RT6/WP6.2 - Linking impact models to probabilistic scenarios

of climate

Deliverable D6.9

Report on an intercomparison study of

modelled, Europe-wide forest fire risk for

present day conditions

C. Giannakopoulos*, P.LeSager, E. Kostopoulou

National Observatory of Athens

Greece

A. Vajda, A. Venäläinen

Finnish Meteorological Institute

Finland

Responsible partner: National Observatory of Athens (NOA)

*email: cgiannak@meteo.noa.gr

Contract number GOCE-CT-2003-505539

http://www.ensembles-eu.org

2

1- Introduction

Several fire indices have been used to estimate forest fire risk. However, their

construction varies widely from one index to another, reflecting different approaches. In

addition, the reliability of an index may depend on the region where it is applied, since

indices or their readings are usually fine tuned for specific regions of interest. In the

present study, two indices are considered: the Canadian Fire Weather Index (FWI) and

the Finish Forest Fire Index (FFI). The Canadian FWI depends on temperature,

precipitation, relative humidity and wind measurements, while the Finnish FFI relies on

potential evaporation, precipitation and snow coverage.

The two indices are compared through inter-correlation and their ability to define the

beginning and end of the fire season is assessed. The indices skills are evaluated over the

Mediterranean region using real fire observations from Italy for the period 1984-2001.

The performance of the indices is also estimated in boreal forest environments. Daily

forest fire indices (FFI and FWI) have been computed for three selected locations in

Finland over the fire season (April-September).

2- Brief description of the fire indices

The Canadian Fire Weather Index (FWI) is based on weather readings taken at noon

standard time and rates fire danger at the mid afternoon peak from 2:00 – 4:00 pm.

Meteorological variables required are:

� Air temperature (in the shade)

� Relative Humidity (in the shade)

� Wind speed (at 10 m above ground, averaged over at least 10 minutes)

� Rainfall (for the previous 24 hours)

The FWI System consists of six components: three fuel moisture codes (Fine Fuel

3

Moisture Code, Duff Moisture Code, Drought Code) and three fire behaviour indices

(Initial Spread index, Build Up Index, Fire Weather Index). Calculation of the index

requires previous day records of the fuel moisture codes. FWI is divided into four fire

danger classes:

o Low 0 – 7

o Medium 8 – 16

o High l7 – 31

o Extreme > 32

The calculation of the Finnish Forest Fire Index (FFI) is based on surface moisture

estimation and requires the following input:

� Potential evaporation from 24-hours centred on time of calculation

� Accumulated precipitation for 24 hours

� Flag specifying the presence or absence of snow cover

Previous day records of the above are required to compute the index. FFI is divided into

three fire danger classes:

o Low 1 – 4

o Medium 4 – 5

o High > 5

It is noteworthy that both indices reflect similar weather conditions, since relative

humidity, temperature and wind can be combined to determine evaporation (Singh and

Xu, 1997). Both indices depend on previous day conditions regarding one or more of

their components and they both define fire danger classes.

3- Correlation between fire indices in the Mediterranean

A subset of ERA-40 reanalyses meteorological data is used to compare FWI and FFI

through correlation on a wide domain. The subset of data is centred over the

Mediterranean region, on a 1° by 1° grid, and consists of 6-hourly data of temperature,

4

precipitation, evaporation and wind for the year 1961. Daily values of the variables have

been utilised to compute the fire indices. Relative humidity required for computing FWI

is estimated by the Romanenko equation as suggested by Singh and Xu (1997). No snow

coverage has been considered when computing FFI. In all following correlations, the sea

locations have been neglected.

Considering all locations and all days, the average correlation coefficient (r) between FFI

and FWI equals 0.75, reflecting high correlation between the two indices. Figure 1 shows

the daily correlations between the two indices for the entire region. The noticeable

positive trend might be due to differences in the spin up applied to index computation

(one year for FWI, none for FFI) and suggest that the first 100 days of the year show

biased lowered correlations. Preliminary investigation showed no dependence of FFI on

the spin up period, but further analysis is needed to confirm this statement. Moreover, the

particularly low fire risk in winter possibly makes the correlation unreliable.

