An accurate representation of spatio-temporal characteristics of
precipitation fields is fundamental for many hydro-meteorological analyses
but is often limited by the paucity of gauges. Reanalysis models provide
systematic methods of representing atmospheric processes to produce datasets
of spatio-temporal precipitation estimates. The precipitation from the
reanalysis datasets should, however, be evaluated thoroughly before use
because it is inferred from physical parameterization. In this paper, we
evaluated the precipitation dataset from the Bureau of Meteorology
Atmospheric high-resolution Regional Reanalysis for Australia (BARRA) and
compared it against (a) gauged point observations, (b) an interpolated
gridded dataset based on gauged point observations (AWAP – Australian Water Availability Project), and (c) a global
reanalysis dataset (ERA-Interim). We utilized a range of evaluation metrics
such as continuous metrics (correlation, bias, variability, and modified
Kling–Gupta efficiency), categorical metrics, and other statistics (wet-day
frequency, transition probabilities, and quantiles) to ascertain the quality
of the dataset. BARRA, in comparison with ERA-Interim, shows a better
representation of rainfall of larger magnitude at both the point and grid scale
of 5 km. BARRA also more closely reproduces the distribution of wet days and
transition probabilities. The performance of BARRA varies spatially, with
better performance in the temperate zone than in the arid and tropical
zones. A point-to-grid evaluation based on correlation, bias, and modified
Kling–Gupta efficiency (KGE
Availability of accurate precipitation datasets is an essential requirement for the modelling of natural processes, hydro-meteorological analyses and forecasting, and monitoring climatic variations and changes (Kirschbaum et al., 2017; Kucera et al., 2013; Robertson et al., 2013). Comprehensive knowledge of occurrence and distribution of precipitation is however hindered by the sparseness of the gauging network. Variations in the density and coverage of the gauging network make it difficult to capture information on the spatial and temporal variability of rainfall. This is particularly the case in areas covered by deserts, mountains, and oceans and in large areas with low population densities (Salio et al., 2015; Thiemig et al., 2012). This presents a challenge for the Australian continent, where the gauges are mostly located along the densely populated coastal regions. The station network is less dense in the central region, which represents the more arid part of the continent (Johnson et al., 2016).
The difficulties inherent in existing observation networks have prompted the development of various gridded datasets with a consistent spatial and temporal scale. One such precipitation dataset for Australia is the interpolated precipitation product from the Australian Water Availability Project (AWAP). Because of the uneven gauge network distribution, it is not consistent in terms of accuracy (Jones et al., 2009). Global and regional reanalysis datasets provide another source of precipitation data at a consistent spatial and temporal resolution. Such reanalysis datasets are generated using a numerical weather prediction (NWP) model and a data assimilation scheme to incorporate the available observations, thereby providing a consistent method of representation of the atmosphere at a regular interval over a larger spatial and temporal domain (Parker, 2016). A range of global reanalysis datasets such as National Centers for Environmental Prediction – Climate Forecast System Reanalysis (NCEP-CFSR; Saha et al., 2010), the European Centre for Medium-Range Weather Forecasts (ECMWF) ReAnalysis Interim (ERA-Interim; Dee et al., 2011), and the Japanese 55-year Reanalysis (JRA-55; Kobayashi et al., 2015) are readily available.
Global reanalysis datasets have been evaluated for various applications at global and regional scales. Of the global dataset exclusively based on reanalysis, ERA-Interim is generally considered to provide better performance compared with other reanalysis datasets (Beck et al., 2017, 2019). ERA-Interim has been found to reproduce the climatology of global monsoon precipitation (Lin et al., 2014), and in general it demonstrates high temporal and spatial correlations with interpolated observations (Donat et al., 2014). In a recent global evaluation of gridded precipitation datasets using gauge observations by Beck et al. (2017), the ERA-Interim and JRA-55 reanalysis datasets were found to reproduce long-term trends and temporal correlation more reliably than those achieved by satellite datasets. On the Australian continent, ERA-Interim reproduced the observed spatial patterns of long-term rainfall along with other climatic variables and showed an overall better performance compared with NCEP-NCAR (National Centers for Environmental Prediction – National Center for Atmospheric Research) reanalysis (Fu et al., 2016). An evaluation by Peña-Arancibia et al. (2013) of reanalysis datasets, satellite products, and an ensemble of these datasets in Australian and Asian regions showed that the ERA-Interim performed better than other individual datasets across a range of metrics for the Australian region.
