climate data for the australian water availability...

37
Climate Data for the Australian Water Availability Project Australian Water Availability Project Final Milestone Report David A. Jones, William Wang and Robert Fawcett National Climate Centre Australian Bureau of Meteorology October 2007 Corresponding author: Dr David Jones, National Climate Centre, Australian Bureau of Meteorology GPO Box 1289 Melbourne, Australia 3001. [email protected]

Upload: others

Post on 03-Jul-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

Climate Data for the Australian Water Availability Project

Australian Water Availability Project

Final Milestone Report

David A. Jones, William Wang and Robert Fawcett

National Climate Centre

Australian Bureau of Meteorology

October 2007

Corresponding author:

Dr David Jones, National Climate Centre, Australian Bureau of Meteorology

GPO Box 1289 Melbourne, Australia 3001. [email protected]

Page 2: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

1

Climate Data for The Australian Water Availability Project

Table of Contents

1. Introduction ...............................................................................................................3 2. The Meteorological (in situ) Data ............................................................................3 3. The Spatial Analysis Methodologies........................................................................6 4. Quality of the Analyses...........................................................................................16 5. Near-Surface Wind Run.........................................................................................27 6. Summary and Conclusions.....................................................................................29 7. References................................................................................................................31 Appendix: Climatological Maps for 1971-2000........................................................33

The Australian Water Availability Project (AWAP) is a partnership between the Bureau of Rural Science, CSIRO (EOC and Land and Water), and the Bureau of Meteorology. The overall objective and outcome statement for AWAP are described in AWAP Work Plan 2004-2006, Version 6, 27 September 2004.

Page 3: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

2

Summary

The Bureau of Meteorology has generated a range of improved meteorological analyses and

remotely sensed datasets for Australia as a contribution to the Australian Water Availability

Project. The meteorological data include analyses of rainfall, temperature, vapour pressure

and wind run at daily and monthly timescales.

Robust topography-resolving analysis methods have been developed and applied to in situ

observations of rainfall, temperature and vapour pressure to produce analyses at a resolution

of 0.05º×0.05º (approximately 5 km × 5 km) for the period 1980 to the present. The resulting

analyses represent substantial improvements on operational analyses currently produced by

the Bureau of Meteorology. Coarser-resolution analyses of wind run have also been

developed.

This report provides a detailed overview of the project outputs and the analysis systems which

have been used to generate them. Careful attention has been paid to developing systems and

datasets which are robust and useful for the monitoring of both climate variability and climate

change. These systems are now running in real-time and are expected to form the basis for

ongoing monitoring of Australia’s surface climate by the Australian Bureau of Meteorology.

Page 4: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

3

1. Introduction

The past year has seen much of Australia suffer from severe meteorological and hydrological

drought. This has seen water resources fall to record lows and agricultural production

massively decline in many regions (e.g., MDBC 2007). There has never been a greater need

for up-to-date information about the state of Australia’s climate and water resources to ensure

that water resources are used in a sustainable manner.

A key to the better management of water is ensuring that demand does not exceed the long-

term supply. At the present time Australia does not have a comprehensive or consistent source

of information on water balance, covering the relationships between rainfall, evaporation,

evapotranspiration, run-off and drainage to surface and ground water. The National Heritage

Trust supported Australian Water Availability Project was designed to address this knowledge

gap. The objective of this project being “to develop an operational prototype of a new and

integrated approach to monitoring and predicting soil moisture and other components of the

water balance” .

An important aspect of this project is the integration of in situ (conventional) meteorological

observations with remotely sensed data, thereby allowing the description of water and climate

at finer space and time scales. The conventional data are temporally dense (Automatic

Weather Stations - AWS - routinely record one minute data) but spatially sparse. In contrast,

the remotely sensed satellite data are often sampled on scales of a few kilometres or finer, but

infrequent, with consecutive views hours or days apart (that is, spatially dense but temporally

sparse).

The Bureau of Meteorology has developed improved spatial analyses to support the

Australian Water Availability real-time hydrological analysis system for Australia. The

meteorological analyses cover precipitation, maximum and minimum temperatures, vapour

pressure and near-surface wind run (Jones et al. 2006). The meteorological analyses (both

daily and monthly averages) are augmented by a series of new remotely sensed data including

global solar radiation from geostationary satellites and Normalised Difference Vegetation

Index (NDVI) and land surface temperature from the Advanced Very High Resolution

Radiometer (AVHRR) on polar orbiting satellites. The remotely sensed data are described in a

separate report.

Page 5: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

4

The Meteorological (in situ) Data

The data and analyses described in this report are largely based on conventional

meteorological observations from weather stations (“Meteorological Data”). New spatial

analysis techniques have been used to generate high-resolution surfaces from these

observations. The methods employed are described in this report.

Variable

Source

Temporal Resolution

Spatial Resolution

Precipitation Analysis of rain gauge data

Daily and monthly average

0.05°×0.05°

Daily Maximum Temperature

Analysis of thermometer data

Daily and monthly average

0.05°×0.05°

Daily Minimum Temperature

Analysis of thermometer data

Daily and monthly average

0.05°×0.05°

Vapour Pressure “Humidity”

Analysis of dewpoint data

Daily and monthly average of 9am and 3pm

0.05°×0.05°

Near Surface Wind Run

NCEP/NCAR numerical model analysis

MesoLAPS numerical model analysis

Daily and monthly average

Daily and monthly average

2.5°×2.5°

0.042°×0.042°

Table 1: Description of the meteorological data.

