Towards validation of the Canadian precipitation analysis (CaPA) for hydrologic modeling applications in the Canadian Prairies

Highlights

• Change point analysis shows that CaPA corresponds with actual observations (CDCD).

• CaPA has same spatial pattern, dependency and autocorrelation properties as CDCD.

• Hydrologic forcing in model shows that CaPA is reliable for hydrologic modeling.

Abstract

This study presents a three–step validation technique to compare the performance of the Canadian Precipitation Analysis (CaPA) product relative to actual observation as a hydrologic forcing in regional watershed simulation. CaPA is an interpolated (6 h or 24 h accumulation) reanalysis precipitation product in near real time covering all of North America. The analysis procedure involves point-to-point (P2P) and map-to-map (M2M) comparisons, followed by proxy validation using an operational version of the WATFLOOD™ hydrologic model from 2002 to 2005 in the Lake Winnipeg Basin (LWB), Canada. The P2P technique using a Bayesian change point analysis shows that CaPA corresponds with actual observations (Canadian daily climate data, CDCD), on both an annual and seasonal basis. CaPA has the same spatial pattern, dependency and autocorrelation properties as CDCD pixel by pixel (M2M). When used as hydrologic forcing in WATFLOOD™, results indicate that CaPA is a reliable product for water resource modeling and predictions, but that the quality of CaPA data varies annually and seasonally, as does the quality of observations. CaPA proved most beneficial as a hydrologic forcing during winter seasons where observation quality is the lowest. Reanalysis products, such as CaPA, can be a reliable option in sparse network areas, and is beneficial for regional governments when the cost of new weather stations is prohibitive.

Introduction

In many regions of the world, particularly mid-to-high-latitude regions, weather station networks are sparse, making precipitation analysis a challenge and generating significant errors for hydrologic modeling applications (Nicotina et al., 2008). For example, in Canada, there are few stations in the north where, ironically, much of the major hydroelectric and water resource infrastructure is located. The major challenge in hydrologic modeling large watersheds, such as those in Canada, is that weather station density is higher in the south near the American border, and decreases drastically moving northward. Similarly, Coulibaly et al. (2013) conducted a study on spatial analysis of the Canadian National Hydrometric Network (CNHN) based on the World Meteorological Organization (WMO) guidelines for hydrometric network density. Their results show that approximately 12% of the terrestrial area of Canada is covered by hydrometric networks. In other words, this is the only area that meets the WMO 2008 standards. Also, 49% of the area is poorly gauged, while the remaining 39% is ungauged. This shows that most the terrestrial area of Canada did not meet the WMO standard, and as such, precipitation patterns may not be fully represented in modeled flows from these areas. Providing an accurate representation of amount and the spatial variability in precipitation to a hydrologic model is therefore crucial for accurate hydrologic prediction.

For practical applications, interpolation of precipitation estimates from weather stations is relied on to provide values for areas and across regions that are ungauged. Traditionally spatial interpolation techniques based on deterministic and geostatistical principles have been used (Ly et al., 2011). Deterministic interpolation methods involve the Thiessen polygon (THI) Inverse Distance Weighting (IDW), polynomial interpolation (PI), spline Interpolation (SI) and Moving Window Regression (MWR) techniques (Ly et al., 2013). The general disadvantage of these afore mentioned techniques is their inability to quantify interpolation uncertainty. Geostatistical methods on the other hand involve techniques such as Ordinary Kriging (OK), Simple Kriging (SK), and Universal Kriging (UK), etc; with basically every Kriging method dealing with analysis of random fields $Z(u)$ (Z is random and u is the geographical coordinate index) (Bivand et al., 2008). Kriging involves estimation and modeling of spatial dependency, autocorrelation and verification of the stationarity assumption. Geostatistics methods, relative to deterministic techniques, generally provide improved predictions resulting from considering the spatial configurations of observations (Bivand et al., 2008). Even with geostatistical techniques, the impact of spatial smoothing is a challenge (Boluwade and Madramootoo, 2013), resulting in the underestimation of large values and over estimation of small values of the variable of interest. Interpolation accuracy is also only as good as the number of spatial datasets used in producing the resulting interpolated surface (Bivand et al., 2008). Therefore, in large mid- to high-latitude regions of Canada, there is need for better methods of precipitation representation using more advanced datasets from weather prediction models at appropriate scales for hydrologic modeling applications.

