Soil moisture observations and machine learning reveal preferential flow mechanisms in the Qilian Mountains

Abstract

The complexity of the spatial distribution and temporal occurrence of preferential flow (PF) makes it challenging to understand the mechanisms of PF. This study aims to identify the spatial and temporal patterns of PF occurrence using machine learning (Classification and Regression Trees and Random Forests) in the Qilian Mountains, Northwest China. Our results show that detected PF events transport much more rainfall down to the subsoil than non-PF events. Different vegetation types exhibit variations in the main soil layers where PF occurs, which is closely related to the distribution of roots. The PF proportion varies significantly both vertically and horizontally. Based on the Random Forests, we found that the spatial distribution of the PF proportion is mainly controlled by the saturated hydraulic conductivity and residual soil moisture, which cannot be identified by conventional correlation analysis methods. With these soil properties, the spatial distribution of the PF proportion can be estimated with reasonable performance. Using the Classification and Regression Trees method, we identified the temporal occurrence pattern of the PF for different vegetation types and all observation stations. Results indicate that the dominant factors controlling the temporal occurrence of the PF varied for different vegetation types. The thresholds at which these factors initiate the PF also varied. Finally, we found that the PF occurs particularly under wet conditions (except for hydrophobic soils), under denser vegetation, and under conditions of high rainfall amount and intensity, regardless of vegetation type. Our study confirms that both site factors (e.g., soil properties and vegetation) and temporal factors (e.g., initial soil moisture and rainfall characteristics) control the occurrence of the PF in mountainous regions such as the Qilian Mountains and that the Classification and Regression Trees has great potential to study the temporal occurrence of the PF.

Keywords

• Preferential Flow
• Machine learning
• Occurrence pattern
• Mountainous areas