landsat-8 for small, shallow lakes like Vadkert due to its finer spatial resolution.

Despite TSS being optically active, we observed low correlations with Sentinel-2 bands, likely due to in-situ methodological constraints (range-type TSS values at some sites rather than precise measurements). For CaOeq, correlations were moderate (strongest single-band with B2), and candidate multi-band fits failed withheld-year checks; consequently, we did not produce lake-wide retrievals for CaOeq or TSS.