The available global reanalysis datasets such as NCEP-CFSR, ERA-Interim, and
JRA-55 cover the Australian region, but their horizontal resolutions are
relatively coarse (
The Bureau of Meteorology Atmospheric high-resolution Regional Reanalysis for Australia (BARRA) includes a 12 km modelling and assimilation system (BARRA-R; in this paper, we refer to BARRA-R as BARRA for convenience), which is the first 12 km regional reanalysis conducted over Australia, New Zealand, and south-eastern Asia (Su et al., 2019). BARRA is expected to provide an improved understanding of the past weather than previously possible, particularly for extreme events, and this should support better planning and management of climate-related risks in the future. BARRA makes use of local surface observations and locally derived wind vectors which are not available to global reanalysis models. However, precipitation observations are not assimilated in BARRA. Precipitation is modelled using a microphysics parameterization and a mass flux convection scheme. The modelled precipitation can be erroneous (Bukovsky and Karoly, 2007; Parker, 2016). Therefore, it is necessary to understand the nature of the uncertainty and inaccuracies involved and to use this understanding to correct for any systematic errors.
The criteria for evaluation of the reanalysis precipitation need to be rigorous and relevant to the intended purpose of use (Parker, 2016). The type of reference datasets (point or grid), the spatial and temporal scale, and the metrics used vary widely among the studies. In general, gridded precipitation products including reanalysis datasets, satellite precipitation estimates, and/or interpolated datasets are evaluated against benchmark datasets (Baez-Villanueva et al., 2018; Beck et al., 2017; Gebremichael, 2010; Isotta et al., 2015; de Leeuw et al., 2015; Peña-Arancibia et al., 2013; Zambrano-Bigiarini et al., 2017). Suitable benchmark datasets comprise point measurements (Baez-Villanueva et al., 2018; Chiaravalloti et al., 2018; Salio et al., 2015; Thiemig et al., 2012; Zambrano-Bigiarini et al., 2017) and/or high-quality gridded datasets (Chiaravalloti et al., 2018; Isotta et al., 2015; Peña-Arancibia et al., 2013). Catchment-scale representativeness of precipitation datasets is assessed by evaluating catchment average precipitation (e.g. Thiemig et al., 2012). Evaluation studies for hydrological applications focus on a daily timescale analysis (Baez-Villanueva et al., 2018; Thiemig et al., 2012; Zambrano-Bigiarini et al., 2017), though this has been extended to a sub-daily time interval for the purpose of hydrologic risk assessment (Chiaravalloti et al., 2018).
Given the complex statistical behaviour of precipitation, a range of evaluation metrics are available which have their own assumptions and limitations. Typical metrics include unconditional scores (e.g. correlation, bias, root-mean-square error, and mean absolute error), categorical metrics (e.g. probability of detection, false alarm ratio, and skill scores), and distributional statistics (Gebremichael, 2010). In addition, statistical properties like quantiles, wet-day frequency, and transition probabilities help to ascertain if the dataset under evaluation preserves the statistical properties related to sequencing. Any single metric cannot adequately represent the nature of all errors in the precipitation products. Therefore, it is essential to evaluate multiple metrics which describe different aspects of precipitation to identify the possible sources of mismatch between datasets (Baez-Villanueva et al., 2018).