The meteorological variables and analyses are summarised in Table 1. Maximum and

minimum temperature, rainfall and vapour pressure fields (from dewpoint) have been

generated through the geostatistical analysis of observations at weather stations (a mix of

manual and AWS observations). The analyses are derived from daily and monthly data

contained in the Australian Bureau of Meteorology climate databank (Australian Data

Archive for Meteorology - ADAM). ADAM is updated in real-time, thereby allowing

analyses of new data to be produced in a timely fashion. Over the analysis period (1980 to

2006) the rainfall network contains an average of 5760 stations, while there is an average of

721 temperature stations and 670 dew point (“vapour pressure”) stations (see Figure 1). The

near-surface wind run analyses come from a numerical weather model-based assimilation

scheme as described below.

Page 6: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

5

(a)

(b)

(c)

Figure 1: The networks of (a) rainfall, (b) temperature and (c) dewpoint temperature (“vapour

pressure”) stations contributing to the analyses from 1980 to 2006.

Page 7: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

6

For rainfall, the daily data represent the precipitation (including rain, snow, hail and dew)

accumulated in the 24 hour period to 9am. The maximum and minimum temperatures are the

highest and lowest temperature for the 24-hour period starting/ending 9am, respectively. This

convention means that the minimum and maximum temperatures will have generally occurred

on the same calendar day (in the morning and afternoon, respectably). The vapour pressure is

that observed at 9am and 3pm local time. These two times have the best data coverage for

Australia. We note that the vapour pressure has a rather weak diurnal cycle through the day

(Jeffrey et al. 2001, Appendix). The vapour pressure has been calculated at stations using

observations of dewpoint temperature following Murray (1967).

It is not possible to sensibly analyse wind from surface anemometer observations alone. This

is because wind is particularly sensitive to local factors such as station location and exposure.

This sensitivity leads to large errors of representativeness with individual observations being a

poor indicator of the overall field of wind (Daley 1993). The nature of this problem varies

with the landscape. For example, mountainous areas will have complex wind patterns where

individual wind observations may be quite misleading, whereas flat farming areas will tend to

be less complex and individual wind observations may be broadly representative.

The near-surface wind field is related to the air temperature and pressure (Holton 1992). For

this reason representative wind analyses can be created using a multivariate analysis approach,

such as is used for numerical weather prediction (NWP). In this project we have generated

wind products from two sources. A long historical sequence back to 1980 has been generated

from the National Centers for Environmental Prediction (NCEP)/National Center for

Atmospheric Research (NCAR) reanalyses (Kalnay et al. 1996) described below. These

coarse resolution data have been augmented with very high resolution Bureau wind analyses

at approximately 4 km resolution for the period from 2005 to the present. These analyses are

based on the Bureau’s regional weather forecast model MesoLAPS (Puri et al., 1998) as

described below.

2. The Spatial Analysis Methodologies

The maximum temperature, minimum temperature, vapour pressure and rainfall fields are

based on the spatial analysis of meteorological observations. The observations are from the

Bureau of Meteorology’s ADAM climate database. The underlying station networks have

generally improved slightly over the period of analysis (Figure 2), meaning that the

underlying analyses should improve with time.

Page 8: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

7

Generating daily analyses of meteorological data which are consistent with monthly and long-

term analyses is not straight-forward (Rayner et al. 2004). Temporal averaging dampens the

small scale variability while random observational errors are also reduced by time averaging.

In addition, the relationship with topography is more robust on longer time scales. This means

that the monthly average fields are smoother and observations relatively less error prone. As a

result the monthly average is simpler to analyses than daily data. An added complication is

the fact that some stations have digitised monthly data but not digitised daily data and vice

versa (Jones and Trewin 2002).

It is reasonable to require that the analysis fields

• provide a more accurate representation of short-term climate than does the long-term

average (or climatology);

• be consistent with long-term averages (or climatology); and

• have values which are limited to a physically realistic range.

The last of the points while seemingly trivial is important as it is possible for interpolated

surfaces to become quite unrealistic in data voids when meteorological gradients are strong.

We have developed an anomaly-based approach for analysing rainfall, vapour pressure and

temperature. The new approach is similar to the system described by Hunter and

Meentemeyer (2005) and Xie et al. (2007). It uses a decomposition of the meteorological

variable being analysed into a long-term average component and an anomaly component. The

basis for this decomposition is that anomalies tend to be rather smoother that the raw fields

and that climatology provides information beyond that contained in individual station

observations alone.

The analysis methodology represents an extension of the background/increment method of

analysis which is popular for meteorological (weather and climate) analysis (e.g., Koch et al.

1983; Daley 1992; Jones and Trewin 2000). It is important that systems which are to be used

operationally as part of this project are robust and hence the similarity with these systems is a

positive.

An advantage of the anomaly approach is that anomalies tend to be weakly related to altitude

(owing to the tendency for the atmosphere anomalies to be approximately barotropic) and so

can be adequately analysed with a two-dimensional analysis procedure. In addition the

average of all analyses will, by design, be consistent with the underlying climatology.