Figure 1. Daily correlations between FWI and FFI for the Mediterranean region for the

year 1961

5

To get a clearer picture, a spatial analysis was also performed. Figure 2 shows the local

correlation coefficients (FFI/FWI) for two time periods: full year and summer. It shows

that correlation decreases where there is no fire risk, and increases in regions known for

fire occurrences. This is particularly evident in Figure 3, which shows the local ‘r’

between danger classes derived from both indices. Correlation is undefined in areas with

no fire risk all year around, and no colour shading is shown.

Figure 2. Correlation coefficients between time series of FFI and FWI at each location of

the data set. Top: full year, bottom: summer

6

Figure 3. As in Figure 2 only here the time series of danger classes derived from FFI and

FWI are presented

Figure 4 shows the time series at two different locations, one with and one without fire

risk. The time series of the indices in the first graph (Figure 4a) are characterised by

similar behaviour as opposed to the second graph (Figure 4b), which suggests that in

cases with very low fire risk, a spin up may be needed when computing FFI. These

findings confirm that regions (or time period) with low fire risk show low correlation (if

any) between FWI and FFI. The lack of a spin up period when computing FFI lowers

even more the correlation.

7

Figure 4. Typical time series of FFI and FWI in regions with (a) and without (b) fire risk

(a)

(b)

8

3.1 - Comparison of Fire Indices against observations

In this section, regional fire observations (burnt areas, number of fires) from 15 Italian

stations (Figure 5) are compared with fire indices derived from meteorological

observations at the same locations. Data are available for the period 1984-2001.

Figure 5. Locations of the studied stations in Italy

Monthly correlations were estimated between real data and the fire indices and presented

in Figure 6. Blue lines show the correlation between the two indices and confirm the

findings from previous section. Note that a spinning is automatically applied to FFI, since

more than one year of data is available. The trend detected in Figure1 is not shown in the

graphs of Figure 6. The black/red lines represent the correlation between FWI/FFI with

the number of fires. As expected, the correlations are quite low revealing the large

uncertainties involved in fire predictions. None of the indices performs particularly well

at predicting fire, since high fire risk is not always associated with fire occurrence.

External factors, such as human behaviour (land use policy, fire suppression strategy etc)

9

are not taken into account in the indices computation. Similar results are obtained for the

monthly correlation between areas burnt and fire indices (Appendix: Figure 13). Figure 7

shows that areas burnt and fire indices are not correlated in any year. Annual correlations

for the number of fires are also low (Appendix: Figure 14).

Figure 6. Monthly correlations between number of fires and fire indices (FWI: black,

FFI: red) and between the two fire indices (blue) for each of the 15 Italian stations

10

Figure 7. Annual correlations between areas burnt and fire indices (FWI: black, FFI: red)

and between the two fire indices (blue) for each available year

3.2 - Fire Indices Comparison: Determining Fire Seasons

Based on the Italian data, the skill of both indices was assessed in terms of their ability to

define fire seasons. According to Good et al. (2005) and Moriondo et al. (2006), defining

fire seasons with FWI is more robust than with temperature and allows avoiding false

alarm. The method consists of:

• Calculation of 7-day running averages to smooth daily variability

• Use FWI=15 as threshold for beginning and end of the season

• Defining the beginning of fire season as 2 consecutive weeks with FWI>15 (start-

11

period)

• Defining the end of fire season as 4 consecutive weeks with FWI<15 (end-period)

The four parameters (threshold, running average period, start- and end-period) were

further analysed so as to become applicable for use in FFI as well. Figure 8 provides the

fire seasons according to both indices at station code1544, and for a set of season

parameters. The y-axis represents day numbers. A threshold between 4.5 and 5 for FFI

gives results similar to FWI (as expected by the danger classes). A running average

period between 5 and 7 days seems to be more reliable. Under these conditions, the FFI-

derived season begins too early in the early 90’s and around 2000. Both indices

determine the end of the season at the end of the year 1996. This is definitively too late,

and the period used to define the end of the season may need to be shorter.

