You are on page 1of 30

MOD14 algorithm

A contextual fire detection algorithm


(Giglio et al., 2003)
MODIS fire product version 4 (MOD14)

Improved version of the version 3


(Kaufman et al., 1998)

Increased the sensitivity to small cool fires.


MOD14 algorithm
The algorithms consists of

1) Masking out false-alarm causes by cloud,


desert, sun glint, and water body,

2) Selection of the fire candidates using the


threshold values of the thermal radiation.

3) Fire detection using the contrasts in the thermal


radiation between the fire and neighboring pixels.

The threshold values in 1) and 2)


The performance functions and parameters in 3)
: constants
MOD14 algorithm

1 km
MOD14 algorithm
Cloud/water

Heavy fire

Fire by contextual algorithm - compare with background T


ΔT: difference between T4 and T11

(5) or (6)
MOD14 algorithm
Commission errors of the MODIS active fire detection:
small
3% in Brazil (Schroeder et al., 2008)
3% in Siberia (Csiszar et al., 2006)

Omission errors were more common,


especially for small fires:

MODIS had a 50% detection rate:


a 0.04 km2 cluster in Brazil (Morisette et al., 2005b),
a 0.03 km2 cluster in southern Africa (Morisette et al.,
2005a)
a 0.05 km2 cluster in Siberia (Csiszar et al., 2006).
MOD14 algorithm
The merits:
1) Relative local thresholds based on
neighboring thermal averages and
variations can detect small or cool fires
that are undetectable using the simple
absolute threshold of the scene.

2) Many experiences and previous studies


have validated and upgraded the algorithm.

3) The algorithm is globally applicable


in the routine with constant parameters.
MOD14 algorithm
The demerits
1) Any surface feature that produces a sharp
radiometric transition or edge
(between desert and vegetation, cloud edges)
can potentially cause errors (Giglio et al., 2003).
2) Fires are likely missed in heterogeneous land cover,
which complicates the selection of an appropriate
background temperature (Lasaponara et al., 2003;
Schroeder et al., 2005; Wang et al., 2007).
3) Manual tuning of a series of thresholds according
to the available spectral bands is necessary for a
specific sensor (Li et al., 2005).
4) The algorithms are not optimized for regional
applications (Li et al., 2003), especially regional
characteristics of fires and backgrounds being
unaccounted.
A stochastic model of the
fire characteristics that is specific to the fire cases
is necessary for For detail discussions on the
detectability of each fire and
the improvement of the algorithm based on the
characteristics of each fire.

Liew et al. (2005) modeled each of


the fire fractional area and thermal radiation
of the active fires in Indonesian
peatland, and derived the optimal
threshold value of the thermal bands for
the fire detection using the stochastic fire model.
Summary of Hotspot Detection Algorithm
Dr.W.Takeuchi Dr.K.Nakau Dr.S.C.Liew Dr.K.Kushida Mr.W.Park Mr.K.Honda Dr.K.Kimura
Univ. of Tokyo JAXA CRISP Hokkaido Univ. Seoul National Univ. Hokkaido Univ. Hokkaido Univ.
Method of MOD 14 ver.5 Statistical Model Stochastic Model Stochastic Model FPI (Fire Potential Likelyhood Model Image Processing
Algorithm (latest version) Index) map used Technology
Spatial Multi-Criteria
Detection Analysis and
Spatial Statistics

Condition Global Siberia and Alaska • Tested for Mongolia and Siberia Tested for South Alaska Alaska
(Target Area etc.) Kalimantan and Korea (can be applied
Sumatra to Siberia or North
• Need framing Korea)
Dataset of
MODIS fire pixels

Validation • Cross validation ① 30% more • ① better than ① similar level with ① similar level with ① similar level ① similar level with
① VS.MOD14 with NOAA AVHRR sensitive than MOD14 MOD14 MOD14 with MOD14 MOD14
② VS.Ground • ③Terra ASTER MOD14 • ① and ③ are ② done ② With ② with AFS (Alaska
Truth ② Observation by done for ③ not considered AFS(Alaska Fire SErvice) data
③ VS.Satellite plane (JAL) Kalimantan and Fire service) ③ not considered
Image ③ partially done Sumatra fires data
③ not considered

Operational Operational and Research Research Research Research Research


Research To be operational
around early 2009

Web Access OK OK NO NO NO NO NO
An active wildland fire detection algorithm
based on a priori information
on fire and backgrounds
Objectives

1. Evaluate the distributions of the fire


characteristics such as the fire fractional
area and thermal radiation
of the active fire and background pixels
in boreal forests in Siberia and Mongolia
using ASTER images.

2. Build and evaluate a MODIS active fire


detection procedure using multi-band
optimal thresholds based on a stochastic
model of the fire and background
characteristics.
Table 6. Minimum cost function obtained with bi-band threshold
method using both 2.165 µm and 11.3 µm radiances.
Table 7. Omission errors with the threshold at 0.3%
and 0.0001% commission errors for bi-band
threshold method using both 2.165 µm and 11.3 µm radiances.
Table 8. Errors of fire detection on the MODIS simulation
images by bi-band threshold method when the threshold
was set at 0.3% and 0.0001% commission errors.
MOD14:
one pixel false alarm
at a cloud shadow
edge
Conclusion
1. We built and evaluated a fire detection algorithm
based on the stochastic model of fire and
scene-background characteristics in Siberian
and Mongolian boreal forests.

2. The detectability was close to equivalent to that


of the MOD14 algorithm in the objective areas.

3. In order to apply our algorithm to other regions,


fire stochastic model and background information
need to be upgraded for the region.

4. Since our algorithm and the contextual algorithms


were based on the different procedures for the
fire detection, the intercomparison and combination
with them have possibilities for further
precise fire detection.
(1) Optimal conditions for the algorithm application
(Taget area, vegitation, etc.)
Siberia, Mongolia, and other regions.

(2) Validation results of the algorithm


(MOD14, other satellite data, ground observation)
Close to equivalent (Fire stochastic model and
background information are necessary).

(3) Roadmap and necessary items


for real utilization in space agencies
(3-1) Fire stochastic model and background information
upgraded for the region.
(3-2) Transplant the algorithm to the system
of the space agency.

You might also like