Page 9: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

8

0

1000

2000

3000

4000

5000

6000

7000

8000

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

Year

Num

ber

of R

eportin

g R

ainfa

ll Sta

tions

(a)

0

200

400

600

800

1000

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

Year

Nu

mb

er o

f R

epo

rtin

g S

tati

on

s

Vapour Pressure

Temperature

(b)

Figure 2: The number of stations contributing to the (a) rainfall and (b)

temperature and vapour pressure analyses by year.

We have used the Barnes successive-correction method for the analysis of the anomalies

(Koch et al. 1983; Seaman 1989; Jones and Weymouth 1997) and three-dimensional

smoothing splines for the analysis of climatological averages (Hutchinson 1995). These two

Page 10: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

9

techniques have been used widely in meteorological applications previously and have been

shown be robust (e.g., Hutchinson 1995; Jones and Trewin 2000; BoM 2000). The Barnes

method has a number of advantages including being computationally efficient, robust (coping

with strong gradients and data voids) and tuneable. In some contrast, the spline method is

particularly suited to analysing the rather smooth climatological relationships between

meteorological variables and latitude, longitude and altitude but not well suited to noisy data.

These analysis methods are both “statistically” optimal, in that the analysis fields have the

smallest error subject to the constraints on the final analysis. The analysis methods have been

implemented in a modular fashion, which allows for the use of alternative methods for

generating the climate (background) and anomaly (increment) fields in future.

Application to Temperature and Vapour Pressure

In the following, data values at a station k (k = 1,…,S) are denoted with a suffix k, while

analysis values are defined at locations (x,y,z), the coordinates being latitude, longitude and

altitude. Values are defined at a series of time points, denoted by t (t = 1,…,N). The over-bar

denotes a simple temporal average, taken over a 30-year period, following World

Meteorological Organization convention. We have chosen to use the 1971-2000 period as the

base climatology rather than 1961-1990 to improve data coverage. This representation leads

to

)()( tTTtT ′+= (1)

in general,

)()()( tTtTtT kkk ′+= (2)

at stations, and

),,(),,(),,,( tyxTzyxTtzyxT ′+= (3)

for the analysis as a whole. We note in (1), (2) and (3) that the anomaly may be for a day or a

month, both taken from the monthly climatology (1971-2000).

Page 11: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

10

(a)

(b)

Page 12: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

11

(c)

Figure 3: Climatological (monthly) average for February (1971-2000) (a), daily

anomaly analysis (b) and summation (c) for maximum temperature on 1st

February 2007. Maps (a) and (b) represent the first and second terms on the

right-hand-side of equation (3).

Application to the period 1971 to 2000 and beyond

The anomaly analysis is calculated using an optimal two-dimensional Barnes successive

correction analysis procedure, as described in Jones and Trewin (2000) analysed to location

(x,y). This leads to

�� =

=

′=′S

kkkS

kk

tTyxWyxW

tyxT1

1

)(),(),(

1),,( . (4)

The weights Wk(x,y) in equation (4) are obtained using the iterative Barnes algorithm

described by Jones and Weymouth (1997). The weights have been found via exhaustive cross-

validation following Seaman (1989).

The climatological (temporal-mean) analysis is calculated using three-dimensional smoothing

splines following Hutchinson (1995);

Page 13: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

12

),,(),,( ,...,1zyxFzyxT

STT= . (5)

The subscripts on the spline function F(x,y,z) are intended to indicate that the spline function

is dependent on the climatological station data STT ,...,1 . The combined analysis is a simple

sum of the climatological analysis (5) and the anomaly analysis (4);

)(),(),(

1),,(),,(),,(),,,(

1

1

,...,1tTyxW

yxWzyxFtyxTzyxTtzyxT

S

kkkS

kk

TT S�

� =

=

′+=′+= . (6)

An example of this procedure is shown in Figure 3 for daily temperature. The rainfall,

temperature and vapour pressure climatologies for January and July are shown in the

Appendix.

Incomplete Records

The incomplete climate records at stations introduce difficulties in defining the 30-year mean

(1971-2000) at stations and hence in forming (1), (5) and (6). While some stations will have

30 years of data (or nearly so), many stations have incomplete records. The incomplete

records may, however, provide additional useful information for both the climatology and

anomaly analysis. The use of incomplete station records introduces a trade-off between spatial

sampling and temporal completeness when deriving climate means and anomaly analyses.

Let Nk be the number of observations available at the station for the calendar month under

consideration. For a station with a complete or nearly complete record (Nk � 30), we have

)(1

1

tTN

TkN

tk

kk �

=

= . (7)

Using data for stations with short records requires us to form an estimate of the true 30-year

average. We have achieved this using a trade-off between the station’s temporal average

(using all available observations) and an estimate of the temporal average calculated by

interpolation to the station location of the smoothing spline analysis (5). The spline analysis is

obtained using only stations with complete or near-complete records. We thus have available

two alternative and independent estimates for the station average at station k with Nk < 30;

)(1 30

1

tTN

TkN

tk

kk �

<

=

≈ (8)

and

Page 14: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

13

),,(),,()(

1),...,(

1),(

1),...,(

1 30

1

301

11

1

301

11

1

301

11

1

kkktT

NtT

NtT

NtT

N

kkkk zyxFzyxTTSN

tS

S

kN

tk

k

kN

tk

k

N

t����≈

=

≈+

=+

+

≈−

=−

=

=≈ . (9)

We have chosen to use a linear combination of these two different estimates, with the weights

chosen through an optimisation process using cross-validation;