Figure 8. Fire seasons according to FWI (black) and FFI (grey) for location code 1544.

In all cases, start- and end-periods are 14 and 28 days respectively. In each column, a

different running average period is applied to both FWI and FFI. In each row, a different

threshold for FFI is used to define the fire period (it is fixed to 15 for FWI)

12

Figures 15 to 17 in the Appendix provide some preliminary results as far as the

parameters of the fire season are concerned. They suggest shorter end-period (approx. 21

days instead of 28) and longer start-period (approx. 21 days instead of 14) might provide

more reliable seasons. It becomes notable by comparing Figures16 and 17 that the season

end depends on the season beginning. Figure 9 shows the fire season at a different

station, where FFI performs better than FWI at defining fire season. FWI cannot identify

any fire season during the period 1993-1995. On the other hand, the fire season around

2000 is too long with FFI (the full year!) something that FWI does not show. Additional

investigation is needed to understand the reasons for such differences in this station. The

dependence on location calls for a systematic analysis of fire season using larger datasets

such as the ERA-40 reanalysis data.

Figure 9. As in Figure 8 only here the results for location code 1531 are presented

13

4- Temporal and spatial variation of fire indices in Finland.

Daily forest fire indices (FFI and FWI) have been computed for three locations in Finland

(Figure 10). The research focused upon the fire season (April-September) over the period

1961-1997 for Helsinki-Vantaa and Jyväskylä and over 1961-1976 for Sodankylä.

Figure 10. Locations of the studied stations in Finland

Considering the occurrence of days with forest fire index value ≥3 during the period

1961-1997 (Helsinki-Vantaa, Jyväskylä) and 1961-1976 (Sodankylä) the two indices

indicate roughly similar occurrence frequency with a gradual decrease in dissimilarities

and in fire risk conditions northwards: between 37-45% of total number of days in fire

season (April-September) in Helsinki-Vantaa and 20-21% in Sodankylä. This spatial

pattern of fire risk can be explained by the decrease of potential evaporation northward

and by the decrease in the length of fire period (due to the longer duration of snow cover)

in Lapland compared with southern Finland.

The annual variation of differences between the FFI and FWI values was considerable at

each location (Figure 11). The occurrence of both index values above the defined

10oE 20oE 30oE 40oE

60oN

70oN

Helsinki−Vantaa

Jyväskylä

Sodankylä

0 500 1000 1500

14

threshold (moderately dry for FFI and high risk for FWI) coincides in most of the cases

(56-66% of total occurrence). The number of days when only the FFI indicate decreased

moisture conditions is 23-25% of total count. The number of days with high fire risk

indicated only by FWI varies considerably among the three locations tending to increase

northwards (8.9% for Helsinki Vantaa and 20.7% for Sodankylä). On a monthly basis the

values indicate good resemblance of the two fire risk rating indices in summer season

(May-August), while during April the occurrence of FWI “high” fire risk exceed

significantly that of FFI “moderately dry” conditions for all the studied stations. The

apparent south-to-north as well as the monthly variation in the occurrence of the high FFI

and FWI values suggests a slight overestimation of high and extreme fire risk for humid

weather condition by the FWI index, i.e. in Finnish Lapland, particularly in early spring

and autumn period.

Figure 11. The temporal variation of occurrence of days with FWI indicating high and

extreme fire risk and FFI ≥3.

