Quantifying the uncertainty in spatially- explicit land-use model predictions arising from the use of substituted climate data Mike Rivington, Keith Matthews.

Quantifying the uncertainty in spatially-explicit land-use model predictions arising from the use of substituted climate data

Mike Rivington, Keith Matthews and Kevin Buchan.Macaulay Institute,

Craigiebuckler, Aberdeen, Scotland.

[email protected]

MODSIM 2003, Townsville, Australia.

Land Allocation Decision Support System

“There is a serious limit on the application of agricultural, hydrological and ecosystem models if [appropriate spatially and temporally representative] biophysical data are not directly available”

- Gerrit Hoogenboom (2000)

Data +Parameters

Results

Rationale for quantifying uncertainty

Land use models often have a fundamental requirement for biophysical data:

• Weather data is highly spatially and temporally variable• Soils data can be highly spatially variable and temporally variable

with management

• Use of non-representative biophysical data impacts on quality of model output– Restricts calibration and validation

Decision Support Systems often use models for predictions:• Model output will be fundamentally flawed if inappropriate

biophysical data is used

Therefore need to quantify the impacts of using ‘non-ideal’ biophysical input data on model output to determine consequences on DSS performance

Choosing an appropriate weather data source

• Knowing the decay in weather data similarity enables distance thresholds to be set beyond which alternative methods of data supply must be used

• By knowing the impacts of using alternative Met Stations on model output, it is possible to evaluate options for alternative weather data sources:– Data generators– Spatial interpolation– Calculated, i.e. solar radiation derived from max and min

temperature

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

$

R

B

W

Identifying a suitable weather data source

Met stations withsolar radiation data

Aims and Purpose

• To investigate the spatial dissimilarity in solar radiation– How the dissimilarity increases with distance from original data

source

• Determine the impacts on a cropping systems model (CropSyst) output arising from the use of substitute weather data from Met stations at increasing distance from the point of model application

Materials and Method

Spatial dissimilarity in solar radiation• Create database of daily synchronised weather data

(rain, max + min temp, solar radiation) for Met stations with +3 years solar radiation data for 24 locations

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Abd = Aberdeen (29)

Abp = Aberporth (36)

Ald = Aldergrove (25)

Alt = Altnaharra (3)

Auc = Auchincruive (19)

Avi = Aviemore (12)

Bra = Bracknell (24)

Bro = Brooms Barn (16)

Caw = Cawood (29)

Den = Denver (18)

Dun = Dunstaffnage (23)

Eas = East Malling (34)

Esk = Eskdalemuir (24)

Eve = Everton (29)

Haz = Hazelrigg (17)

Inv = Inverbervie (4)

Ler = Lerwick (47)

Loc = Loch Glascarnoch (5)

Myl = Mylnefield (25)

Rot = Rothamstead (28)

Sto = Stornoway (13)

Sut = Sutton Bonnington (26)

Tul = Tulloch Bridge (6)

Wal = Wallingford (26)


Spatial dissimilarity in solar radiation:• Create database of daily synchronised weather data

(rain, max + min temp, solar radiation) for Met stations with +3 years solar radiation data for 24 locations

• Use a search and optimisation replacement strategy for missing data

• Calculate yearly solar radiation RMSE values (1–365 daily difference in solar radiation) between site 1 and sites 2-23 for all corresponding years, i.e:


365

2EO

Where:O = observed daily solar radiation for site 1E = observed daily solar radiation for site 2

Yearly RMSE =

This process was applied for each site, i.e. - O for a fixed site - E for the nearest site, then second nearest, third nearest etc, for all sites in the data base with corresponding years data.

This produced a matrix data set from which it was possible to derive the mean, maximum and minimum RMSE for each site.


Impacts on CropSyst output• Run a standardised Spring Barley scenario within CropSyst for each

year of available weather data per site– weather data the only variables

• Compare yield estimates from one site with those from the three nearest sites for all corresponding years of available weather data

• Metrics used:– Difference in mean yield / n– Probability of means being equal– Difference in total yield / n– Absolute difference / n (sum of over and under estimates)– Size of maximum single error

(where n = number of corresponding years)

Results

Spatial dissimilarity in solar radiation• Range in rates and patterns of dissimilarity

– Geographical variation

Rothamstead Distance (km)

0 200 400 600 800 1000

So

lar

rad

iatio

n R

MS

E (

MJ/

m^2

/da

y)

0

2

4

6

8

10

Geographical variation - South-East Britain

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Altnaharra Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

^2/d

ay)

0

2

4

6

8

10

Geographical variation - Northern Britain

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Results



– Met Station network density

Wallingford Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

2/da

y)

0

2

4

6

8

10

Met Station network density

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Aberporth Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