���

���

�−+�

���

�≈

����

<

=≈

=

≈+

=+

+

≈−

=−

=

� ),,()1()(1

)(1

),...,(1

),(1

),...,(1

30

1

30

1

301

11

1

301

11

1

301

11

1

kkktT

NtT

NtT

NtT

N

N

tk

kk zyxFwtT

NwT SN

tS

S

kN

tk

k

kN

tk

k

N

t

k

. (10)

Cross-validation showed that for stations with Nk ≥ 12, it is better to use the simple temporal

average (8) of the available observations, while for Nk < 12 it is better to take a weighted

combination of the two estimates (8) and (9). This optimisation also revealed that the final

analysis accuracy is only weakly dependent on the exact form of the weights in (10). Figure 4

shows the optimal linear weights for this sum obtained for the analysis of rainfall (the same

weights have been applied to all fields for consistency). We note that in (5) and (9) the

stations used are thus those with Nk ≥ 12, rather than Nk � 30, consistent with the weights.

Climate Weightings

0

0.2

0.4

0.6

0.8

1

1.2

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

N

Wei

ght

Climate Weight

Spline Weight

Figure 4: Linear weights for deriving the temporal average in (10).

Application to Rainfall

Rainfall differences from climatology contain a substantial topographical signature and are

generally no smoother (and indeed they may even be noisier) than the raw rainfall fields

themselves. In contrast, the rainfall percentage of mean shows rather less topographical

Page 15: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

14

signature than simple rainfall anomalies. For rainfall, we have defined the anomalies using

division rather than subtraction. The equivalent equations to (1), (2) and (3) for rainfall are

)()( tRRtR ′×= (11)

in general,

)(tRRR kkk ′×= (12)

at stations and

),,(),,(),,,( tyxRzyxRtzyxR ′×= (13)

for the analysis.

The form of the final analysis comparable to (6) becomes

),,()(),(),(

1),,(),,(),,,( ,...,

1

1

1zyxFtRyxW

yxWzyxRtyxRtzyxR

SRR

S

kkkS

kk

×

����

����

′=×′= �� =

=

. (14)

As before, climate averages at stations are calculated using (7) and (10) applied to rainfall.

The rainfall climatology for January and July is shown in the Appendix.

Measuring The Accuracy of the Analysis Process

The accuracy of the spatial analyses has been determined through cross-validation at stations.

Validation results have been calculated for the five years 2001-2005 (see Jones and Trewin

2000). This period has been chosen as it is fully independent of that used to define the

climatology. The 2001 to 2005 period is sufficiently long that it covers a range of climate

regimes and so should be a robust estimate of the analysis accuracies overall. We note that the

climatological analyses are not significantly influenced by individual stations so the impact of

not cross-validating this step is insignificant (though using 2001-2005 avoids this

approximation).

Cross-validation has been achieved by randomly deleting 5% of the stations in the network,

performing an analysis using the remaining 95% of station observations and then calculating

the analysis errors for the omitted stations. This process was repeated 20 times for each

month/day providing independent verification statistics for every station. The errors at

stations have been used to generate maps of analysis error and to generate all-station average

errors. As noted previously the network has not changed significantly since 1980, so this five

Page 16: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

15

year period should be representative of the overall analysis accuracy for the 1980 to present

period.

Jones et al. (2006) describe a range of issues with the method of cross-validation. Importantly,

cross-validation will give somewhat inflated analysis errors, as the method involves a

degrading of the data network compared to reality (e.g., Jones and Trewin 2000; Jeffrey et al.

2001).

In addition, calculating analysis errors by independent cross-validation against station

observations introduces a bias due to observation “error” (see Daley 1993; Jones and Trewin

2002). Consider a cross validated estimate of a station value T at station k and time t, denoted

by ),,,(ˆ tzyxT kkk . This is calculated using the 95% of the network which is retained in the

cross-validation step. This is a cross-validated version of (6). The cross-validated analysis

error is given by

)(),,,(ˆ)( tTtzyxTtE kkkkk −= . (15)

Aggregating across time, we can calculate a station Root Mean Square (RMS) analysis Error

(RMSE) at the station

[ ] [ ]��==

−==N

tkkkk

N

tkk tTtzyxT

NtE

NRMSE

1

2

1

2 )(),,,(ˆ1

)(1

. (16)

Note that N will vary from station to station and according to whether the analysis is for daily

or monthly data. The observation Tk(t) can be divided into a “true” component and an

“observational error” component ek(t). The true component is what would be measured if the

observation at station k was completely accurate, while the error component is the error

introduced due to factors such as instrument miscalibration, misreading by the observer,

errors in spatial representativeness arising from specific factors at the observation site and so

on. Hence we have

)()()( tetTtT kTrue

kk += , (17)

giving

( ))()(),,,(ˆ)( tetTtzyxTtE kTrue

kkkkk +−= (18)

and

Page 17: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

16

( )[ ]�=

+−=N

tk

Truekkkkk tetTtzyxT

NRMSE

1

2)()(),,,(ˆ

1. (19)

On the assumption that the observational errors are statistically independent of the

interpolated and true values (Daley 1993), this can be further simplified to

( ) [ ] [ ]

( ) ( ) .

