Evaluation of satellite-derived burned area products for the fynbos, a Mediterranean shrubland

(1)

CSIRO _PUBLISHING

@fr?B

www.pub.lish,

_csi.o.

_uuZ;ou.nlffi

International Jour@

----=\

-r l-re }}|;-ry, -r-r, oao-oo-r _\

Evafuation

_{of satetite-derived}

burned

_{area products}

for the fynbos,

_{a Medite;;**; shrubland}

Helen M. de Klerb,B,E,

Adam M. Wilsonc and Karen

SteenkampD

Stellenbosch _{University, private}

Bag X1, : Services, _{p Bag xSOl 4,stellenbosch,} rrversity

of Connecticut, 75 N. JSA.

Meraka lnstitute, pO _{Box 395,}

Additionaf keywordsl

MODIS, Sourh Africa, Western Cape. Manuscript ....tu:.d_-!

_{T uary 2llt,accepted 2 May 20tl published

online xx xxxxx xxxx Introduction

@ IAWF 2011

(2)

Int. J. IYildland Fire

_ - 4r in many Mediterranean ecosystems, f,rre in the dwarf Mediterranean shrubland typical of the ffnbos (Rebel o et al. 2006) is a vital ecological process and important for conserva_ tion management _{(Cowling 1 9g7; Richards}_{on et al.1994; Bond} and Keeley 2005). The flnbos biome, which occurs in the mainly winter rainfall region of the south_westem tip of South Africa (Rebelo et al.2006), is not only highly fire prone, but is

fit:

ldunl:lgtd_n9eds

periodic

nre to reproOu..

i..g. Keeley

and Bond 1997; Rebelo et a\.2006).The fynbos biome covers a geographically small area, but is of greai conservation imoor. taace owing to its high levels ofspecies richness and endem'ism (these aspects are thoroughly discussed by Rebelo et al. 2006). W rlson e t al. (20 I 0) and van W ilgen et a l.(20 1 0) suggest that the south-western Cape is vulnerable to changes in thJ fire regime driven by human activity and climate chaise.

Iynbos is structurally and spectrally diffJrent from grassland and miombo (Wessels e/ a/. 201 0) where most testing of satellite -bumed area algorithms has been conducted (e.g. frigg and Flasse 2001; F.oy et at.2005a,2005b; Giglio'et"at. zdOby. tn addition, the fynbos biome includes areas-of topographically complex landscapes _{where shadowing may affeci the accuracy} of the reflectance values reaching the sateilite (de Klerk 200gj. Glint from extensive coastlines and light soils can also lead to spurlous reflectance measurements (de Klerk 200g). The West_ em Cape Nature Conservation Board (CapeNatuie) manages formally protected areas of the Western Cale of South Afrla, many of which fall within the fynbos biome. Good datasets oi fire. b.oundarie^s _{mapped by operational fire agencies in the field,} which we refer to as manager-mapped fire boundaries, are available- for these protected areai lsee full description in 'Materials

and methods') (de Klerk 200g; van Wilgen and Forsyth 2008). This enabled us to test the Meraka Institute,s new MODIS bumed area product, based on the Giglio et al, (2009) algorithm released latein200g,aswell as the MCD45A1 bumed area product, based on the Roy e/ a t. (2005a) algorithm, for which data arc available on the MODIS website.

, .In this study, we compared the two MODIS_derivedproducts rn me rcrmally protected areas within the funbos and succulent karoo biomes (Mucina and Rutherford 2b0q with carefully T,tpf.ll5^. boundaries produced by reserve managers (see de Klerk ZUUE fbr details of the database). General validation methods (after Giglio et al. 2009; Roy and Boschetti 2009) were used and included (l) an inventory accuracy assessment of the number and size of individual bum scars mapped by the reference and classified datasets; and (2) a subsefof standard confusion matrix parameters, which provided a per pixel evalu-ation ofthe spatial fidelity ofa classificevalu-ation', u.curu"u.

The additional question of whether these products can be used as a,'reliable, surrogate for mapping burned areas outside of formally protected areas, where .utr*Uy there are no ana_ logue fire history records available,was also an important one, with landscape conservation initiatives on private land increas_ ing in the Westem Cape (CapeNature 200d).

