A grid-based interpretable machine learning method to understand the spatial relationships between watershed properties and water quality

Abstract

Understanding the spatial relationship between watershed properties and water quality is essential for watershed management. However, it remains challenging to identify such a relationship because of its nonlinearity. To understand the temporal and spatial effects of the driving of water quality, a grid-based interpretable machine learning approach was developed to explore the relationship between water quality (i.e. nitrogen, phosphorus, and chemical oxygen demand) and watershed properties in the Minjiang River Watershed (MRW). A grid-based model was developed with 1837 input features for streams and reservoirs, respectively, based on land use, population, fertilizer, and Noah land surface model in the MRW. Compared to the water quality in the stream, the water quality in the reservoirs may be more sensitive to the change environmental settings. The reservoirs may amplify the effects of climate variability on water quality. The soil moisture may modify the water quality, especially the dryness of the top soil may control the pollutant that entry into the watershed. The effects of the urbanization may modify the distributions of important feature that control the regional water quality. The point source pollution per inhabitant may be reduced with increasing urbanization. This study provides an in-depth understanding of the relationship between water quality and watershed properties, temporally and spatially.

Keywords

• Driving factors
• Water quality
• Watershed
• Interpretable machine learning