The development of Numerical Weather Prediction (NWP) models (Haltiner, 1971) provides a platform to predict precipitation at short time steps. NWP uses current observation of weather combined with model predictions to forecast future events (Belair et al., 2003, Mailhot et al., 2006). It is a procedure that applies 3-D differential atmospheric equations to estimate future states of the atmosphere that require initial conditions (Cote et al., 1998, Girard et al., 2010). The Regional Deterministic Prediction System (RDPS) of Environment Canada is an operational short-term NWP system developed for North America, four times daily with a horizontal resolution of 10-km (Fortin et al., 2015). RDPS relies on the Canadian Global Environmental Multi-scale model (GEM; Fortin et al., 2015). GEM is an integrated forecasting system and data assimilation platform which is based on the hydrostatic primitive equations with time discretization following an implicit two-time-level semi-Lagrangian procedure (Lespinas et al., 2015). The performance assessment of GEM as a short-term forecast product has been reported by Cote et al. (1998). Their results further demonstrated the validity and versatility of GEM model. GEM’s short range forecast is suitable for a background field for reanalysis products such as the Canadian Precipitation Analysis (CaPA) (Lespinas et al. 2015).

CaPA is a joint Canadian project between the Meteorological Research and Department (MRD) and Meteorological Services of Canada (MSC), both divisions of Environment Canada (Mahfouf et al., 2007). CaPA produces 6 h and 12 h accumulation of precipitation covering all of North America on a 10 km grid. CaPA assimilates the GEM’s short-term forecasts, radar precipitation estimates (Fortin et al., 2015), satellite observation (Boluwade et al., submitted for publication) and point estimates from weather stations using an internal quality control procedure (Lespinas et al., 2015, Fortin et al., 2015). CaPA offers potential as a forcing product for hydrologic modeling across Canada, particularly regions in the mid- to high-latitudes with sparse gauge networks. However, there is a need to first verify the reliability and accuracy of CaPAs precipitation estimates such that Canadian provincial agencies, public institutions and interested parties may have confidence in using CaPA for operational decision making. Assessing the reliability of CaPA as an additional data source (with actual observations) must therefore involve comparisons point–to–point (P2P), map-to-map (M2M), and as a forcing for an operational hydrologic model.

Comparing on a point basis, change point analysis is a reliable technique to determine the timing and partitioning of precipitation series (Fortin et al., 2004, Fu et al., 2014, Gallagher et al., 2012). Change point analysis based on Bayesian techniques can be used to help determine similarities and aberrations among the datasets, where Bayesian statistics outperform frequentist methods, which is beneficial in quantifying uncertainty in Bayesian model parameters (Banerjee et al., 2004). Comparing on a spatial or regional basis, the datasets must first be interpolated, which is also a necessary first step for fully distributed models such as WATFLOOD™ that require gridded inputs. Variogram analysis and kriging techniques that take care of spatial configurations of the observations can provide the spatial dependency and spatial autocorrelation properties of the datasets (Bivand et al., 2008). Furthermore, the reliability of CaPA as an alternate precipitation forcing can be tested by using the reanalysis data in a hydrological model directly, and comparing performance metrics to those obtained from forcing with gauge observations.

The primary focus of this paper is to assess and quantify the reliability of CaPA precipitation estimates as a hydrologic model forcing relative to Canadian observation data on the Canadian Prairies. The objective is to assess CaPA as a reliable alternative for CDCD using techniques based on P2P, M2P and as a forcing for interpolated hydrologic model. This paper is organized as follows: Section 2 describes the study area with descriptions of the datasets (CaPA and CDCD), Bayesian change point for P2P, spatial pattern, autocorrelation and interpolation for M2M and hydrologic modeling process: using WATFLOOD™ Section 3 presents and discusses the results. Finally, Section 4 details the conclusions and recommendations.