Materials and methods

Study area and protected area data

CapeNature (http://www. cape nattre.co.za/,accessed 1 2 August 2011) manages formally protected areas _{of the Westem Cape of}

H. M. de Klerket al.

Burned'area _{- reference data from reserve managers}

Burned area - WAMTS/Cigtio _{algorithm datasets}

The Meraka Institute in South Africa (http://www.csir.co.za./

merckal, accessed _{12 August 2011) releasid}_{a new MODIS_}

(3)

Fynbos fire mapping with MODIS satellite products

Int. J. Ilildland Fire

0 25 50 100 150 200

-Kilometers

Fig' 1' Land managed for conservation purposes by capeNature, which includes statutory Provincial Nature Reserves (Intemational union for the conservation ofNature and Natural Resources, _{IUC\ category II), state Forests (also IUcN category II) and wilderness Areas (IUCN category I)} (solid black) as well as privately owned, proclaimed Mountai; catchment Areas (solid grey). The insert shows the location ofthe westem cape province (sofid black) in south Aftica(hashing). Dataforthis figure are in UTM 34 S (central meiidian,2l;refercncelatitude, 0; scale factor, 0.9996; false easting, 500 000; false northing, 10 000 000) and WGSS4 (Hartebeesg4) datum.

same time series of images. Information onpersistent changes in this bum-sensitive vegetation index are combined with active

South Aliican Council for Sbientific and Indushial Research (CSIR, http://www.csir.co.za, accessed l2 August 2011), Meraka Instifute's Wide Area Monitoring Information System (WAMIS) website (http://www.wamis.co.za, accessed l2 August 201 l) for the period from I September 2007 to 30 April 2009.

Burned area - MCD45A1/Roy algorithm datasets

Bumed Area Product; Boschetti et a\.2009; http://modis_fire. umd.edu/BA_methodology.html _{, accessed}_{12 August 2011). It} is based on the premise that bumed areas are characterised by

(4)

Int. J. IYildland Fire

deposits _{of charcoal and ash, removal of vegetatron}

and an altered vegetation structure _{(Roy el al. 1999). iulti_temporai} land surface reflectance data are used to tocate _{rafiO chanjes in} the daily surface reflectance _{dynamic, tfrrougfr'tfr. use of a} bidirectional _{refl ectance}_{distribution nrnctlon'lnnOF) model} (Roy et a\.2005a). The algorithm is applied iJependently to

geolocated _{500-m pixels (Boscheni}

"ih. ZOOS). A statistical measure _{of the difference}_between

the observeiBRDF values . and the predicted BRDF_values, _{per pixel, at the viewing and} illuminating angles of the obseivation, i, ur.a to quanti$u

!!1ee lrom a previously- observed state @oy et al. 2002,

2005a; Boschetti et al. 2009). Large Aiscrefancies between

predicted _{and measured}_{values are uttriUot.O}

to thange (Roy et

.al. 20t05a). A temporal constraint is used to differentiate oerween temporary changes, such as shadows, _{that are spectrally} :t:rlT from more persistent fire induced

"il** (Roy et al. 2005a).

-^"1191"::re reprojecte,d to UTM 34 S (centrat meri dian,2l; rererence latrtude,0; _scale_{factor,0.9996;false}_easting,

500 000;

false northing, 1000000O) _{and WGSg4 (Hartebeesga)}_datum,

l_.-rlTt!d to a grid cell size of 463.31m, using one of the WAMIS images as a .master _{grid' to ensure}_exact_{ci_regrstration} among all datasets. _{All area calculations}_{were performed}_in UTM 34 S, as this provided good aerial and distainse accuracy across _{the study area of the Western}_Cape.

Al1 grid cells that fell l: :::1.^^:l _rrovrnce_weret"rmally protected areas in the" Westem Cape_ignored

for.these analyses, u, ,unug"r_rappid bumed area data are available _{only inside'prot .i.i u..ur,} Eva _{I u ation of ,performa}_nce,

We were interested _{in the ability of the satellite products to} accurately _{(1) identi$r bumed areas}_{and measure}

their size; and

(2) map their boundaries. _{Our evaluation consisted of an}

'inventory'

analysis _{that identified whether u pu.ti.rtu, fi..} :u:.o1 yu: lapped by all databases and comparJd _{the sizes of} individual fires with a least_squares _{.egr.sslon, as well as a} confusion _{matrix that assessed}_geographi-c_ua.uru"y.

