Detection of Watershed Non-stationarity from Process-Based Model Evaluation using Approximate Bayesian Computation

Abstract:

A key assumption in virtually all modeling studies in surface hydrology is that the watershed properties (dominant flow-paths, soil and vegetation properties, etc.) are time-invariant, and hence can be estimated from a sufficiently long rainfall-discharge data record. However, this “stationarity” assumption has been criticized in the recent literature. Human interference and climate change substantially affect land-use type and distribution, vegetation and soil properties (among others), and thus render the runoff response to rainfall non-stationary. In this presentation I will introduce a statistical methodology that will help to assess catchment non-stationarity. Our approach has its roots in diagnostic model evaluation, and uses approximate Bayesian computation (ABC) and Markov chain Monte Carlo simulation with DREAM_(ABC) to detect gradual/abrupt changes in the watersheds response to rainfall. Non-stationarity of catchment behavior is not easily visible in the rainfall-runoff response, yet manifests itself if summary metrics of the discharge data are computed for different time periods. Models that assume stationarity are unable to accurately simulate these gradual/abrupt shifts in catchment behavior, and consequently, diagnostic inference with ABC cannot produce a “behavioral” (posterior) model space. Our preliminary studies demonstrate that the presented methodology is able to successfully differentiate between watersheds that are classified as stationary and those that have undergone significant changes in land-use and urbanization and thus are deemed non-stationary.