of nine different regional algorith

of nine different regional algorithms applied to 1-km VGT daily
surface reflectance imagery. Roy et al. (2002, 2005b) developed a
predictive bi-directional reflectance modeling approach to map
burned areas on a daily basis using 500-m MODIS imagery. Most
recently, Tansey et al. (2008) modified one of the regional GBA-2000
algorithms for global use to produce the L3JRC 1-km global burned
area product for 2000–2007.
Although the majority of existing burned-area mapping methods
do not exploit active fire information, a minority are hybrid algorithms
which supplement the “standard” remotely-sensed indicators used for
burn mapping (surface reflectance, surface temperature, NDVI, etc.)
with active fire maps. Roy et al. (1999), for example, used AVHRR data
to map savanna burns in southern Africa from a temporal composite of
the range of a spectral index. Burned and unburned pixels were
differentiated using a threshold based on the mean and standard
deviation of the range of this index for pixels where active fires were
detected. Similarly, in the Fraser et al. (2000) HANDS algorithm, which
was designed for mapping boreal forest burns with AVHRR data, the
expected change in successive 10-day NDVI composites for burned
pixels was derived using an AVHRR active fire mask. A similar method
was developed by Pu et al. (2004) for mapping burned areas in
California, again with AVHRR data. George et al. (2006) used two
different vegetation indexes derived from 16-day MODIS nadir BRDFadjusted
reflectance composites to detect burn scars in central Russia
over a twelve year period. Disturbed landscape segments were
identified using a contextual algorithm and NDVI differencing, and
those segments containing active fires were classified as having
burned. Lastly, Loboda et al. (2007) developed a method for mapping
burned areas on an annual basis using 500-m MODIS 8-day surface
reflectance composites and 1-km MODIS active fire masks. As with the
Roy et al. (1999) approach, thresholds for burned pixels were derived
from statistics computed for pixels where active fires were detected.
While the spatial and temporal information available from active
fire data is intuitively useful for burned area mapping, active fire maps
generally have several characteristics which complicate their use in
hybrid algorithms, particularly those intended for use in multiple
biomes. First, the minimum detectable size of an active fire is up to
~1000 times smaller than the minimum detectable size of a burned
area (Giglio et al., 2006); selecting burned training pixels based on the
occurrence of an active fire is therefore susceptible to contamination
from pixels containing small, undetectable burned areas. Second,
active-fire false alarms (i.e., commission errors) will also contaminate
burned training samples. Third, whereas using active fire locations to
guide the selection of burned training pixels is comparatively
straightforward (e.g., Roy et al., 1999), considerably more care is
required in selecting unburned training pixels, as the absence of
detected fires at a particular location does not guarantee that the
location did not burn. Active-fire omission errors can occur because
fires are too small to detect, or are obscured by clouds or overstory
vegetation, or were not actively burning at the time of the satellite
overpass. This can lead to the inclusion of burned pixels in an
unburned training sample selected on the basis of their (lack of)
proximity to hot spots.
In this paper we present an automated hybrid algorithm for
mapping burned areas using MODIS imagery which largely overcomes
the above issues. The algorithm, which is a simplified version of an
earlier prototype used by Giglio et al. (2006, Appendix A), detects
persistent changes in a daily vegetation-index time series derived
from MODIS surface reflectance observations. Maps of active fires are
used to generate regional probability density functions suitable for
classifying these persistent changes as either burned or unburned. The
algorithm identifies the date of burn, to the nearest day, for pixels
within individual MODIS Level 3 tiles (Section 2) at 500-m spatial
resolution.
Although our algorithm is conceptually similar to the hybrid
approaches mentioned above, it contains several innovations
intended to more fully exploit the information provided by active
fire maps, enabling the algorithm to function more robustly over a
wide range of biomes. Among these are the ability to identify training
samples of both burned and unburned pixels, in part through a region
growing phase, which also permits the algorithm to function in the
presence of extremely large burned areas, and the use of both spectral
and textural information to help discriminate between burned and
unburned pixels. In addition, the algorithm operates within a Bayesian
framework in which prior probabilities are adjusted based on the
proximity of burned training pixels, further exploiting the information
derived from active