The inventory accuracy assessment was produced _{by com_}

paring the proportion _{ofbum scars}_{in the manager_mapped}

data ,l?1,*:* identified by the classified darasets _ltJrmeJ.detection _* ability' and calculated as the number of burn scars in the

classified datasets _{divided by the number in th. manager_}

mappe.d _{data) and by comparing}_{the sizes}

of individual burned

areas in manager-mapped _{and classified datasets}_{usrng the}

coefflrcient _{of determination}_{(R2) 1e.g. Giglio et al. 2009i Ftoy} and Boschetti 2009). _{The size comparisons"were}_calculated_first for all fires that occurred in eitheaor both datasets _{itermed .all} fire events'), _{and then for those fires that o""rrrr.J Jnty in both} datasets (termed ,shared _{fire events').}

A confusion matrix provides a geographic accuracy assess_ ment by evaluatine the snatial _{fidelity o=f}mapping _{ona per pixel} basis _{(e.g. Gigtioir at.2009; Roy and eo^r'rh.iti-ibos).}

The confusion _{matrix is a standard}_technique

used to .uropurc u cla^ssified _{image against}_{a set of reference}_{data (eg. Campbell} 1996). Although compilation of the error matrix is generally s,tr$ghtforward, preparationof _{the maps}_{for comparison}

may be difficult (Campbell 1996). There _{are many ..urir"rit ut .uo u,} oenved _{trom a confusion}_{matrix and it is suggested}

that a suite of measures be used to examine different aspects of the

H. M. de Klerk et al.

comm1ss10n.

(5)

o All fire events X Shared fire events - _{Linear (All fire events)}

B2 = 0.8697

- _Linear_(Shared_{fire events)} Rz :0.8407

5000 10000 15000 20000 25ooo 3ooo0 35000

Manager mapped fire areas (ha)

o All fire events X Shared fire events - _{Linear (All fire events)}

Rz : 0.2688

- _Linear_(Shared_{fire events)} R2 = 0.7201

r---_-,o_

-4 -3 -2

Manager mapped fire areas (log ha)

S-*;-1;" IT_"i:ion _{wAMIS product ufiith-(a)}of the area of individual fire events, mapped as bumed areas by reserye managers (reference dataset) and the_{linearand (b) log16-transformedaxes,} The least-squares regressions are presentedas solid lines. The dashed I : l. liqe indicates pe'ifect agreement,

(a)

€ zsooo

o (u E

E zoooo

& rsooo

E o s 1 0 0 0 0 (b) o q o o o o) U)

Several fireJgy,pnts _{that were mapped by the managers were not} mapped by:WAMtS (see the low vaiues of 31.-Z-ql.tyo for detection abilify, Table 1). Note that although several points lie on the l: I perfect agreement line, there are a few notable exceptions.

(6)

Int. J. Wildland Fire q E } E 5 d { o a d 6 . 9 h P 6 H 5 o \ o v h , 9 ^ 8 Y -6 O i 3 . 5 S x ! ? -d 9 . 9 4 . t r d o \ o F 3 s ' 6

- ? E a

' g B €

F O 9 o F d -o 9 2 , 1 ^ l & d -F o Jq .'-1 -1 \ c.l \? \

n + 5 N 9 S

a r € q r

e h s e e s

s F e l I $

N E N N q

P 3 S f f i E a

a H F - d

! 5 9 ; " ; 5

@ + € @

-N s s 6 3 e

H.M. deKerket al.

or where bumed areas were either under_mapped by the managers_or over_mapped by the WAMIS product,-were q-e1era]lV _{low (14.6-23.4%, Tabte 1). Arr excep?ion was 2009} (3 1.9 -32.8%o, _{Tab le 1 ), when managers}

possibly under_mapped a fire in the GrootWinterhoek by a slubstantial amount (Fig. 3e).

Low omission and commisiion and good spatial match as des.cribed _{above, led to overall good val-ues for foth producer} and consumer accuracy (which varied from Si.Z to gl.ey", Table I).

MCD45A1

ned areas conservativelv. pped an area of 711,9ha, r of only 2962ha, giving lo/o; Table l). This was rveral fire events (e.g. the b) and missed a few fire

manaser-mapped

_nres

_rhat

_MCD45A

_{il,Ji'"ffi +:i',ffii}

fl."y:ll Th*ing

low percentages

of detection

ability,

Tabte

t ).

