关键词: 氨氮, 空间插值, 机器学习, 随机森林, SHAP

Abstract:  Ammonia nitrogen is one of the main inorganic pollutants in groundwater, which mainly comes from agricultural, industrialy and domestic pollution. Excessive ammonia nitrogen will endanger human health. Temporal and spatial distribution of ammonia nitrogen is affected by factors such as meteorology, hydrology, hydrogeology, and land use type, so groundwater ammonia nitrogen analysis based on limited sampling points will generate great uncertainty. In this study, firstly, the Songhua River basin in the Sanjiang Plain was taken as an example, and soil organic matter mass fraction, soil total nitrogen mass fraction, soil cation exchange capacity (CEC), soil pH value, groundwater depth, thickness of clay layer in vadose zone and land use type were selected as potential influencing factors, a machine learning model for fitting ammonia nitrogen concentration was established. Secondly, significant influencing factors were identified using the shapley additive explanations (SHAP) method of interpreting machine learning models. Finally, a machine learning prediction model was established according to the significant influencing factors, and the data of groundwater ammonia nitrogen in the study area was interpolated. And the temporal and spatial variation of ammonia nitrogen was analyzed. The results showed that groundwater depth, land use type, CEC and soil organic matter mass fraction were the main influencing factors of groundwater ammonia nitrogen in this area. The area of groundwater ammonia nitrogen in the Ⅰ-Ⅲ water quality level showed an increasing trend. The proportion of area increased from 31% to 87%. And the area of Ⅳ-Ⅴ water quality showed a decreasing trend. The proportion of area decreased from 69% to 13%. The overall water quality was improved from 2011 to 2018.

Key words: ammonia, spatial interpolation, machine learning, random forest, SHAP