Results (Indonesian) 1:

sembilan algoritma daerah yang berbeda, diterapkan 1-km VGT setiap haripermukaan reflektansi citra. Roy et al. (2002, 2005b) dikembangkanprediktif bi-directional reflektansi model pendekatan untuk petadibakar daerah setiap hari menggunakan citra MODIS 500-m. SebagianBaru-baru ini, Tansey et al. (2008) diubah salah satu daerah GBA-2000algoritma untuk global digunakan untuk menghasilkan L3JRC 1-km global dibakardaerah produk untuk 2000-2007.Meskipun sebagian besar metode yang ada pemetaan daerah dibakartidak mengeksploitasi informasi aktif api, minoritas hibrida algoritmayang menambah jarak jauh merasakan indikator "standar" digunakan untukmembakar pemetaan (reflektansi permukaan, suhu permukaan, NDVI, dll.)dengan aktif api maps. Roy et al. (1999), misalnya, menggunakan AVHRR datauntuk memetakan Sabana membakar di Afrika Selatan dari gabungan fosil darikisaran spektral indeks. Yang dibakar dan pasanglah pikseldibedakan menggunakan ambang berdasarkan mean dan standarpenyimpangan jangkauan indeks ini untuk piksel yang mana kebakaran aktif ituterdeteksi. Demikian pula, dalam algoritma tangan Fraser et al. (2000), yangdirancang untuk pemetaan hutan boreal luka bakar dengan AVHRR data,diharapkan perubahan dalam 10 hari berturut-turut NDVI komposit untuk dibakarpiksel berasal menggunakan masker aktif api AVHRR. Metode yang serupadikembangkan oleh Pu et al. (2004) untuk pemetaan daerah dibakar diCalifornia, lagi dengan AVHRR data. George et al. (2006) digunakan duaIndeks vegetasi yang berbeda yang berasal dari 16 hari MODIS nadir BRDFadjustedreflektansi komposit untuk mendeteksi bekas luka bakar di Rusia tengahselama periode dua belas tahun. Terganggu lanskap segmen yangdiidentifikasi menggunakan algoritma kontekstual dan NDVI pembedaan, danmereka segmen yang mengandung kebakaran aktif diklasifikasikan sebagai memilikidibakar. Akhirnya, Loboda et al. (2007) mengembangkan metode untuk pemetaantempat-tempat yang dibakar secara tahunan menggunakan 500-m MODIS 8-hari permukaanreflektansi komposit dan MODIS 1-km aktif api masker. Seperti denganPendekatan Roy et al. (1999), ambang batas untuk dibakar piksel diperolehidari statistik dihitung untuk piksel yang mana kebakaran aktif terdeteksi.Sementara informasi ruang dan waktu yang tersedia dari aktifdata api intuitif berguna untuk daerah terbakar pemetaan, maps api aktifumumnya memiliki beberapa karakteristik yang menyulitkan penggunaannya dalamalgoritma hibrida, terutama mereka yang dimaksudkan untuk digunakan dalam beberapabioma. Pertama, ukuran minimum terdeteksi api aktif adalah hingga~ 1000 kali lebih kecil daripada terdeteksi ukuran minimum dibakardaerah (Giglio et al., 2006); memilih dibakar pelatihan piksel berdasarkanterjadinya kebakaran aktif rentan terhadap kontaminasidari piksel yang mengandung kecil, tidak terdeteksi dibakar daerah. Kedua,Alarm palsu aktif-api (yaitu, Komisi kesalahan) juga akan mencemaridibakar pelatihan sampel. Ketiga, sedangkan menggunakan lokasi aktif apiguide the selection of burned training pixels is comparativelystraightforward (e.g., Roy et al., 1999), considerably more care isrequired in selecting unburned training pixels, as the absence ofdetected fires at a particular location does not guarantee that thelocation did not burn. Active-fire omission errors can occur becausefires are too small to detect, or are obscured by clouds or overstoryvegetation, or were not actively burning at the time of the satelliteoverpass. This can lead to the inclusion of burned pixels in anunburned training sample selected on the basis of their (lack of)proximity to hot spots.In this paper we present an automated hybrid algorithm formapping burned areas using MODIS imagery which largely overcomesthe above issues. The algorithm, which is a simplified version of anearlier prototype used by Giglio et al. (2006, Appendix A), detectspersistent changes in a daily vegetation-index time series derivedfrom MODIS surface reflectance observations. Maps of active fires areused to generate regional probability density functions suitable forclassifying these persistent changes as either burned or unburned. Thealgorithm identifies the date of burn, to the nearest day, for pixelswithin individual MODIS Level 3 tiles (Section 2) at 500-m spatialresolution.Although our algorithm is conceptually similar to the hybridapproaches mentioned above, it contains several innovationsdimaksudkan untuk sepenuhnya memanfaatkan informasi yang diberikan oleh aktifapi maps, memungkinkan algoritma untuk fungsi lebih bersemangat lebih dariberbagai macam bioma. Di antaranya adalah kemampuan untuk mengidentifikasi pelatihansampel dibakar dan pasanglah piksel, di bagian melalui wilayahfase tumbuh, yang juga memungkinkan algoritma untuk fungsikehadiran daerah dibakar sangat besar, dan penggunaan keduanya spektraldan tekstur informasi untuk membantu membedakan antara dibakar danpasanglah piksel. Selain itu, algoritma beroperasi dalam Bayesiankerangka di mana sebelumnya probabilitas disesuaikan berdasarkandekat dibakar pelatihan piksel, lebih lanjut mengeksploitasi informasiberasal dari aktif