An interpretable Earth observation framework for estimating lake dissolved organic carbon in Sweden using Sentinel-2 and ERA5-Land data

Inland-water dissolved organic carbon (DOC) is a key regulator of aquatic carbon cycling and water quality, yet it remains difficult to observe at the spatial and temporal scales targeted by Earth observation (EO). This study developed an interpretable machine learning (ML) framework to estimate lake DOC and interpret model behavior across Sweden using Sentinel-2 MultiSpectral Instrument (MSI) reflectance (RS), ERA5-Land environmental variables (E), and geographic features (G). Using 1,595 spatiotemporal match-ups from 349 lake locations during the period 2017–2024, this study benchmarked five regression algorithms across four feature combinations. XGBoost with the full E + RS + G configuration achieved the best estimation performance (RMSE ≈ 2.46 mg C L −1 ; Bias ≈ −1.6 %) and enabled 10 m quarterly DOC estimation and mapping for examining spatiotemporal changes in Swedish lakes. Robustness was further evaluated using a held-out test set, spatially separated validation, interpolation, and coordinate-only baselines to assess potential spatial confounding. SHapley Additive exPlanations (SHAP) analysis indicated that DOC predictability in the Sweden-wide dataset is strongly spatially structured: geographic features capture broad-scale spatial context, while vegetation, hydroclimatic variables, and visible-to-near-infrared reflectance provide more directly interpretable process-related signals. Overall, the framework provides an operational EO–ML workflow for DOC estimation in Sweden, with explicit robustness tests to assess the respective roles of auxiliary spatial context and EO–environmental predictors.