Helsinki-Vantaa

0

20

40

60

80

100

120

140

1961 1964 1967 1970 1973 1976 1979 1982 1985 1988 1991 1994 1997

Year

No

. o

f cases

only FWI

only FFI

FWI and FFI

Jyväskylä

0

20

40

60

80

100

120

1961 1964 1967 1970 1973 1976 1979 1982 1985 1988 1991 1994 1997

Year

No

. o

f cases

only FWI

only FFI

FWI and FFI

Sodankylä

0

20

40

60

80

100

1961 1963 1965 1967 1969 1971 1973 1975

Year

No

. o

f cases

only FWI

only FFI

FWI and FFI

15

The daily variation of FFI and FWI was analysed in comparison with the variation of fire

risk defining meteorological parameters for the 1997 fire season at Helsinki-Vantaa

station (Figure 12). In this case May represents the threshold, where FFI and FWI values

reach the “high risk” and “extreme risk” classes in most days, indicating the start of fire

season. In general, FWI shows a more pronounced, quick response to the variation of

precipitation events, whilst FFI indicate smoother changes. The largest deviations

between the two indices were observed particularly at the beginning (April) and at the

end (September) of the study period. In early April, shortly after the snowmelt, the soil

moisture is high and it is partly enhanced by the frequent rainfall events and the low daily

mean temperature (Figure 12a, b). Such meteorological conditions are usually associated

with low fire danger. It seems that the Canadian method has some deficiencies in

predicting fire risk for this period, with considerable variations between low and high fire

risk classes. Furthermore the FWI system shows a systematic overestimation of extreme

fire risk compared to the FFI system: during the 1997 fire period none of extreme fire risk

(FFI=6) and only 22 cases of high fire risk (FFI=5) were predicted by the Finnish index,

while the Canadian index predicted extreme fire risk in 34 cases (Figure 12c, e). During

the studied period, the Fine Fuel Moisture Code (FFMC) values ranges from 27.3 to 93,

with a mean of 79.5. About 47% of observed values were ranged in the high and extreme

classes (Figure 12f). The relationship between the FFMC and flammability is highly non-

linear, thus small variations in FFMC at the high end of the scale represent important

variations in flammability (Harrington et al., 1983). This suggests that almost half of days

during the fire season demonstrate conditions suitable for fire occurrence. The Drought

Code (DC) responds slowly to environmental changes. In spring, its initial value depends

on the intensity of precipitation occurred during the preceding autumn and winter

(Dimitrakopoulos and Bemmerzuk, 1998), and the value is at its lowest (DC=4). As the

season progresses DC begins rising, and reaches its largest values in autumn (DC=333),

which range within the “high risk” class (Figure 12g).

16

Figure12. The variation of daily total precipitation (a), mean air temperature (b), Finnish

Forest Fire Index (c), relative humidity (d), Canadian Fire Weather index (e), Forest Fuel

Moisture Code (f) and Drought Code (g) at Helsinki-Vantaa for the period April-

September 1997

FFMC

20

40

60

80

100

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Co

de v

alu

e

DC

0

50

100

150

200

250

300

350

400

450

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Co

de v

alu

e

FWI

0

1

2

3

4

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Ind

ex v

alu

e

FFI

0

1

2

3

4

5

6

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Ind

ex v

alu

e

Precipitation

0

5

10

15

20

25

30

35

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Pre

cip

itati

on

(m

m)

Temperature

-5

0

5

10

15

20

25

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Tem

pera

ture

(°C

)

Relative humidity

0

20

40

60

80

100

97/4/1 97/4/21 97/5/11 97/5/31 97/6/20 97/7/10 97/7/30 97/8/19 97/9/8 97/9/28

Date

Rela

tive h

um

idit

y (

%)

b

c d

e

g

f

c

a b

d

e f

g

17

The spatial distribution tendencies of the daily FFI and FWI indices are generally similar

(Figure18 Appendix), although several differences can be observed in local features. For

example, slightly overestimated fire risk by FFI compared to the FWI values is observed

during June 22-23 in Lapland. In general, significant deviations occur in prediction of the

moderate fire danger conditions i.e. the regions marked with moderate risk by FWI

correspond to moderately wet conditions (FFI=3) by the Finnish method that indicates an

underestimation of fire risk by FFI.

