Transcript MODIS Active Fire Validation Wilfrid Schroeder Ivan Csiszar Louis Giglio
MODIS Active Fire Validation
Wilfrid Schroeder
(Univ. of Maryland, NOAA/NESDIS/STAR, Camp Springs, MD)
Ivan Csiszar (
NOAA/NESDIS/STAR, Camp Springs, MD
) Louis Giglio (
Science Systems and Applications, Inc., Lanham, MD
) Chris Justice (
Univ. of Maryland, College Park, MD
)
MODIS Science Team Meeting, 27 th Jan 2010 Land Breakout Session
Background
Sample Size:
18 ASTER scenes
Region:
South Africa
Proof of concept using fixed threshold method applied to ASTER band 9 to derive 30m resolution active fire masks Morisette et al. 2005 Sample Size:
131 ASTER scenes
Region:
Northern Eurasia
Development of active fire validation protocol Csiszar
et al
. 2006 Sample Size:
100 ASTER scenes
Region:
Global
Development of robust active fire detection algorithm for ASTER Giglio
et al
. 2008
Background
Sample Size:
115 ASTER scenes
Region:
CONUS
Validation of NOAA/NESDIS operational fire monitoring system including analyst data Schroeder
et al
. 2008 Sample Size:
167 ASTER + 123 Landsat ETM+ scenes
Region :
Brazilian Amazonia
Generalization of moderate-coarse resolution fire data validation (MODIS + GOES) using higher resolution imagery Schroeder
et al
. 2008 Sample Size:
24 ASTER + 8 Landsat ETM+ scenes
Region :
Brazilian Amazonia
Assessment of short-term variation in fire behavior – implications to active fire validation Csiszar
and Schroeder
2008
Current Status
Sample Size: ~
2500 ASTER scenes
Region :
Global
Stage III validation of MOD14 Schroeder
et al
. (in preparation) • Daytime & nighttime data • Data equally distributed across the globe • Multi-year analysis (2001-2006) • ASTER SWIR anomaly May ‘07 • Omission/commission errors derived as a function of percent tree cover
Temporal Consistency of MOD14 Detection Performance Using a subset of points covering the range of 20-40% tree cover No statistically significant difference over time (i.e.,
D t
= 0;
p
< 0.01) 100 80 60 40 20 0 01<>20 2001-2002 2003-2004 2005-2006 41<>60 81<>100 121<>140 161<>180
Fire Cluster Size (number of 30m ASTER fire pixels)
>200
ASTER (RGB 8-3-1) 26 Jan 2003 00:09:09UTC SE Australia
ASTER (30m Fire Mask) 26 Jan 2003 00:09:09UTC SE Australia
Overall Probability of Detection Summary curve using all data points (125K MODIS pixels with >0 ASTER fire pixels including16K MOD14 fire pixels) 40 30 20 10 0 0 100 90 80 70 60 50 100 200 300 400 500 600 Fire Cluster Size (number of 30m ASTER fire pixels) 700 Day Night 800 900
Daytime Probability of Detection as a Function of Percentage Tree Cover** 100 80 60 40 20 0 0 50 100 150 200 250 300 350 400 Fire Cluster Size (number of 30m ASTER fire pixels) vcf <20 vcf 20<>40 vcf 40<>60 vcf >60 450 500 All ** average value calculated using a 20x20km window centered on the target pixel
ASTER (RGB 8-3-1) 21 June 2003 17:38:35UTC Manitoba, Canada
ASTER (30m Fire Mask) 21 June 2003 17:38:35UTC Manitoba, Canada
Commission Errors as a Function of Percentage Tree Cover** No nighttime commission error (
n
=722) 18 16 14 6 4 2 0 12 10 8 <15 With Scars W/O Scars 2% overall fire-unrelated false alarm rate 15<>30 30<>45 45<>60 VCF Range (% tree cover) 60<>75 >75 ** average value calculated using a 20x20km window centered on the target pixel
Daytime Commission Errors as a Function of Land Cover Type** (IGBP classes) 16 14 12 10 14.3 14.3
7.7
8.9
8 6 4 2.6
2.0
2 0.0
0.0
0.0
0.0
0.1
0.2
0.3
0.5
0.6
0.7
1.0
0 Eve rg re en D N eci ee du dl ou e s le N af ee Pe fo dl rm Ba re e an rr st le en en af t sn o fo re r sp st ow ar a se nd ly ice ve ge ta te d G ra W ssl oo an dy ds sa O va pe nn n as sh ru bl an ds Sa va D nn C C as eci ro lo pl se du an d ou ds sh s ru Br bl oa an d ds le af fo re st C ro pl an d/ N at ur al ve ge ta tio U n rb m an osa a Eve ic nd rg b re ui M lt en up ixe B d ro fo ad re st -le af fo re Pe st rm W an at en er t w et la nd s ** predominant class using a 20x20km window centered on the target pixel
