Figure 1. Fire size and temperature marking a pixel
The objective of this application demonstration is to assess the operational constraints related to the processing of fire products for the scientific community. Making use of the large archive datasets available at ESRIN through thematic application is one of the activity planned for the future Data User Programme.
A timely and frequently updated fire atlas over the world's tropical forests and savannahs is needed for land use, forestry, atmospheric chemistry, global climate, fire management applications. Fire product has been identified by IGBP as an important input for global change analysis (Townshend 1992). Hao & Liu 1994 recently published their results on burned biomass at five degree square scale for the late 1970s. They requested a better temporal survey of the biomass burning distribution in the tropics using satellite imagery.
Andrea et al. 1991 estimate that the biomass burned every year over Africa is 390 million tonnes for forest and 2430 million tonnes for savannahs, and Hao & Liu give a total figure of 2500 million tonnes. This represents around half of the world's biomass burning.
According to Malingreau 1990 and Hao & Liu 1994, the only solution for a rapid and efficient survey of biomass burning is the use of permanent operational satellite.
On the African continental scale, fire events are correlated to bioclimatic and phytogeographic regions as well as to human activities and meteorological conditions (Malingreau et al. 1990). It is worth noting that the peak in fire activity occurs in the afternoon. This peak activity is related to human daily activity as well as to the fire physics: during the day the vegetation dries favouring flaming. Therefore daytime fire detection is often applied even if this is more difficult than during nighttime. Fire detection is based on sensibility at fire temperatures (400- 1000 K) of channel 3 (3.7 micrometers) of the AVHRR. During daytime the radiance seen by the satellite is a composition of terrestrial emission, solar reflectance and atmospheric attenuation. The terrestrial emission could be decomposed in two different types of signals: the fire (which is a function of its temperature and size) and the background signal (also a function of its temperature and size).
A fire detection algorithm shall identify the pixels that were marked due to the presence of one or many fires and do not count the ones marked due to different environmental effects (e.g. background temperature, background reflectance, cloud ...).
All NOAA-11 AVHRR afternoon passes from Niamey and Nairobi stations have been systematically processed from July 1992 to June 1994. A total of 4000 images of 4-minute lengths have been used for this experience. The ESA / ESRIN fire algorithm development was initiated in May 1991, as part of the SHARK research package activities. The first extensive results were presented in the CD-ROM Ionia (Arino et al. 1993, Arino & Melinotte 1993) where over one thousand images with fires (West Africa, Central Africa, East Africa, South Africa, Australia) have been used to test and tune the algorithm. In 1994, applying the detection systematically to all archived data over Africa led us to create a fire index to cancel the artifact of data unavailability due to acquisition problem. The fire detection algorithm and the fire index construction are described below.
Algorithm description
Our work refers to the
algorithm proposed by Kaufman et al. 1990 which updated the
Dozier 1981 algorithm to deal with environmental effects for
daytime detection with AVHRR data.
Channel 3 saturates around 320 K. This saturation can result from environmental reasons other than a fire. Figure 1 provides a quick insight of the pixel marking due to sub-pixel fire presence at two temperatures: 500 and 800 K. Without accounting for background and atmospheric effects, the two fires, despite their size, will saturate channel 3. A pixel could obviously contain more than one fire.
Figure 1. Fire size and temperature marking a pixel
The algorithm consists of a detection with channel 3 and then a series of thresholds on the calibrated brightness temperatures and reflectances to avoid 'false alarms'. Tests 1, 2 and 3 are an adaptation of the Kaufman algorithm to work within the radiance range found over Africa. Tests 4 and 5 are original tests developed 'ad hoc' to account for the remaining environmental effects not removed by the adapted Kaufman algorithm. Test 4 was introduced in 1991 and Test 5 in 1993 when extensive processing showed the limitations of the four previous tests.
The algorithm was found to be both coherent and competent, i.e. detected fires can be considered as highly likely, however some fires can be missed. Making the algorithm more resilient to misinterpretation results in greater fire 'loss'. Therefore a trade-off was performed and the algorithm has been frozen in its actual state. The remaining unresolved problem will be cleared by a product visual inspection: Test 6.
Visual inspection of the products (Test 6)
No
algorithm is perfect: unrealistic detection is still performed
with our algorithm while some highly probable fires are lost. The
fire product visual inspection is the sixth test within the
processing chain.
The remaining error sources have been analysed visually by inspecting all Ionia Quicklooks (Melinotte & Arino 1995) where detected fires were reported. The visual inspection of all products by an operator cancels dubious cases: in case of flagrant error (e.g. large uniform zone taken as fire, line shape fires ...), the operator rejects the overall image. For the two years of study (4000 products), the visual inspection process lead to the rejection of around 10% of the images where an environmental problem marked the pixel instead of a fire.
