Estimating dissolved organic carbon in Lake Mälaren and Lake Erken, Sweden: A comparative study of spatial transferability and model generalisability using XGBoost and Sentinel-2 data

Sammanfattning

Accurate monitoring of dissolved organic carbon (DOC) in lakes is essential for understanding aquatic carbon dynamics, assessing lake ecosystem health, and supporting water resource management. While remote sensing offers scalable alternatives to in-situ monitoring, machine learning models often struggle to generalise across systems. This thesis investigates the potential of the eXtreme Gradient Boosting (XGBoost) machine learning algorithm for DOC concentrations in two lakes in central eastern Sweden: Lake Mälaren and Erken. The XGBoost model integrates Sentinel-2 surface reflectance imagery, environmental variables from ERA5-Land reanalysis data, geographic coordinates and in-situ measurements. A locally trained XGBoost model for Lake Mälaren yielded accurate DOC predictions, characterised by low error metrics and consistent performance, suggesting high model reliability under varied feature combinations. The SHapley Additive exPlanations (SHAP) analysis identified latitude and catchment runoff as the primary predictive variables, whereas spectral reflectance features contributed the least. To evaluate spatial transferability and model generalisation, two approaches were applied: (1) the Cross-Lake Generalisation Model, which was trained on Lake Mälaren and independently tested on Lake Erken, and (2) the Lake-to-Lake Transferred Model, which used Lake Erken data for training and testing but utilised the hyperparameter configuration derived from the Lake Mälaren. While the Cross-Lake Generalisation Model showed underfitting and poor generalisation, the Lake to-Lake approach offered more stable, though still limited, predictive accuracy. These results highlight that while model architecture can transfer, successful application across lakes depends on local data structure, input alignment, and ecological context. SHAP proved essential in interpreting model logic and assessing generalisation. Overall, the findings demonstrate that while XGBoost offers strong predictive performance in well-characterised lakes, its broader applicability remains constrained by differences in ecological context, data distribution, and variable relevance. The SHAP analysis offered valuable insight into system-specific predictor relevance and shifts in feature importance, enhancing interpretability. The results underscore the need for expanded predictor testing and the development of more transferable modelling frameworks to support scalable DOC monitoring across diverse freshwater systems.