)(1

)(),,,(ˆ1

22

1

2

1

22

Obsk

Truek

N

tk

N

t

Truekkkkk

EE

teN

tTtzyxTN

RMSE

+=

+−= ��== (20)

In (20)( )2TruekE is the “true error variance” and is a measure of the accuracy of the analysis

in estimating the true field. This value is the true analysis error. The second term ( )2ObskE is

the “observational error variance” and measures the accuracy of the observations. Clearly,

even a perfect analysis will have a non-zero cross-validated error because observations

have some level of error. To obtain a zero cross-validated error, the observations need to be

“perfect”. While it is common practice for the cross-validated differences between

independent observations and analyses to be treated as “analysis errors”, and they are

defined as such here, it is important to keep in mind that they also contain an observation

error component. Daley (1993) and Jones and Trewin (2000) describe how the

observational errors can be estimated.

In defining the analysis errors averaged across time and stations we have used the

additional measures of bias and Mean Absolute Error (MAE). These are both defined in

the usual way (e.g., Jones and Weymouth 1997),

[ ]�=

=N

tkk tE

NBIAS

1

)(1

(21)

and

�=

=N

tkk tE

NMAE

1

)(1

(22)

The all station average is these terms extended across stations,

[ ]� �= =

��

���

�=

S

k

N

tk

k

k

tENS

BIAS1 1

)(11

(23)

Page 18: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

17

and

� �= =

��

���

�=

S

k

N

tk

k

k

tENS

BIAS1 1

)(11

. (24)

3. Quality of the Analyses

Cross-validated analysis statistics for the five years 2001 to 2005 are provided in Tables 2 to 4,

with maps of RMSE in Figures 6 to 8. For reference, we also provide national average

statistics for the operational Barnes analysis system used at the Bureau of Meteorology (see

Jones and Weymouth 1997; Jones and Trewin 2000). There are no operational analyses for

vapour pressure, so a comparison is not possible for this variable. We note that these

verification results are not exactly comparable to those provided by Jones and Weymouth

(1997) and Jeffrey et al. (2001), because of slightly different cross-validation procedures and

different verification periods.

The RMSE for monthly maximum and minimum temperatures are typically between 0.5 and

1ºC, while those for daily temperatures are a little larger at around 1 to 2ºC. There is a strong

correspondence between station density (Figure 1) and analysis error (Figures 5 and 6), with

larger errors occurring through the poorly observed western interior.

The new analyses are a substantial improvement on current Bureau practice for maximum and

minimum temperatures at both the monthly and daily time scales. For maximum temperatures,

the RMSE is reduced by around 40% for daily and nearly 60% for monthly analyses. The

percentage improvement for minimum temperatures is about half as much, but still substantial

at around 0.5ºC for RMSE.

Mean (ºC) Bias (ºC) RMSE (ºC) MAE (ºC) Monthly Maximum Temperature AWA 24.8 0.00 0.7 0.5 Operational 24.8 0.02 1.6 1.1 Monthly Minimum Temperature AWA 12.7 0.00 1.0 0.7 Operational 12.7 0.00 1.5 1.1

(a)

Page 19: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

18

Mean (ºC) Bias (ºC) RMSE (ºC) MAE (ºC) Daily Maximum Temperature AWA 24.8 0.02 1.2 0.9 Operational 24.8 0.00 1.9 1.3 Daily Minimum Temperature AWA 12.7 −0.05 1.7 1.3 Operational 12.7 0.03 2.1 1.5

(b)

Table 2: Verification statistics for the five-year period 2001 to 2005. Monthly

maximum and minimum temperatures (a) and daily maximum and minimum

temperatures (b).

The improvement in the temperature analyses is quite general across Australia. The most

substantial improvements are in regions of significant topography. For example, near the

Victorian Alps and Snowy Mountains the RMSE is reduced from more than 2ºC to around

0.6ºC for daily data (not shown). Similar improvements have been reported by Hunter and

Meentemeyer (2005) for California.

The spatial maps of analysis error for maximum and minimum temperatures (Figures 5 and 6)

show little evidence of increased values near significant topography. This confirms that the

anomalies at both timescales tend to be approximately barotropic. These results support the

two-step approach with the use of a two-dimensional anomaly analysis.

Spatially the RMSE highlights regions where spatial analysis is particularly difficult or the

network insufficiently dense. There is some evidence that analysis errors for maximum

temperature are larger near the coast around northwest Australia and about the Nullarbor Plain,

with the areas near Shark Bay and Eucla standing out in particular. These two coastal regions

often experience very strong gradients in maximum temperatures between the inland deserts

and coasts, and are difficult to analyse with a relatively sparse network. It is possible that

much of the coast of Western Australia and parts of the Northern Territory experience similar

issues during the warmer period of the year.

Page 20: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

19

(a)

(b)

Figure 5: Cross-validated RMSE for monthly maximum (a) and minimum (b)

temperatures. The units are ºC.

Analysis errors for minimum temperature are greater than those for maximum temperatures.

This is because minimum temperatures tend to have larger errors of representativeness and

shorter length scales (see Jones and Trewin 2000). In addition minimum temperatures often

show complex and variable relationships with topography (e.g., Trewin 2005). These factors

mean that a denser network is required for minimum temperature to achieve the same analysis

Page 21: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

20

accuracy as that for maximum temperature. This will be clearly important for analysing

events such as frosts where a difference of 1 to 2ºC may be very significant. We note that the

RMSE for monthly maximum and minimum temperatures are now not much larger than the

theoretical lower bounds calculated by Jones and Trewin (2000, 2002) in parts of inland

eastern Australia. This suggests that in these regions future analysis improvements will

require the use of quite different analysis procedures and new datasets, such as those obtained

by remote sensing.