Algorithm limitations
In addition to the
environmental problems discussed above, other types of conflicts
or artifacts need to be taken into account when analysing the
results:
Fire Index Atlas
To avoid introducing acquisition
artifacts and reduce the above-listed limitations into the Atlas
it was decided to create a monthly Fire Index at one degree
square scale. The Fire Index should therefore guarantee stable
fire statistics from one year to the other. The number of fires
within one degree square for a given month has been divided by
the number of available images within the same degree square for
the analysis (Figure 2).
The resulting Fire Index ranges from 0 to 30. The Fire Index is
intended to smooth on a monthly basis the bias introduced by the
system. Using monthly product data allows to get confidence in
the results obtained from one year to the other.
Figure 2. Fire Index construction. The example given highlights
the construction of the Fire Index over all Africa for the month
of January 1993
Processing sequence
Selected products are used as
input. Desert areas are automatically discarded. Channels 1 to
4 of the AVHRR are calibrated and geolocation is performed. Fire
detection is performed at full resolution. The Ionia Quicklook
with fire superimposed is used for visual assessment of the data
product quality (e.g. vegetation, cloud/smoke and missing line
presence will help the operator to decide). In case the operator
is not confident with the product, it is rejected as a whole.
Processing load estimation
Given
the large amount of input data (6 to 7 daily scenes in average
over tropical Africa), a quasi-operational software was used. The
processing of a monthly Fire Index over Africa takes around 80
hours on a SUN Sparc Work Station. This processing time does not
take into account the media handling which was carried out by an
automatic Juke Box, neither the Ionia Quicklook inspection which
is evaluated to be less than 30 seconds per product.
Raw fires counting
4000 images have been processed
over two years. Figure 3
shows the full-resolution raw results for the two consecutive
years over Africa. 408 638 fires have been identified between
July 1992 and June 1993 by processing 1984 products. 453 617
fires have been identified between July 1993 and June 1994 by
processing 2190 products. The normalisation by the image numbers
used gives a very stable result from one year to the other of
respectively 206 and 207 fires per image. The overall fire
distribution is bounded by the Sahara desert in the Northen
hemisphere, by the Horn of Africa in the east and by the Kalahari
desert in the Southern Hemisphere. The tropical rain forest
delimitation ie easily identified by the lack of fire both in the
south of West Africa (Liberia, Ivory Coast and Ghana) and in
Central Africa (South Cameroun, Gabon, Congo and North Zaire).
At first glance the result stability from one year to the following year needs to be analysed controversially due to the orbit drift from one year to the other and to meteorological changes.
Figure 3. Fire counts per year: July 1992 - June 1993 and July
1993 - June 1994
Monthly atlas
Figure 4 and Figure 5
show the Fire Index Atlas for the two consecutive years from July
1992 to July 1994. The temporal and spatial pattern of fire
distribution over the whole continent is very stable from one
year to the other. The fire distribution is linked to the dry
season in the Northern hemisphere with a peak in January for the
savannahs south of the Sahara desert. This fire activity peak in
January is well correlated to a strong decrease in surface albedo
over the same regions as was observed by Arino 1990.
Precipitations increase from March to September with a
latitudinal gradient from tropical forest to desert. At the
maximum of the rain season June to September, the fire activity
decreases severely.
Figure 4. Fire Index Atlas: July 1992 - June 1993
Figure 5. Fire Index Atlas: July 1993 - June 1994
In the Southern Hemisphere from June to October the fire distribution shifts from Angola/Zaire to Zambia/Tanzania to finally reach Mozambique in September. The fire distribution follows the progression of the dry season in these countries.
Our results are consistent with those of Cahoon et al. 1992. They observed the fire distribution in 1986 and 1987 at the African continent level during nighttime by using the visible channel of the DMSP. The same pattern and fire distribution repetitivity were observed. The fire density was very low for Cahoon et al., due to the fire diurnal cycle. This reinforces the concept of daytime detection. Both results demonstrate the complementarity of different satellite data for assessing quantitative and qualitative environmental information.
Our results could be used as input to a work on biomass burning such as the one of Hao & Liu. Better than deriving biomass burning activity from tropospheric ozone concentrations (which is the result) as performed by Hao & Liu, the direct observation of fire events provides frequent and timely update of fire distribution. This fire distribution could then be converted into biomass burning parameters.
A Fire Index continental product has been processed from July 1992 to July 1994. Long-term series availability is assured by ESA through its station network (Arino 1993), and the archiving of the '1-km AVHRR Global Land Dataset' (Eidenshink & Faunden 1994).
