Nutrient concentration modelling in the Baltic countries using spatial machine learning

The water quality of surface streams is impacted by various environmental (soil texture, precipitation, and local topography) and anthropogenic (fertilizer and manure deposition) factors of the upstream catchment. Knowledge of relationships between the water quality and the catchment-wide characteristics is of high importance for outlining critical areas for interventions, e.g., nature-based solutions for nutrient capture. 

Various modelling techniques have been implemented to gain insights into the catchment characteristics and the corresponding nutrient concentrations.  The use of machine learning methods for this purpose has increased due to the relaxed requirements of the input data as well as increasingly ubiquitous spatial environmental datasets. However, machine learning models are not spatially-aware by default. Recently, various methods have been proposed to account for spatial dependency across multiple modelling stages. 

We employ the Random Forest supervised machine learning algorithm to model nutrient (nitrogen and phosphorous) concentrations on a point scale. We use national level monitoring data of the Baltic countries between the years 2017-2023 with varying number of observations per site, averaged over the study years. Environmental characteristics (topography, land use, climate, soil properties) describing the corresponding upstream catchment area are used as the explanatory features. As the catchments extend beyond the borders of the Baltics, we use various global datasets for feature creation (e.g., ERA5-Land, SoilGrids). In addition, we apply spatial machine learning methods and assess their applicability for catchment-based modelling. Lastly, we employ explainable AI methods, namely SHapley Additive exPlanations and Partial dependency plots, to validate if our model’s revealed relationships correspond to the domain knowledge.