(a)

(b)

Page 22: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

21

Figure 6: Cross-validated RMSE for daily maximum (a) and minimum (b)

temperatures. The units are ºC.

The monthly rainfall analyses show a modest but significant improvement over current

Bureau practice, and are about half those reported previously for Australian analyses (Jeffrey

et al. 2001). The analysis improvement is most marked in southern Australia where the

climatological signals encapsulated in the climate means are more robust. We note that the

RMSE is substantially larger than the MAE. This is because of a significant skewness in the

distribution of rainfall errors, with a relatively small number of large errors.

There is a marked north-south gradient in the RMSE for rainfall across Australia (Figure 7).

In part this reflects the higher rainfall in the tropical regions which will lead to larger analysis

errors for a given data smoothness and network (see Daley 1993). This pattern has been noted

previously by Mills et al. (1997), Jones and Weymouth (1997), and Jeffrey et al. (2001). This

pattern is further amplified by the tendency for rainfall to be highly convective in tropical

parts and hence to have shorter characteristic length scales (e.g., Mills et al. 1997; Ebert et al.

2007).

Mean (mm)

Bias (mm)

RMSE (mm)

MAE (mm)

MAE/Mean (%)

Monthly Rainfall AWA 53.7 0.2 20.8 11.3 21 Operational 53.7 0.1 23.7 12.6 23

(a)

Mean (mm)

Bias (mm)

RMSE (mm)

MAE (mm)

MAE/Mean (%)

Daily Rainfall AWA 1.9 0.0 3.7 1.1 57 Operational 1.9 0.0 3.8 1.1 57

(b)

Table 3: Verification statistics for the 5 year period 2001 to 2005. Monthly

rainfall (a) and daily rainfall (b).

The RMSEs for daily rainfall are very similar to those for the Bureau’s current operational

system and those reported elsewhere for Australia (Mills et al. 1997; Jeffrey et al. 2001). The

insensitivity of the errors to the analysis method is somewhat surprising, given that the

underlying analysis systems are different. To some extent, these findings may be interpreted

by the observation that for daily rainfall the relationship with topography is not particularly

Page 23: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

22

strong or robust, so the advantage of using topography is not great (though clearly on

individual days and cases this may not be true). This link to topography is particularly weak

in northern Australia where rainfall is more convective.

(a)

(b)

Figure 7: Cross validated RMSE for monthly (a) and daily (b) rainfall. The units

are mm.

Hunter and Meentemeyer (2005) found rather little positive impact from including climate-

topography relationships in daily rainfall analyses in California. A possible way of improving

Page 24: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

23

the analyses might be to develop rainfall-altitude relationships (climatologies) which are

conditional on weather type such as light wind convective situations versus strong on-slope

flow situations.

We also note that the analysis errors for daily rainfall are only weakly dependent on the

Barnes parameters obtained through the optimisation process described by Seaman (1989).

Mills et al. (1997); Weymouth et al. (1999) and Jeffrey et al. (2001) found similar

insensitivities in their analyses of daily rainfall. We interpret this as indicating that the length

scales for rainfall vary markedly from day to day (and also spatially), and hence are not well

approximated by a single parameter set. It is clear that substantial improvements in daily

rainfall analyses will require either far denser networks, or the use of remotely sensed and/or

model-derived data (e.g., Ebert et al. 2007).

Mean (hPa)

Bias (hPa)

RMSE (hPa)

MAE (hPa)

Monthly Vapour Pressure AWA 9am 13.7 0.03 1.1 0.7 AWA 3pm 13.1 −0.05 1.5 1.1

(a)

Mean (hPa)

Bias (hPa)

RMSE (hPa)

MAE (hPa)

Daily Vapour Pressure AWA 9am 13.7 0.02 1.8 1.2 AWA 3pm 13.1 −0.06 2.4 1.6

(b)

Table 4: Verification statistics for the 5 year period 2001 to 2005. Monthly

vapour pressure (a) and daily vapour pressure (b).

Figure 8 shows the distribution of analysis errors for monthly and daily vapour pressure (at

9am). The errors for 3pm are similar (see Table 4) though they tend to be a little larger,

particularly in the north. Following the climatological field (Appendix), vapour pressure

analysis errors increase towards the north where the background means are substantially

higher. There is also evidence of somewhat increased errors close to the coast, where

gradients often tend to be large between moist maritime air and drier continental air, in

agreement with Jeffrey et al. (2001). The lowest analysis errors are found in the well sampled

southeast and southwest parts of Australia.

Page 25: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

24

These vapour pressure analyses are the first of their type to be produced by the Bureau of

Meteorology, and consequently they cannot be directly compared to existing analyses.

Comparison with Jeffrey et al. (2001) suggests these analyses may be a little better. An

important observation is the absence of inflated errors near topography. This suggests that the

vapour pressure/altitude relationship is rather robust and amenable to the two-step anomaly

analysis method we have developed.

(a)

(b)

Page 26: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

25

Figure 8: Cross-validated RMSE for monthly (a) and daily (b) 9am vapour

pressure. The units are hPa.