This study is one of the first comprehensive explorations of the BARRA precipitation estimates. It aims to identify strengths and limitations of the BARRA precipitation to ascertain its efficacy for further hydrometeorological applications in the Australian region. The evaluation is performed against gauged measurements from two data sources: one is based on gauged point observations, and the other is a high-resolution gridded AWAP dataset derived from interpolating gauge measurements, which is widely accepted as being the best synthesis of gauged observations. There are inherent differences in gridded and point rainfall estimates due to the spatial averaging of point observations across each grid cell area. Since BARRA provides direct estimates of area-average rainfall, it is useful to compare these estimates with the best available point and areal precipitation estimates. In addition to observed and interpolated datasets, we also compare BARRA with corresponding areal rainfalls provided by the ERA-Interim global reanalysis. ERA-Interim was selected, as it is used to drive BARRA (Su et al., 2019), and thus this comparison reveals the extent to which the high-resolution regional reanalysis provides additional value over ERA-Interim in capturing finer-scale meteorology. ERA-Interim was also selected, as it has been found to be one of the best-performing reanalysis datasets (Peña-Arancibia et al., 2013).
Despite the availability of BARRA precipitation as hourly values, we select a daily timescale to match the temporal resolution of the best available gridded reference dataset (AWAP). In addition, the accuracy at a daily scale serves as a desirable first step towards further examination at finer timescales because any reliance on sub-daily estimates necessarily depends on their ability to correctly represent daily precipitation. The additional value of evaluation at a daily scale is that it assesses the potential of BARRA to provide estimates of daily rainfall in the sparsely gauged regions across Australia. We focus our evaluation on a range of metrics related to the suitability of rainfalls for hydrological applications and climate studies. We consider the depth, statistical distribution, wet-day frequency and transition probabilities at a daily level for evaluation. The purpose of this evaluation is thus to shed light on the suitability of BARRA for input to hydro-meteorological applications related to climate change studies, water resource management, and the analyses of floods and droughts.
The sub-hourly time series of recorded rainfall from continuous rainfall stations are obtained from the Bureau of Meteorology. Daily rainfall is generated by aggregating the sub-daily rainfall observations. The period of analysis was determined by the duration at which all the datasets used in the study are available, i.e. January 2010 to December 2015. The gauge stations used in the study are chosen based on the availability of information for the entire period over which BARRA data estimates are available. A total of 441 stations are selected, and their spatial distribution is shown in Fig. 1a.
Average annual precipitation over Australian region.
Overview of gridded precipitation datasets used in this study.
An overview of gridded datasets used in this study is presented in the
Table 1. The Australian Water Availability Project
(AWAP) provides a daily high-quality
The global reanalysis ERA-Interim of the European Centre for Medium-Range
Weather Forecasts (ECMWF,
BARRA also uses 4D-Var but at the 36 km resolution and uses the Unified
Model as its forecast model (Su et al.,
2019). BARRA extends spatially over
The study is conducted over a period of 6 years, from January 2010 to December 2015, which represents the full period of available data at the commencement of the study. We undertake a point-to-grid and grid-to-grid evaluation of the reanalysis datasets against suitable reference datasets. The gauged data represent the best estimate of rainfall at a point, and AWAP data provide the best estimate of rainfall over a grid cell. Both estimates are, however, limited by the available gauging density. AWAP estimates are based on all available daily and continuous gauges, and the point data considered here are from selected sub-daily rainfall stations with maximum availability of data over the study period. The selected gauges are distributed unevenly across Australia and are spatially representative of the available gauging network. As the gridded datasets represent an areal average, it may be expected that there are differences between point and gridded estimates, as the latter account for some spatial averaging. While the gauged data and AWAP rainfall estimates represent the best available reference datasets based on measured data, both are imperfect representations of areal rainfalls. The AWAP estimates contain inaccuracies due to the interpolation method, and the point estimates provide only a coarse estimate of rainfall over a grid cell area. The ability of these point and gridded reference datasets to represent actual areal rainfalls is heavily dependent on the gauging density and local orography, and these factors influence the accuracy of the reference datasets to different degrees across Australia. Accordingly, we compare the BARRA and ERA-Interim estimates with both point gauged and AWAP areal data, where the evaluation of both offers different insights to the quality of model estimates.