Monitoring the Fire Index over several years will consolidate the Fire Index as occurred for the NDVI after its creation. Temporal and spatial coverages need to be increased (e.g.: processing of the South American, Asian and Australian continent). The constraints in terms of operation have been evaluated. The user access to products is being defined.
The Fire Index Atlas is now available for deep analysis (e.g. global atmospheric emissions). Burned areas need to be analysed in combination with fire detection. The relationship of the Fire Index to meteorological conditions needs to be established. Finally extensive validation campaigns should be performed by scientists. This atlas proves that quantitative and qualitative information on fire distribution provided by remote sensing can be made available to the scientific community.
New (GOME on ERS-2) and forthcoming atmospheric instruments (Sciamachy on Envisat) will provide us with the possibility to search for a correlation of fire distribution with trace gas presence. ATSR-2 on board ERS-2 has the same detection capacity as AVHRR. The use of its data should increase the detection frequency and therefore improve the monitoring of fire distribution.
Andrea M O, 1991: Biomass burning: its history, use and distribution and its impact on environmental quality and global change. Global biomass burning, 3-31, MIT press, Cambridge, J S Levine Ed.
Arino O, 1990: Surface albedo and short-wavelength radiative budget: satellite contribution. PhD thesis, 182 p, 23 Febr. 1990, INP Toulouse.
Arino O, 1993: AVHRR data acquisition processing and distribution at ESA. In: Eurocourses on Advances in the use of AVHRR data for land application, Ispra, 22-26 Nov. 1993, A. Belward & G. D'Souza Eds..
Arino O, J-M Melinotte & G Calabresi, 1993: Fire, cloud, land, water: the 'Ionia' AVHRR CD-Browser of ESRIN. EOQ No. 41, July 1993.
Cahoon Jr D R, B J Stocks, J S Levine, W R Cofer III and K O'Neill, 1992: Seasonal distribution of African savanna fires. Nature, 359, 812-815.
Dozier J, 1981: A method for satellite identification of surface temperature fields of sub-pixel resolution. Rem. Sen. of the Environment, 11, 221-229.
Eidenshink J C & J L Faunden, 1994: The 1-km resolution global dataset: current progress. IJRS, 15, 3443-3462.
Hao W M & M-H Liu, 1994: Spatial and temporal distribution of tropical biomass burning. Global biogeochemical cycles, Vol.8, No. 4, 495-503.
Kaufman Y J, A Setzer, C Justice, C J Ticker & I Fung, 1990: Remote sensing of biomass burning in the tropics. Fires in the Tropical Biota, Ecosystem Processes and Global Challenges, 371- 399, Ed. J G Goldammer.
Malingreau J-P, 1990: The contribution of remote sensing to the global monitoring of fires in tropical and subtropical ecosystems. Fires in the Tropical Biota, Ecosystem Processes and Global Challenges, 337-370, J G Goldammer Ed.
Melinotte J M & O Arino, 1993: AVHRR CD-Browser 'Ionia'. Proc. 6th Eur. AVHRR Data Users Mtg, Belgirate, 29 June-2 July 1993.
Melinotte, J M & O Arino, 1995: The Ionia '1-km' Net-Browser experience: Quicklook processing and assess statistics. EOQ, No. 50, Dec. 1995.
Townsend J R G (Ed.), 1992: IGBP report No. 20: Improved Global Data for Land Applications. IGBP Secr., Royal Swedish Academy of Sciences, Stockholm, Sweden.
Corrigendum
In J-P. Malingreau & al's article on
'TREES Project Significant Results over Central and West
Africa'(EOQ No.48, June 1995, p. 8), the first line from top
should read: "The territory covered by the mosaic is more than
2.5 million km ..." (not 2500 square kilometers).
Authorship
Due to an error of transmission, the
list of authors of the article 'Observing Storm Clouds by Space-
borne Multi-frequency Microwave Radiometers' (EOQ No. 49, Sept.
1995, pp.7-12) was incomplete. It should read: 1. Alberto Mugnai,
Istituto di Fisica dell'Atmosfera CNR, via Galileo Galilei I-0044
Frascati e-mail: mugnai@hp.ifsi.fra.cnr.it; 2 & 3. N. Pierdicca
& F. S. Marzano, Dept. of Electronic Engineering, Univ. 'La
Sapienza', I-00184 Roma, Italy; 4. Eric A. Smith, Dept. of
Meteorology, Florida State University, 306 Love Bldg Tallahassee,
FL 32306-3034, USA; 5. Greg. J. Tripoli, Dept. of Atmospheric &
Ocean Sciences, Univ. of Wisconsin, Madison, WI 53706, USA.
ESA EOQ Nr. 50.