Some Case Examples

The analysis statistics reveal that the analyses developed through this project are a substantial

improvement (with the exception of daily rainfall). Importantly, these improvements are

evident in most individual analyses. In this section we consider in some detail two recent

examples for monthly and accumulate daily rainfall.

Figure 9 shows the monthly rainfall for June 2004 across Tasmania using the new anomaly

approach, the operational Bureau analysis and from raw station observations. Nationwide, the

monthly analyses have a MAE of 7.9 mm (for the AWA analysis) and 10.3 mm (for the

Barnes operational analysis), revealing a substantial overall improvement. For Tasmania the

MAE values are 20.7 mm and 30.0 mm, respectively, revealing a large improvement in the

accuracy of the rainfall analysis. These can be compared to the Tasmanian station mean for

June 2004 of 151 mm.

Clearly, there is a marked improvement in the representation of Tasmanian rainfall in Figure

9c. The high rainfall in the west, central highlands and northeast are all well captured as is the

tight rainfall gradients in the central parts of Tasmania. To the east the low rainfall about the

coast and near Hobart is also well captured.

(a)

(b)

Page 27: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

26

(c)

Figure 9: Monthly rainfall for June 2004. The current operational analysis (a),

observed totals (b) and AWA analysis (c). The units are mm.

A second example is for the three week period 1st to 21st of June 2007. This period witnessed

severe flooding on the New South Wales coast following a sequence of three major low-

pressure systems. The new analyses show a substantial improvement in the resolution of the

very high rainfall on the coast near Newcastle and secondary maxima south of Sydney and

near the Blue Mountains.

Page 28: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

27

(a)

(b)

(c)

Figure 10: Rainfall for the 1st to 21st of June 2007. The current operational

analysis (a), observed totals (b) and AWA analysis (c). The units are mm.

4. Near-Surface Wind Run

The primary wind-run data developed for the project are compiled from 6-hourly (near

surface) 0.995 sigma level winds extracted from the NCEP/NCAR reanalyses (Kalnay et al.

1996). The 0.995 sigma level is about 40 metres above the surface. These wind analyses are

based on a numerical weather prediction model-based analysis system incorporating global in

Page 29: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

28

situ and remotely sensed data. Given their global extent, these data are on a coarse 2.5°×2.5°

grid (which in the Australian region represents approximately a 250 km×250 km grid). The

analysed wind field is restricted to capturing synoptic-scale wind patterns, with local affects

such as down-slope winds and some sea breezes likely to be missed. It is not possible to

generate sensible analyses for wind using conventional univariate analysis techniques due to

large errors of representativeness at wind observation sites (e.g., Rayner et al. 2004) and a

network which is insufficiently dense for the purpose.

The development of improved wind mapping for Australia is an area of ongoing

research at the Bureau and elsewhere. A gridded hourly analysis of 10-metre wind with

a grid resolution of approximately 4 km×4 km is presently being trialled in the Bureau,

based on short range meso-scale weather forecasts modified with real-time wind

observations. Miller and Benjamin (1992) describe the application of a similar system in

the United States.

Daily and monthly wind run fields are being derived from this system by calculating the wind

speed (the magnitude of the wind vector) at each analysis hour for each grid point and taking

that to be the average wind speed for the entire hour. An appropriate scaling, summed over

the 24 analyses, gives the daily wind run. It is then aggregated to provide a monthly wind run

(expressed for convenience in units of km/day). Data for this new system are only available

from late 2004 onward. Examples of the very high resolution daily and monthly wind run

calculations are shown in Figures 11 and 12 respectively.

Figure 11: A very high resolution wind run analysis for 31 May 2007. The units

are km/day.

Page 30: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

29

Figure 12: A very high resolution wind run analysis for May 2007. The units are

km/day.

6. Summary and Conclusions

In this report we have provided a detailed description of a series of new meteorological

analysis products developed by the Australian Bureau of Meteorology as a contribution to the

Australian Water Availability Project. Careful attention has been paid to developing systems

and datasets which are robust and useful for the monitoring of both climate variability and

climate change. These systems are now running in real-time and are expected to form the

basis for ongoing monitoring and mapping of Australia’s climate by the Australian Bureau of

Meteorology.

The analyses make use of a new two-step analysis system which partitions the analysis field

into a climatological component and an anomaly component. This approach has been found to

be robust, to preserve the background climatology in the long-term and to be computationally

efficient. These systems are seen as a substantial improvement on existing Bureau practice

and are comparable with international practice.

There are ongoing issues which have emerged through this study and which will be the focus

of future development and work. Foremost, there is a need to improve the daily rainfall

analyses, for which all currently available Australian analyses have less than impressive

accuracies. The evidence is that this will require either very different analysis techniques

which make use of data not currently used (such as from remote sensing and numerical

weather prediction) or a substantial improvement in the national rain-gauge network.

Page 31: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

30

7. References

Australian Bureau of Statistics 2005. Year Book of Australia 2005. Available from Australian

Bureau of Statistics, Canberra, A.C.T., Australia. 893pp

Australian Water Availability Project 2004. Work plan 2004-2006. Bureau of Rural Sciences,

12pp.

Bureau of Meteorology 2000. Climatic Atlas of Australian Rainfall. Available from

Australian Bureau of Meteorology, Melbourne, Victoria, Australia. 25pp.

Daley R. 1993. Atmospheric data analysis. Cambridge University Press, New York, U.S.A.,

457pp.