Gridded rainfalls are compared with point rainfalls using a nearest-neighbour approach. The choice of interpolation scheme is especially important when comparing datasets of different spatial resolutions; bilinear interpolation is likely to smoothen the precipitation field, resulting in higher bias for larger rainfall (Accadia et al., 2003), whereas the nearest-neighbour method preserves the magnitude of the precipitation over the grid. At each location of point rainfall, the nearest-neighbour method assigns the precipitation amount from the nearest grid point. In grid-to-grid evaluation, both reanalysis datasets are interpolated to a common AWAP spatial scale using nearest-neighbour method. This means that, for each AWAP grid, the precipitation variable is obtained from the nearest grid of reanalysis datasets. Given the spatial inconsistency in the accuracy of AWAP dataset, the grid-to-grid evaluation is limited to grid points that are nearest to the gauge stations used in this study.
The performance of daily precipitation from the reanalysis data is assessed
considering all days of the year and by seasons (summer – DJF; autumn – MAM;
winter – JJA; spring – SON). In addition, evaluation is stratified across
three broad climatic zones (arid, tropical, and temperate) as defined by the
Köppen–Geiger classification
(Peel et al., 2007). Most
comparisons are undertaken using estimates from every day in the 6-year
period, and thus the number of wet days varies with the different datasets
and locations considered. However, it is recognized that reanalysis products
tend to produce a high number of days with light drizzle, and therefore they
overestimate the frequency of wet days. Thus, following
Baez-Villanueva
et al. (2018), Ebert et al. (2007), and Zambrano-Bigiarini et al. (2017), a
threshold of 1 mm d
Modified Kling–Gupta efficiency (KGE
The evaluation of precipitation data generally involves an assessment of
detection capabilities and biases in the form of continuous and categorical
metrics. The continuous metrics used in this study are the modified Kling–Gupta
efficiency (KGE
The four categorical indices adopted are the probability of detection (POD)
or hit rate, false alarm ratio (FAR), critical success (or threat) index
(CSI), and frequency bias (fBias). These are evaluated over five different
rainfall intensity classes: no rain (
In this study, we evaluate the capacity of the reanalysis datasets to reproduce a range of precipitation statistics related to the frequency of wet days; transition probabilities between wet and dry days; and the distribution of rainfall amounts based on the 90 %, 95 %, and 99 % daily exceedance values.
The transition probabilities considered here are
Figure 1 shows the average annual precipitation over Australia for the period of 6 years (2010–2015) as estimated using the different datasets (gauged point rainfalls, AWAP, BARRA, and ERA-Interim). It is seen that there is a high spatial variability in the average precipitation over Australia. Regions of high rainfall (northern and eastern coasts along with western Tasmania) are similarly represented by all four datasets. ERA-Interim provides a coarser representation of the precipitation field, which expectedly fails to capture the higher spatial variability in the coastal regions and orographic precipitations in the Great Dividing Range and western Tasmania. By comparison, BARRA precipitation captures this variability in the AWAP precipitation. It should be noted, however, that the AWAP data provide a poor estimate of precipitation over central Australia, where there is a paucity of gauging information. BARRA, on the other hand, provides a high-resolution precipitation pattern over the ocean as well as the central Australian region where gauges are sparse.