5- Conclusions and future plans

In general, FWI and FFI determine a fairly similar fire risk for a set of weather readings

(‘r’ approx. 0.7). Higher correlations are found especially for locations under significant

fire risk. The results improve for the lower values of fire risk if a spinning period is used

in the computation of FFI. Similar results were obtained for a boreal forest environment.

Both indices show similar features especially during summer, but some deviations are

typical during early spring and autumn, as FWI probably overestimates the fire risk. The

comparison with meteorological parameters revealed a quick response of this index to

environmental changes, especially to rainfall events.

It needs to be highlighted that fire indices do not constitute fire predictors, however they

can be used to define the fire season as an advanced alternative to temperature-based

determination. It is found that fire-season determination based on indices is satisfactory

but needs to be fine-tuned. Namely, cautious adjustment of the start and end periods is

required. Neither FFI nor FWI provided a robust definition at any location. Hence, a

systematic study using ERA-40 data is planned to be the topic of a future study, where

FFI, FWI, and particularly their components will be investigated in relation to weather

data.

6- Acknowledgements Special thanks go to Drs Marco Bindi and Marco Moriondo from the University of

Florence, Italy for the provision of Italian stations forest fire data.

18

References

Dimitrakopoulos A, Bemmerzuk A, 1998. Evaluation of the Canadian Forest Fire Danger

Rating System (CFFDRS) and the Keetch-Byram Index (KBDI) in the

Mediterranean Climate of Greece. III International Conference on Forest Fire

Research. 14th

Conference on Fire and Forest Meteorology. Vol Y. 995-1009.

Good P, Moriondo M, Giannakopoulos C, Bindi M, 2005. The meteorological conditions

associated with extreme fire risk in Italy and Greece: relevance to climate

model studies. Climate Research, submitted.

Harrington J, Flannigan M, Van Wagner C, 1983. A study of the relation of components

of the Fire Weather Index to monthly provincial area burned by wildfire in

Canada 1953-80. Canadian Forest Service, Chalk River, Ontario. Information

Report PI-X-25.

Moriondo M, Good P, Durao R, Bindi M, Giannakopoulos C, and Corte Real J, 2006.

Potential impact of climate change on forest fire risk in Mediterranean area,

Climate Research, Special issue 13, Vol. 31, 85-95.

Singh V, Xu C, 1997. Sensitivity of mass transfer-based evaporation equations to errors

in daily and monthly input data, Hydrological processes, 11, 1465-1473.

19

Appendix

Figure13. Monthly correlations between areas burnt and fire indices (FWI: black, FFI:

red) and between the two fire indices (blue) for each of the 15 Italian stations

20

Figure 14. Annual correlations between number of fires and fire indices (FWI: black,

FFI: red) and between the two fire indices (blue) for each available year

21

Figure 15. As in Figure 8 here with a smaller period to define the end of the season (14

instead of 28 days)

22

Figure 16. As in Figure 8 here with a smaller period to define the end of the season (21

instead of 28 days)

23

Figure 17. As in Figure 16 here with a longer period to define the beginning of the

season (21 instead of 14 days)

24

Figure 18. Spatial variation of the Finnish Forest Fire Index values (a) and the Canadian Fire Weather Index values (b) for the period

16-25th of June, 2005.

17.06.2005 18.06.2005 19.06.2005 20.06.2005

21.06.2005 22.06.2005 23.06.2005 24.06.2005 25.06.2005

16.06.2005

wetvery wet

moderately wetmoderately drydryvery dry

25

Figure 18. (continued)

17.06.2005 18.06.2005 19.06.2005 20.06.2005

21.06.2005 22.06.2005 23.06.2005 24.06.2005 25.06.2005

lowmoderate

extreme

16.06.2005

high