Ebert E., Janowiak J.E. and Kidd C. 2007. Comparison of near-real-time precipitation

estimates from satellite observations and numerical models. Bulletin of the American

Meteorological Society, 88, 1-18.

Holton J.R. 1992. An Introduction to Dynamical Meteorology. Academic Press, San Diego,

U.S.A., 511pp.

Hulme M. and New M. 1997: Dependence of Large-Scale Precipitation Climatologies on

Temporal and Spatial Sampling. J. Climate, 10, 1099-1113.

Hunter R.D. and Meentemeyer R.K. 2005. Climatologically aided mapping of daily

precipitation and temperature. Journal of Applied Meteorology, 44, 1501-1510.

Hutchinson M.F. 1995. Interpolating mean rainfall using thin plate smoothing splines. Int. J.

Geog. Inf. Systems, 9, 385-403.

Koch S.E., DesJardins M. and Kochin P.J. 1983. An interactive Barnes objective map analysis

scheme for use with satellite and conventional data. J. Climate and Appl. Met., 22, 1487-1503.

Jeffrey S.J., Carter J.O., Moodie K.B. and Beswick A.R. 2001. Using spatial interpolation to

construct a comprehensive archive of Australian climate data. Env. Model. and Software, 66,

309-330.

Jones D.A. and Trewin B.C. 2000. On the description of monthly temperature anomalies over

Australia. Australian Meteorological Magazine, 49, 261-276.

Jones D.A. and Trewin B.C. 2002. On the adequacy of historical Australian daily temperature

data for climate monitoring. Australian Meteorological Magazine, 51, 237-250.

Page 32: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

31

Jones D.A., Wang W., Fawcett R. and Grant I. 2006. The generation and delivery of Level-1

historical climate data sets. Australian Water Availability Project Milestone Report. 32pp.

Jones D.A. and Weymouth G. 1997. An Australian monthly rainfall data set. Technical

Report No. 70, Bureau of Meteorology, Melbourne, Australia. 19pp.

Kalnay E., Kanamitsu M., Kistler R., Collins W., Deaven D., Gandin L., Iredell M., Saha S.,

White G., Woollen J., Zhu Y., Chelliah M., Ebisuzaki W., Higgins W., Janowiak J., Mo K.C.,

Ropelewski C., Wang J., Leetmaa A., Reynolds R., Jenne R. and Joseph D. 1996. The

NCEP/NCAR 40-year reanalysis project. Bull. Amer. Meteor. Soc., 77, 437-471.

Miller P.A. and Benjamin S.G. 1992. A system for the hourly assimilation of surface

observations in mountainous and flat terrain. Monthly Weather Review, 120, 2342-2359.

Mills G.A., Weymouth G., Jones D., Ebert E.E., Manton M., Lorkin J. and Kelly J. 1997. A

national objective daily rainfall analysis system. Bureau of Meteorology Development Report

No. 1, Bureau of Meteorology, Melbourne, Australia. 30pp.

Murray Darling Basin Commission 2007. River Murray System – Drought Update No. 7. 8pp.

Available from http://www.mdbc.gov.au/rmw/drought_updates .

Murray F.W. 1967. On the computation of saturation vapour pressure. J. Appl. Meteor., 6,

203-204.

Puri, K., Dietachmayer, G.S., Mills, G.A., Davidson, N.E., Bowen, R.A. and Logan, L.W.

1998. The new BMRC Limited Area Prediction System, LAPS. Australian Meteorological

Magazine, 47, 203-223.

Rayner D., Moodie K., Beswick A., Clarkson N. and Hutchinson R. 2004. New Australian

daily historical climate surfaces using CLIMARC. Available from Department of Natural

Resources, Mines and Energy, Queensland. Brisbane, Queensland, Australia. 76pp.

Seaman R.S. 1989. Tuning the Barnes objective analysis parameters by statistical

interpolation theory. J. Atmos. Ocean. Tech., 6, 993-1000.

Trewin B.C. 2005. A notable frost hollow at Coonabarabran, New South Wales. Australian Meteorological Magazine, 54, 15-21.

Watkins A.B. 2005. The Australia Drought of 2005. WMO Bulletin, 54, 156-162.

Weymouth G., Mills G.A., Jones D.A., Ebert E.E., Manton M.J. 1999. A continental-scale daily rainfall analysis system. Australian Meteorological Magazine, 48, 169-179.

Xie P., Yatagai A. and Chen C. 2007. A gauge-based analysis of daily precipitation over East

Asia. J. Hydrometeorology, in press.

Page 33: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

32

Appendix: Climatological Maps for 1971-2000

(a)

(b)

Figure A1: The monthly average maximum temperature for January (a) and July (b). The units are ºC.

Page 34: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

33

(a)

(b)

Figure A2: The monthly average minimum temperature for January (a) and July (b). The units are ºC.

Page 35: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

34

(a)

(b)

Figure A3: The monthly average 9am surface vapour pressure for January (a) and July (b). The units are hPa.

Page 36: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

35

(a)

(b)

Figure A4: The monthly average 3pm surface vapour pressure for January (a) and July (b). The units are hPa.

Page 37: Climate Data for the Australian Water Availability Projectdata.daff.gov.au/brs/brsShop/data/awapfinalreport200710.pdfAvailability Project Australian Water Availability Project Final

36

(a)

(b)

Figure A5: The monthly average rainfall for January (a) and July (b). The units are mm.