Boxplot of correlation (
The variation in KGE
Median values of KGE
Figure 3 presents the summary of KGE
Evaluation of daily rainfall in different seasons reveals a mixed
performance between the different reanalysis products. The performance
during winter is better than in summer for both reanalysis datasets. This is
likely due to the ability of NWP models to accurately simulate synoptic
systems which represent the majority of wintertime rainfall. However, this
difference is larger for BARRA than ERA-Interim. There is less discrepancy
in the correlation between BARRA and ERA-Interim. However, the variation in
correlation across stations is higher for the BARRA dataset. BARRA tends to
overestimate the depth of rainfall, and the ERA-Interim underestimates
variability component across all the seasons. The mean and variability for
BARRA during winter (JJA) are closer to gauge estimates than in the summer (DJF)
season. Based on KGE
The summary of the KGE
Figure 4 shows the comparison of the wet-day
frequency and transition probabilities (
Wet-day frequency estimates from gridded datasets (AWAP, BARRA, and
ERA-Interim) and observations at gauge locations
Quantile estimates (mm d
Figure 5 shows the comparison of quantiles obtained from the gridded datasets to the corresponding quantiles derived for the observed point rainfalls. Since all days were considered in computing the quantile (i.e. both wet and dry days were considered), the 99 % value represents “large” rainfalls that are exceeded on average only three or four times per year. On the other hand, the rainfall represented by the 90 % quantile (exceeded on average 36 times per year) corresponds to the more frequent wet-day rainfalls which depend upon the climatology of the station considered. The AWAP rainfalls corresponding to these quantiles are all higher than the point rainfall estimates. The quantile estimates in BARRA and ERA-Interim show very different patterns to those of the AWAP dataset. BARRA estimates for all quantiles (90 %, 95 %, and 99 %) exhibit no consistent bias with location. In contrast, ERA-Interim greatly underestimates the larger precipitation, where the degree of underestimation increases with an increasing quantile. The nature of the differences varies with location and quantile: for example, the ERA-Interim 90 % quantile estimates are biased low compared with the point rainfalls in the northern region, yet there is a tendency for all higher quantiles to be biased high. The ERA-Interim reanalysis appears to be unable to represent higher precipitation magnitudes, and the largest value plateaus at about 40 mm for the rarer quantiles.
Figure 6 shows the boxplot of four categorical
performance indices computed between three gridded datasets with reference
to gauged point rainfalls at a daily scale. The POD is highest for the AWAP dataset, and the POD values for the two
reanalysis datasets are consistently lower and somewhat similar. For
rainfall intensities greater than 20 mm and lower than 5 mm, BARRA shows
higher detection of rainfalls, whereas for the rainfall intensity 5–20 mm,
ERA-Interim shows slightly better detection skill. In contrast, the false
alarm ratio (FAR) is higher for BARRA for the heavy rainfall classes
(
Boxplot of categorical performance indices (probability of detection, false alarm ratio, frequency bias, and critical success index) calculated against gauge data for five classes of rainfall intensity. Each box extends from first to third quartile, medians are marked in each box, and whisker extends to furthest point or 1.5 times the interquartile range, whichever is closer.
Boxplot of correlation (
The boxplot of KGE
In Table 2, a summary of the KGE
The frequency of wet-day and transition probabilities obtained from reanalysis datasets at AWAP grid locations is shown in Fig. 8. Overall, the representation of wet-day frequency and transition probabilities follows a similar pattern to the point-to-grid assessment (Fig. 4). In comparing the gridded datasets with AWAP data as the benchmark, the estimates of wet-day frequency and transition probabilities from BARRA are closer than those of ERA-Interim. However, both reanalysis datasets show an improved correlation and reduced bias in grid-to-grid over point-to-grid evaluation. This is expected, as this bias is attributed to the spatial resolution of the data.
Wet-day frequency estimates from reanalysis datasets (BARRA and
ERA-Interim) and AWAP at gauge locations
The comparison of quantiles in gridded datasets in Fig. 9 shows a similar pattern as observed in Fig. 5 for the gauged point rainfalls. Both reanalysis datasets represent the 90 % quantile reasonably well. However, the difference between the reanalysis products increases at higher quantiles. The degree of bias in ERA-Interim is considerably more pronounced compared with BARRA.
Quantile estimates (mm d
Figure 10 shows the boxplot of four categorical
performance indices computed between reanalysis datasets with reference to
the AWAP dataset at a daily scale. In general, the POD for both datasets is
similar to point-to-grid comparison. For the larger precipitation
intensities, the detection capacity of BARRA is higher, whereas for the
smaller rainfall intensities, the detections from both datasets are close to
each other. The FAR scores for BARRA estimates of higher rainfalls are
slightly greater, indicating the larger occurrence of false alarms while
detecting higher rainfall. For the low rainfall classes, BARRA exhibits
slightly better FAR than ERA-Interim. The CSI score is higher for BARRA
across all rainfall classes and indicates that BARRA is slightly more
skilful than ERA-Interim. At larger rainfall class, the CSI is notably
higher for BARRA, indicating its better performance in representing larger
rainfall. The frequency bias shows that the BARRA precipitation tends to
produce more events of light rainfall while missing out on some larger
rainfalls. The bias in BARRA is, however, smaller in comparison with
ERA-Interim, which shows a marked underestimation of larger rainfall events
(
Boxplot of categorical performance indices (probability of detection, false alarm ratio, frequency bias, and critical success index) calculated against AWAP data for five classes of rainfall intensity. Each box extends from first to third quartile, medians are marked in each box, and whisker extends to furthest point or 1.5 times the interquartile range, whichever is closer.
