A spatially explicit interpretable machine-learning method to track dissolved inorganic nitrogen pollution in a coastal watershed

Abstract

Sustainable water management requires an in-depth understanding of dissolved inorganic nitrogen (DIN) in watersheds. However, it remains challenging in the context of intensifying human-induced and naturally impacted stresses in a changing environment. A spatially explicit interpretable machine-learning method was proposed based on the annual variation in land use/cover, Catchment Land Surface Model, Noah land surface model, population dynamics, and nightlight in a coastal Chinese watershed. The proposed model was effective for delineating DIN processes and sources in watersheds. The d (index of agreement), R2 (coefficient of determination), |PBIAS| (percent bias), and KGE (Kling-Gupta efficiency)during the training period were 0.91–0.98, 0.76–0.96, 0.38–3.12, and 0.73–0.87, respectively. The d, R2, |PBIAS|, and KGE during the testing period were 0.85–0.94, 0.56–0.79, 3.59–6.18, and 0.65–0.86, respectively. NH4+-N in the watersheds may be strongly related to the effects of urbanization within the watersheds while the agricultural activities may modify the patterns of NO3–-N in the watersheds. The NH4+-N was highly related to the urbanization effects, which contributed 20 %–30 % of the riverine NH4+-N in the North River. In contrast, agricultural activities may modify patterns of NO3–-N in watersheds, and agricultural activities in the West River contributed 50 %–70 % riverine NO3–-N in the watershed. Moreover, urbanization may alter the water content, soil properties, and regional climate patterns within the watersheds, while repeated DIN input through agricultural activities in the agricultural watersheds may change the nitrogen processes in the soil and groundwater. Paired with agricultural activities, dams or reservoirs may amplify NH4+-N to watersheds with modifying the advection and diffusion of DIN from agricultural activities, and increasing sediments in the watersheds. This study demonstrated the potentials of the proposed method in tracking DIN pollution and provided new insights into nutrient management in the watersheds under intensifying human-nature interactions.

Keywords

• Land use
• Nitrogen
• Spatially explicit
• Machine-learning
• Jiulong River watershed