Daytime Commission Errors as a Function of Land Cover Type** (IGBP classes) 70 60 50 40 58.8
30 20.0
20 12.6
10 7.9
0.0
0.0
0.1
0.2
0.3
0.3
0.6
0.6
0.9
1.0
1.1
1.3
3.1
0 ow Pe rm Ba an rr en en t sn o r sp ar a se nd D ly ice ve eci O ge pe du ta ou n te s sh N d ru ee bl dl an D e ds le eci af du fo ou re G s st ra Br ssl oa an d Eve ds le W rg af oo re fo dy en re st sa N va ee dl nn e as le af fo re C st Sa lo va se nn d as sh ru bl an ds C ro pl an d/ C N ro at pl ur an M al ds ixe ve Eve rg d ge fo re ta re en st tio B n ro m osa ad -le U ic af rb fo an re a Pe st nd rm b ui an lt en up t w et la nd s W at er ** point value representing the target pixel
Quality Check – Visual Inspection
Typical false detection MODIS/Terra
False alarms can occur more than once at the same location Some burn scars may also affect the Cloud Mask & LST products
Path Forward
• Development of Landsat-5 TM active fire masks to evaluate MODIS/Terra fire data over far off nadir scan angles – Problems with TM data quality must be addressed (radiance bleeding from adjacent fire pixels) • • Use of airborne sensor data – Alternative to orbital sensors – Quality data enabling fire characterization analyses – Potential gap filler : final link between Landsat-class data and surface observations • Provide key insight on the relationship between Landsat-class fire pixels and active fire area (ha, m assessment) 2 , ...) – Possibility for sequential mapping of prescribed/wild fires (ideal for diurnal cycle Reproducing MODIS fire pixel data using ASTER imagery – Potential for fire characterization validation : applicability must be evaluated using reference airborne and field data – Retrospective analysis of large volume of ASTER and MODIS/Terra data : fine look at fire characteristics across different biomes
Landsat-5 TM (RGB 7-5-2)
Landsat-5 TM (Fire Mask)
Path Forward
• Development of Landsat-5 TM active fire masks to evaluate MODIS/Terra fire data over far off-nadir scan angles – Problems with TM data quality must be addressed (radiance bleeding from adjacent fire pixels) • • Use of airborne sensor data – Alternative to orbital sensors – Quality data enabling fire characterization analyses – Potential gap filler : final link between Landsat-class data and surface observations • Provide key insight on the relationship between Landsat-class fire pixels and active fire area (ha, m assessment) 2 , ...) – Possibility for sequential mapping of prescribed/wild fires (ideal for diurnal cycle Reproducing MODIS fire pixel data using ASTER imagery – Potential for fire characterization validation : applicability must be evaluated using reference airborne and field data – Retrospective analysis of large volume of ASTER and MODIS/Terra data : fine look at fire characteristics across different biomes
Reproducing MODIS Fire Pixel Radiance
Fire Pixel Radiance = [ ASTER Sfc Temp, ASTER Fire Mask, MODIS (PSF + SRF), (Atm + Solar) ]
Reproducing MODIS Fire Pixel Radiance
•
Concluding Remarks
Increased capacity to ingest and co-locate different datasets – ASTER, ETM+, TM, CBERS, Airborne imagery used successfully in combination with MODIS data – Optmized use of NASA & international assets (multi-sensor/satellite data integration/fusion) – Efficient data mining codes enabling manipulation of large volume of higher resolution imagery data and active fire information from MODIS • Capacity building towards development/application of sensor networks and next generation datasets – Great potential for transition of research methods/techniques/science codes into operations through NOAA/NESDIS – VIIRS and GOES-R in advantageous position in regards to active fire data validation • Protocols being developed • Field campaigns and fine resolution airborne data still an important component in the validation of active fires – Inter-agency collaboration/coordination is needed (involvement of USFS and other state agencies) – Progress with fire characterization depends on the successfull implementation of field work