Ine reverse scenario was seen in one instance where managers under-mapped a fire in 2009 it the GrootWinterhoek, which WAMIS also identified as being considerably larger tnan the area mapped by the managers (Fig. 3e). fne UbO+ial product seldom identified a burned area that did not burn, leading to very low euors of commission

-of g.l*19.1% ffrbfe ll uirO friei, consumer's accuracy _{of g0 .9 to 919% lfatL t;. There were no} D45A1 thatwere notalso ry be responsible for the 1.76-0.80), even though rresentrng perfect agree_ atasets.

Discussion

The high levels of omission seen in MCD45A1 compared with WAMIS (40.1-80.8%, _{v. 12.443.g%,Table 1; see fewerpoi,rts} along the x-axis in Figs 2, 4) were-expected because of the approach of the MCD45A1 product (eg. Roy and Boschetti 2009). MCD45A1 aims to map burned u..u, .onr.ruutively, in that it.specifically aims to reduce errors of ,ornrnirrion by ,only selecting.fire affected pixels where there are burn candidates that provide persistent evidence of fire occurrence, (Roy et al. 2005a), P.oy et al. (2005a) observed that this l.ud. to lure.. missed orunde.-rupp-.d

. f i r e a n d a t t h e e d g e s o stiictimjllem-entaiionrjf - (oy et al.2005a). The stng shong fire affected ich neighbouring pixels s (Roy ef aL.2005a), In axation of the temporal to reduce ,false rejec_

;#Ti??,*"fi:H::n

perhaps

_the

_adaptation

_{or the}

_tempora,

_{;il:li3,:l'r3; il?ijfli}

application, _{as suggested,}_{applied in specific cases}

(Roy et a/. 2005a), may improve on the error oi omission ral anC ttre

q6 qH 6d

K E K h S S

h N g h R E

H E $ R 8 *

@ € r r -o f + a h N r @ r ) O $ € * ' + € v r r - N N

i S N S F *

n a o N 9 -a s +

P e c * * f r

> > > x x x

* s g e g g

RHRRHR

(7)

Int. J. Wildlqnd Fire

}

I

J .

I

'..r'', ttrl'-t

r

Jrl a i 0 0 5 1 2 3 4 5 6 f f i K i t o m e t e r s , i

, A

" n

Fig' 3' Differences between burncd areas mappcd by the managers (black linc) and the WAMIS product (horizontal hashing) and the MCD454lproduct (diagonal hashing); where WAMIS maps a very sirnilar fire extcnt to the managers in (a) Jonkershoek and (6) Cedcrberg _{fires in Fcbruary of2009; (c) whcre WAMIS undermaps}_{thc fire extent rnapped}

by managcrs in the southem Cederberg firc in February 2008; and (/) whcre WAMIS misses small fires mapped by managers, _{such as the 274 ha that bumed in GrootVadersbosch}_{in October 2008. (e) A case wherc both satellitc bumed} area products (WAMIS and MCD45A1) show a larger area than mapped by the managers rs seen in thc GrootWinterhoek firc of Febmary 2009. protected Area boundaries are in white.

producer's accuracy' The algorithmt*"t"tJr\ttur _{low- com-}

2005a). Moreover, the algorithm appears to be very reliant on bustion completeness (i'e. when a cool fire doesn't burn all plant _{the posiltre presence of black ash and the authors comment that} material in the burned area) or when a fire is smaller than a _{this^ may pose a problem when ,more reflective underlying} MODIS 500-m pixel and at land-water interfaces (Roy et al. _{surfa."s 1in MoDIS bands 2, 5 and 6) are exposed by the action}

(8)

Int. J. llildland Fire

H.M. deKlerk et al.

0 0 . 5 1 2 g - 4

Fig,3. (Continued)

offire' (Roy et al.2005a). Such an instance _{will arise when the} fire occurs _{on highly reflective soil C; iib* ffi;l.ii*ilr.ty,} this may occur when hot fires produce frigfiy,-.fi;.iive white ash _{(Roy and Landmann}_2005)._{if *ioO uoO"ruildirriput,}

trr. u.t and charcoal qapidly after tire fire, the po.ffii.-l.n..r*.u values _{may be higher than the algorithm u^ntffii..}

lnoy "r ol.