^2/d

ay)

0

2

4

6

8

10

Met Station network density

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Results




– Length of data record

Tulloch Bridge Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

^2/d

ay)

0

2

4

6

8

10

Length of data record

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Dunstaffnage Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

^2/d

ay)

0

2

4

6

8

10

Length of data record

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Results




– Length of data record

• Nearest neighbour does not always provide the best substitute

Dunstaffnage Distance (km)

0 200 400 600 800 1000

Sol

ar r

adia

tion

RM

SE

(M

J/m

^2/d

ay)

0

2

4

6

8

10

Best substitute – not always nearest neighbour

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Impact on CropSyst output• Best substitute weather data – not always the nearest

neighbour

– At 4 sites nearest substitute provided best fit for metrics– At 2 sites second nearest substitute provided best fit for metrics– At 3 sites third nearest substitute provided best fit for metrics– Remaining sites had a range of best substitutes per metric

Results

Site Substitute Dist (km)

Mean Yield Diff / n (t/ha)

Total Yield Diff / n (t/ha)

Absolute Diff / n (t/ha)

MaxError (t/ha)

Inverbervie Aberdeen 32 -0.0068 -0.0272 0.1892 -0.33

Mylnefield 65 0.0361 0.2164 0.5084 0.89

Aviemore 102 -0.6264 -1.8791 1.8791 -5.53

Auchencruive Eskdalemuir 88 -0.1081 -1.8408 1.1547 -2.97

Dunstaffnage 121 -0.0088 -0.1591 0.5992 1.49

Aldergrove 140 -0.0029 -0.0494 0.3619 1.38

Hazelrigg Cawood 109 -0.0904 -1.3566 1.3566 -3.27

Eskdalemuir 147 -0.1805 -2.3465 2.3465 -6.24

Sutton B’ton 166 -0.07 -0.9796 1.0031 -2.09

Cawood Hazelrigg 109 0.1025 1.5375 1.5375 3.27

Sutton B’ton 111 0.0131 0.3274 0.6693 2.46

Denver 171 0.012 0.1919 0.3656 1.88

Assessment Metric

Best substitute weather source

Impact on CropSyst output• Best substitute weather data – not always the nearest

neighbour

– At 4 sites nearest substitute provided best fit for metrics– At 2 sites second nearest substitute provided best fit for metrics– At 3 sites third nearest substitute provided best fit for metrics– Remaining sites had a range of best fitting substitutes per metric

• Substitute data will introduce a minimum error of approx

+/- 0.4 t/ha • Smallest maximum error was 0.33 t/ha

– Between Aberdeen and Inverbervie (32 km apart)

• Mean maximum error was 1.90 t/ha

Results

Site SubstituteDistance

(km)Absolute diff / n

(t/ha)

Aberdeen Inverbervie 32 0.189

Aldergrove Auchencruive 140 0.342

Denver Brooms Barn 40 0.354

Cawood Denver 171 0.366

Denver Rothamstead 103 0.403

Loch Glas Aviemore 86 0.413

Aldergrove Dunstaffnage 206 0.413

Brooms Barn Rothamstead 81 0.429

Altnaharra Loch Glas 68 0.436

Bracknell Wallingford 33 0.447

Mean 0.379

Substitute sites providing the ten lowest absolute differences / n in yield (t/ha).

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Conclusions

Spatial dissimilarity in solar radiation• Clear trend of increasing dissimilarity of solar radiation

with distance BUT– Sufficient variation to exclude assumption that the nearest

substitute will provide the best alternative– Rate of dissimilarity increase is a function of:

• Met station network density• Geographical location

• Distance threshold of when to use alternative (non-observed) data source varies with location

Abp

Abd

Wal

Tul

Sut

Sto

Myl

loc

Ler

Inv

Haz

Eve

Esk

Eas

Dun

Den

Caw

Bro

Bra

Avi

Auc

Alt

Ald

Rot

Locations in dense Metstation network ofgeographically similarcharacteristics can usea higher distancethreshold

Locations in low densityMet station network withdiverse geographicalcharacteristics need a lowdistance threshold

Isolated locations shoulduse an alternative method,i.e. calculated solarradiation

Conclusions

Impact on model output• Highly complex relationship between distance to substitute data

source and model output

• Best substitute may be beyond nearest or even second nearest Met Station

• Choice of substitute has different impacts on model output, depending on which assessment metrics are used

• Substitute data can have a substantial impact on model output

• Impact needs to be quantified to enable appropriate model output interpretation

Thank you for your attention

website:

www.macaulay.ac.uk/LADSS

Quantifying the uncertainty in spatially- explicit land-use model predictions arising from the use of substituted climate data Mike Rivington, Keith Matthews.

Documents