The evaluation of daily precipitation from BARRA and its comparison against existing datasets reveals a mixed performance under varying benchmark dataset and the metrics. The general observations on the performance of BARRA are summarized in Table 3, and the key insights from these results are discussed below.
Summary of performance of BARRA precipitation.
The BARRA dataset exhibits similar spatial patterns of mean annual rainfall as the AWAP dataset in regions where gauge density is highest (Fig. 1). However, the reliability in the evaluation of BARRA in the central regions of Australia is confounded by the lack of gauges. The coarser resolution of ERA-Interim misses the small-scale variability; in contrast, the more realistic representation of complex topography in BARRA helps to capture fine-scale features such as orographic precipitation (Su et al., 2019). A good level of agreement in spatial patterns of rainfalls with AWAP and similarity with ERA-Interim estimates (over both land and ocean) at a coarser scale suggest that the BARRA dataset might provide useful information on the distribution of rainfall in regions where direct measurement of precipitation is not available.
In general, a slight overestimation is observed in mean rainfall from
high-resolution gridded data when compared with gauged point rainfall
estimates (bias in mm d
In addition, the temporal variability is represented well by BARRA, while this is underestimated by ERA-Interim (Figs. 3 and 7). As with the assessment of spatial variability, the reason for this underestimation is the coarser resolution of ERA-Interim. The variability ratio varies across climatic zones, with the largest difference in the tropical zone. BARRA greatly overestimates variability in the tropical region, whereas the opposite is true for ERA-Interim (Table 2). The discrepancy in the performance across climatic zones is further discussed in Sect. 5.5.
All the gridded datasets exhibit higher frequencies of wet-day occurrence compared to gauge point rainfalls. This difference is consistent with the physical reasoning that the likelihood of rainfall occurring over an area is always higher than that over a point location. Moreover, an increase in the area produces a higher difference in the wet days, and this is consistent with the increasing levels of bias with grid cell size between AWAP, BARRA, and ERA-Interim (Fig. 4). In comparison with AWAP as a benchmark, BARRA closely represents the wet-day frequency and exhibits less uncertainty at higher estimates compared with ERA-Interim (Fig. 8).
The variation in transition probabilities
The comparison of quantiles (Figs. 5 and 9) demonstrates that the BARRA dataset is able to represent the higher quantiles more accurately than ERA-Interim. The BARRA estimates of quantiles show less bias, but variance increases with the magnitude of precipitation amount. This uncertainty could be due to displacement error as the precipitation becomes more localized with an increase in magnitude. ERA-Interim, on the other hand, greatly underestimates the higher quantiles. This could be attributed to its coarser resolution, resulting in the averaging of rainfall over a larger area. These results are consistent with published findings (Isotta et al., 2015; Jermey and Renshaw, 2016) that high-resolution regional reanalysis improves over ERA-Interim in the representation of large rainfall.
The categorical evaluation (Figs. 6 and 10) also illustrates the improved performance of the BARRA dataset over ERA-Interim during larger rainfall events. Because of the higher spatial resolution, BARRA provides a more accurate representation of larger rainfalls. Therefore, the probability of detection is higher for the larger rainfall classes. Despite exhibiting a higher hit rate, BARRA also shows higher false alarm ratios. The increased false alarm in BARRA is due to the higher number of large rainfall events reported by the BARRA dataset. With larger rainfall events, the false alarm is also likely to increase. Such a trend is usually observed in the assessment of reanalysis and satellite rainfall estimates (Zambrano-Bigiarini et al., 2017). The critical success index, the metric which penalizes both misses and false alarms, represents the overall skill of the data. The performance of BARRA lies between the skilful AWAP and less skilful ERA-Interim. The greater CSI for BARRA compared with ERA-Interim suggests that its improved hit rate outweighs the frequency of false alarms. The difference in the CSI between BARRA and ERA-Interim is greater for larger rainfall classes, with the former yielding a better score.