2005b), flre _{\oV algorithm is more likely to detect fires on} su1fa19s _{that^have a high pre_fir" ,.n..iu*.}

ttu-n on less relecuve surtaces because it seeks large changes in reflectance from-pre-bum _state_{(mainly orory gra;; t; uii.i,_ir'tn.y,"c}

Landmann _{2005). This -uy opluin why the VfCO+Saf ufgo_}

(9)

Int. J. IYiklland Fire

, 0 0 . 5 1 _{2 3 4 5 6 7 8 I t O 1 1 t 2} ==rE--Fhr l , r ' \ t , 1 ,

product help to account for the influence ofvarious vegetation structure and fire types on the post_fire reflectance character_ istics (Giglio et aI.2009). The fact that the WAMIS product testing included a vegetation type similar to the _{$rnbos, the} Californian chaparral (Giglio er at. 20091, probably improved WAMIS performance in the fynbos.

The difference between these two bumed area products was particularly evident when considering the burned areas of large fires (e.g, Fig. 3b). Giglio et al. (2009) would probably attribute the success of the WAMIS product in mapping larger bumed areas to 'the inclusion of a region growing phase, which also permits the algorithm to function in the presence of extremely large burned areas,, as well as the accurate identification ofboth burned and unburned pixels through the incorporation ofactive fire pixels and through the use of both spectral and texfural information.

Both.products struggle to accurately map the extent of fire events ln topographical complex areas, as is seen where WAMIS under-mapped certain fires, such as in the southeur Cederberg in February 2008. Both products missed several small to medium fires that had complex boundaries and occurred in topographically complex areas, such as Grootvadersbosch (27 4ha, Fig. 3Q, Riviersonderend (1695 ha) and Limietberg (5 I 6 ha).

Conclusions

We have compared two MODIS derived burned area products with reliable field-mapped reports for the protected areas in the fynbos biome of South Africa. Our results showed that the two burned area products have different strengths; both may be useful for specific applications. The MCD4SAI had low emors

(10)

Int. J. Ilildland Fire (a) 25000 1 5 0 0 0 H. M. de Klerk er a/. 1 0 0 0 0 o o o F o (g E o

o

a

-3

-2

(b) G I o (E I (d E o (d v o

lQ(x

5000

X

All fire events Shared fire events Linear (All fire events) Rz = 7987

Linear (Shared _{fire events)} Rz = 0.775

I _ - l

20000 25000

X ,

1 0 0 0 0 1 5 0 0 0 Manager mapped fire areas (ha)

o All fire events X Shared fire events -*--- Linear (All fire events)

R 2 :0 . 3 7 1 s

- _{Unear (Shared}_{fire events)} Ra = 0.2612

Manager mapped fire areas (log ha)

tr'ig' 4' Regression _{ofthe area (ha) ofindividual fire events, mapped as bumed areas by reserve}

managers (reference dataset) and the

il:?:11J,:".lJJi#lJ?J[:ffi1(a)roe,o-transrormeiut.

rr,"r.u,,-.quu..,i.g."-,,i*i,i,",.nt"da,u,oridrine.

rhedashed

o x u #

However, the less common omission errors and improved producer's accuracy _{ofthe WAMIS product}_{are tikely to make it}

a more useful data source to supplement _{and check the field}

mapping _{of bumed areas.}

Although WAMIS shows much promise, we would not

recommend _{that it replace the gathering of manager_mapped}

fire boundaries _{within the flinbos prote-cted}

ur'"-*l Ho*.urr, WA-[4IS _{may be useful in the'fo]lowing}_contexts:

1t io r,ig],fight medium _{and large fires missed}_{by man-agers}

lbecause of gaps in :"T:Tj]T:lrlgnal lnlnsements, or fires that rained our deep ln wllderness areas before manager became _{aware of themi;}

(11)

References

Bond WJ,.Keeley JE (2005) Fire as a global ,herbivore,: the ecology and evolution of flammable ecosystem _{s, Trends in Ecologt & Evolution 20,} 387-394. doi: 10. 1016/J.TREB.2005.04.025

