Climatic, physiographic, and anthropogenic factors controlling spatial and temporal variability of water balance within the Budyko framework

Abstract:

Quantification of the partition of water balance input into output is essential to understanding water availability. Within a water balance model known as the Budyko framework, water availability is controlled by precipitation, potential evapotranspiration, and other factors represented by a single parameter ω. The ω can represent climatic factors such as climate seasonality and snow fraction, physiographic factors such as vegetation productivity and topography, and anthropogenic factors such as agricultural drainage, groundwater withdrawal, and urbanization. Previous studies have analyzed the controlling factors of the parameter over space. However, most of them focused on a specific type of watersheds. It remains unclear if and how the spatial representation of the parameter can vary based on watershed characteristics. In addition, a non-stationary pattern of streamflow may be caused by different factors over time; however, few studies investigated the temporal variability of the controls of water availability. In this study, temporal and spatial variability of the parameter in a Budyko-type Fu’s equation is modeled in four sets of watersheds, including reference, agricultural, regulated, and urban watersheds in the contiguous United States. Multiple linear regression and random forest models are used to quantify variable importance. The geographically weighted regression method is used to capture the spatial non-stationarity of representation of ω. The purpose of this analysis is to: 1) determine temporal and spatial controls of water balance; 2) identify the most important impact factor on water availability other than precipitation and potential evapotranspiration; 3) compare the strengths of conventional statistical method and machine learning models within the Budyko framework. It is hypothesized that controls of water balance in reference watersheds are climate and physiographic factors, while in non-reference watersheds, primary controls depend on watershed characteristics. Random forest is expected to outperform multiple linear regression due to its effectiveness in capturing data non-linear data structures.