Most metrics suggest that the performance of BARRA is superior in the
southern (temperate) region compared to the northern (tropical) region. The
summary of KGE
This study shows that the general pattern of performance between the reanalysis datasets is similar in the point-to-grid evaluation. However, the evaluation against the gridded AWAP estimates showed a markedly better performance by BARRA in terms of overall bias, variability, wet-day frequencies, transition probabilities, quantiles and categorical metrics. As expected, BARRA better represents areal rather than point rainfall, as it represents the average precipitation field over a grid cell.
The purpose of the current study is to document the performance of the BARRA dataset at a daily scale and to provide a comparative analysis of its strengths and limitations relative to other available datasets. The analysis includes point-to-grid and grid-to-grid evaluations at the gauge locations. A range of metrics representing correlation, daily precipitation statistics, and categorical performance are explored and compared on an annual as well as a seasonal basis.
The high-resolution nature of the BARRA dataset provides more detailed and accurate estimates of rainfall across the Australian region than the coarser ERA-Interim. BARRA precipitation exhibits good agreement with the average annual rainfall from AWAP as well as to the gauge dataset. The correlation statistics of the BARRA estimates are slightly lower than for the global reanalysis (ERA-Interim). However, the depth and variability in daily precipitation from AWAP are better reproduced by BARRA than by ERA-Interim. We can conclude that BARRA precipitation is representative at the spatial scale of AWAP, considering that AWAP data provide the best basis for comparison with the reanalysis datasets.
BARRA provides largely unbiased estimates of larger rainfall quantiles, whereas ERA-Interim is clearly underestimated. Categorical evaluation also shows a better correspondence of the larger events in BARRA compared to ERA-Interim. This has important implications for hydrological modelling, as simulations of runoff processes are heavily dependent on how realistically the precipitation field is represented both spatially and temporally. Information on large rainfalls is also important for many engineering investigations where the design loading conditions of interest are dependent on the characteristics of daily rainfall. The superior performance of BARRA in representing large rainfall strengthens a case for its use over ERA-Interim, where information about extremes is required.
BARRA closely reproduces the frequency of wet days and dry–wet transition probabilities. This evaluation broadly supports BARRA precipitation for its ability to reproduce precipitation statistics at a daily scale. BARRA precipitation could be useful for assessing variation in the spatial and temporal characteristics of precipitation in a consistent manner that is not influenced by differences in gauge density. In addition, it could potentially be used as a source of data in the central arid zone, where AWAP estimates are poor or not available. Considering the limitations in the availability of gauged datasets, BARRA could be considered to be a valuable reference dataset for hydro-climatic analysis across the whole of Australia, particularly where gauging density is low.
The core attraction of the BARRA dataset is the availability of sub-daily precipitation estimates. Such information is not available in the AWAP dataset, and the spatial resolution of the estimates is higher than the currently available global reanalysis and satellite datasets. Accordingly, future work will be directed towards an analysis of BARRA estimates of sub-daily precipitation to assess its ability to represent precipitation at finer timescales.
Codes used for the analysis are available in the Supplement.
The supplement related to this article is available online at:
SCA designed the research and performed the analysis. All co-authors provided ideas and feedback following discussions. SCA prepared the paper, with contributions from all co-authors.
The authors declare that they have no conflict of interest.
We would like to thank colleagues at the Bureau of Meteorology (Peter Steinle, Robert Smalley, and Alex Evans) for discussions at various stages of research. We are also grateful to Dörte Jakob for commenting on early results and for providing feedback on drafts of the paper.
This research has been supported by Seqwater (grant no. CO034338) and the Australian Bureau of Meteorology Scholarship Program (grant no. TA38977).
This paper was edited by Giuliano Di Baldassarre and reviewed by Korbinian Breinl and one anonymous referee.