BoschettiL, RoyD, HoffmannAA(2009)MODIS Collection 5 BurnedArea Product-MCD45. User's Guide. Ver. 2.0. November2009. (University

2\e^r_!r2, Forest Ecologt and Management 147,67_74. doi:10.1016/ s0378_ 1 127(00)00434_5

Li 7. Fraser.R, JinJ, Abuetgasim _{AA, Csiszar I, Gongp, pu R, Hao W (2003)} Evaluation of algorithms for fire detection and mapping across Nofh America from sarellite. Journal of Geophysicat R)iealch ir}g, 4076. do i : 1 0. 1 029 / 200 | JD00 1 3 7 7

Roy_

?l Boschetti L eO09) Southern Africa validation of the MODIS, L3JRC and GlobCarbon bumed area products. IEEE Transacnons on Geoscience _{and Remote Sensing 47, 1032_1044. doi: I 0. I 1 09/TGRS.} 2008.2009000

RoVfP, Landmann T (2005) Characterizing the surface heterogeneity of fire effects using multi+emporal reflective wavelength data. Interna_ tional Journal of Remote Sensing 26, 41974t8. doi:10.10g0/ 0 t 4 3 t 1 6 0 5 0 0 1 t 2 7 8 3

Roy DP, Giglio L, Kendall JD, Justice CO (1999) Multiremporal active_fire based burn scar detection algorithm. Internqtional Joinal of Remote Sensing 20, 1 03 1 1 038. doi: I 0. 1 080/0143 tl 699213073

Roy DP, Lewis PE, Justice CO (2002) Burned area mapping using multi_ temporal moderate spatial resolution data - a bi-directional reflectance

Shlisky A, Waugh J, Gonzales p, Gonzalez M, Manta M, Santoso H, Alvarado E, Ainuddin Nuruddin A, Rodriguez_Trejo DA, Swaty R, Schmidt D, Kauflnann M, Myers R, Alencar A, Kearns F, Johnson D, Smith J, Zollner D, Fulks W (2007) Fire, ecosystems and people: threats of Maryland: College park

ampbell JB (l9%fziffiffiiion to remote sensing., (Taylor & Francis: London)

CapeNature (2009)'stewardship operational procedures manual, (Cape_ Nature: Cape Town)

Chuvieco E, Ventura G, Martin Mp, Gomez I (2005) Assessment of multi_ temporal compositing techniques of MODIS and AVHRR images for burned land mapping. Remote Sensing of Environment 94, 45C_462. doi: 1 0. 10 16lJ.RSE.2004. I 1.006

Diaz-Delgado _{R, Pons X (2001) Spatial pattems offorest fires in Catalonia} (NE Spain) along the period 197 5-1995 : analysis of vegetation recovery

(12)

and strategies for global biodiversity conservation. Global Fire Initiative Technical Report 2007_2. The Nature Con.ervancy. (Arlington, vA) Available at http://www.nature.org/initiatives/fire/ .,,,|,t:r1fir"^_...orystems_and*people.pdf [Verifi ecl 7 funuiry.-Ot t1 shllslry A, Alencar AA, Nolasco MM, Cunan LM (2009) Overvrew: global

fire regime conditions, threats, and opportunitie; for fire management in the tropics. In ,Tropical Fire Ecology. _{6C. Vfe Cociranejpp. OS_A:.} (Praxis Publishing Ltd: Chichester, UK)

SukhininAl, French NHF, Kasischke ES, Fiewsoir JH, Soja AJ, Csiszar IA, Hyer EJ, Loboda T, Conrad SG,. Romasko Vl, pjichento fa, Vist<iv SI, Slinkina OA (2004) AVHRR-basedmapping of fires in Russia; New products for fire management _{and carbon cycle-studie}_{s, Remote}

Sensing of Environment 93, 546_s64.doi: _10._{1Ol6/i.RSE.2004.08.01}₁ Trigg S, Flasse S (2001) An evaluation of different bispectrai spaces fo,

discriminaling burned shrub-savann ah, _{International Journal of Remote} Sensing 22,2641-2647. doi;10.1080/0143 I I 601 10053185

(13)

A U T H O R QUERTES

A!!: We can keep this as 'lo916'; however, it need only be 'log' (as base 10 is assumed). ,ln' is the natural logarithm. AQ2: Please confirm the place of publication.