Attention maps of computed as absolute SHAP values (a) and (d) show median attention maps across the test period; (b) and (e) show the difference between attention maps during periods with high nitrate concentrations; (c) and (f) show the difference between attention maps during the periods associated with high levels of hydrological connectivity.Abstract
Catchment wetness and related hydrological connectivity determine nitrate mobilization, transport and transformation. However, quantifying connectivity at this scale remains challenging. In this study, we tackle this challenge by developing a multi-branch Deep Learning framework trained on streamflow observations and SMAP-Hydroblocks, a unique satellite-based soil moisture data set at 30 m resolution (aggregated here to 1 km) over the US. We hypothesize that spatial patterns of soil moisture are indicative of the relationship between hydrological connectivity and nitrate transport at the catchment scale, and can be a use to predict nitrate concentrations in the streams. We tested this hypothesis using the developed framework to simulate daily nitrate concentration simultaneously at the outlets of eight catchments with varying land cover and concentration-discharge (C-Q) patterns. Our model satisfactorily represents nitrate dynamics in the study catchments with a median 0.62 Nash-Sutcliffe Efficiency. Spatial patterns of soil moisture on average account for 32% of model feature importance and model performance deteriorates on average by 14% if these patterns are not considered. Leveraging explainable AI (XAI) we confirm that model decisions align with known physical processes across catchments with contrasting C-Q behavior. Attention maps identified near-stream hotspots as the regions with the highest predictive power for nitrate export suggesting a potential link between the soil moisture patterns and catchment-scale hydrological connectivity. This proof-of-concept study demonstrates the potential of XAI combined with high-resolution remote sensing products for improving nitrate predictions and mapping critical areas for nitrate export..
