Time‐Varying Sensitivity Analysis Reveals Relationships Between Watershed Climate and Variations in Annual Parameter Importance in Regions With Strong Interannual Variability

Abstract

As weather and climate extremes continue to set new records (WMO, 2017), it is becoming increasingly important to study the effects of these extremes on hydrologic cycles (Bates et al., 2008) as well as our ability to simulate these changes within hydrologic models (Nijssen et al., 2001; Xu, 1999). In the United States, Cal-Abstract Climate change impacts on hydroclimatology are becoming increasingly apparent around the world. It is unknown how annual variations in precipitation and air temperature alter the modelinferred importance of hydrological processes and how this varies across watersheds. To examine this, we used parsimonious rainfall-runoff model and applied time-varying sensitivity analysis across 30 Californian watersheds for a 33-year period (1981-2014). We calculated annual total order sensitivity indices for five performance metrics: Kling-Gupta Efficiency (KGE) and model error in simulating hydrologic signatures (runoff ratio, slope of flow duration curve, baseflow index, and timing of streamflow centroid). Sensitivity of hydrological signatures to the parameters differed by signature, while parameter importance with respect to KGE was much more spatially and temporally variable. Variations in parameter sensitivity with respect to KGE were either correlated with air temperature (snow-dominated sites in the Sierra Nevada) or precipitation (lower elevation sites < 1,300 m). Across error metrics and signatures, parameter sensitivity strongly differed between wet and dry years for a subset of our study sites. While parameter importance varied through time, parameter sensitivity variations across watersheds were much more pronounced. This suggests that parameter controls on model performance are much more a reflection of watershed properties as opposed to being dominantly shaped by shifts in precipitation and air temperature. These findings emphasize the importance of understanding simulated watershed responses to fluctuating annual conditions, as this conceptual knowledge is necessary to anticipate similarities and differences in response across even relatively proximal watersheds in the face of growing extreme conditions. Plain Language Summary Computer models are useful tools that can be used to analyze what inputs matter most to ensuring that model output is similar to observed conditions. In order to understand how very wet years versus very dry years impact which model inputs (called parameters) are most important to matching good predictions, we simulated streamflow using a simple mathematical model many thousands of times. We used these outputs to determine how variations in predictions of streamflow (output from the model) are related to variations in model parameters, to identify which parameters are most important for ensuring reasonable model predictions. We focused on watersheds that are close together, but that may experience different average conditions and interannual variability. We found that which parameters are important varied across different watersheds and also varied with total a...

Identifying Dominant Processes in Time and Space: Time‐Varying Spatial Sensitivity Analysis for a Grid‐Based Nitrate Model

“…In wet years however, the dominance gradually shifted to dynamics of the gravitational water drainage given the increased relative sensitivity of fInterf and fPerco (Figures 6 and 7). Similar increased sensitivity in deeper layers and aquifers with greater annual precipitation was also reported in Basijokaite and Kelleher (2021). Such a shift in dominance emphasizes that monitoring and modeling should be cognizant of changing wetness conditions, and highlights the generic importance of long‐term datasets in water quality models.…”

“…However, one should recognize the uncertainty in SA, since insights into catchment functioning are only possible if it is conducted based on adequate information and methodological setting (Gupta & Razavi, 2018). In terms of SSA, similarly, different settings can result in contrasting conclusions (Baroni et al, 2017); for example, the parameters characterizing deep aquifers have been reported to be more influential in low-flow conditions for discharge simulation (Guse et al, 2016;Massmann & Holzmann, 2012;Pfannerstill et al, 2015), while Basijokaite and Kelleher (2021) and our study found increased sensitivity during wet periods instead. Here we take DMC as an exemplar to briefly introduce how these key factors/methodological settings altered the results during SSA implementation.…”

“…Such time‐dependent nature of parameter sensitivity has been recognized already ∼50 years ago (McCuen, 1973). Consequently, many studies have applied SA in different time periods to explore the temporal dynamics of parameter sensitivity in lumped (Basijokaite & Kelleher, 2021; Ghasemizade et al., 2017; Herman et al., 2013) or distributed models (Guse et al., 2014; Pfannerstill et al., 2015; Reusser et al., 2011). For example, Basijokaite and Kelleher (2021) found that parameter controls on model performance were shaped by a shift in precipitation and air temperature.…”

“…Consequently, many studies have applied SA in different time periods to explore the temporal dynamics of parameter sensitivity in lumped (Basijokaite & Kelleher, 2021; Ghasemizade et al., 2017; Herman et al., 2013) or distributed models (Guse et al., 2014; Pfannerstill et al., 2015; Reusser et al., 2011). For example, Basijokaite and Kelleher (2021) found that parameter controls on model performance were shaped by a shift in precipitation and air temperature. However, relatively few studies have incorporated analysis of time‐variance of parameter sensitivity into a grid‐based SSA (Herman et al., 2013a), because it further increases computational demand.…”

Analysis of the Relative Importance of the Main Hydrological Processes at Different Temporal Scales in Watersheds of South-Central Chile

“…Analysis of the temporality of the most important processes that influence various climate scenarios has become a focus of study for many researchers. For example, Diop et al [9] investigated the long-term streamflow trends at three time scales (monthly, seasonal, and annual) in the upper Senegal River basin, Howden et al [10] presented a method to detect changes in the mean and variance of hydrological variables and explore the hydrological processes involved in the non-seasonal behavior of time series, and Basijokaite and Kelleher [11] analyzed the relationship between the most important processes in a watershed and their seasonal and annual behavior. In addition, various studies have demonstrated that changes in streamflow time series can be attributed to climate [12][13][14][15] and/or anthropogenic factors [16,17].…”