MONITORING BIOMASS BURNING USING ATSR-2 DATA

ABSTRACT

AVHRR data have long been used for detecting biomass burning. However, the afternoon overpass time coincides with the build up of clouds in tropical regions. A second, complimentary fire sample at mid-morning may now be possible using the ATSR-2 sensor which is equipped with similar thermal wavebands. An automated, contextual fire detection algorithm, originally developed for AVHRR data, was modified and applied to ATSR-2 data acquired for lowland Bolivia during the 1995 fire season. Fixed threshold components of the algorithm were tested and the results used to make further amendments to the algorithm. Preliminary validation results are presented which suggest that ATSR-2 can be used to detect biomass burning within different tropical environments.

1. INTRODUCTION

Biomass burning has major large scale impacts on biogeochemical cycles, atmospheric chemistry, vegetation dynamics and the spatial boundaries between ecosystems (e.g. Levine 1991; Crutzen and Goldammer 1993). Currently these impacts are not well understood and there is a recognised need for regional and global fire data within the framework of global change research (e.g. IGBP, 1994). Using data from the Advanced Very High Resolution Radiometer (AVHRR) sensors, it is possible to detect active fires at suitable spatial and temporal scales and so build up maps of regional and global fire activity. The majority of fire detection studies have so far used data from the early afternoon AVHRR overpass. With the launch of the second Along Track Scanning Radiometer (ATSR-2) in April 1995, there is the potential to monitor fire activity at mid-morning. This may prove particularly advantageous in the tropics where cloud cover tends to build up during the afternoon. Discussion of nightime fire detection and other temporal sampling considerations is outside the scope of this study but are considered by Robinson (1991), Langaas (1993) and Malingreau and Eva (1995). The ATSR-2 and AVHRR sensors have comparable spatial resolutions (at nadir, 1km and 1.1km respectively) and the same thermal bandwidths making it possible to adopt AVHRR fire detection techniques for ATSR-2 that have, to some extent, already been validated (see table 1).

The objective of this investigation is to assess the potential of ATSR-2 data for monitoring active fires.

The paper describes how a contextual algorithm, originally developed for AVHRR data, has been applied to ATSR-2 data covering lowland Bolivia and then amended based on the results. Bolivia was chosen because it presents a variety of forest, savanna and agricultural landscapes with different fire types and regimes, and little is known of south American fire regimes outside of Brazil.

_______________________________________________________

_______________________________________________________

Table 1. ATSR-2 and AVHRR Channels

2. FIRE DETECTION METHODOLOGY

AVHRR channel 3 (3.55m m - 3.93m m) is sensitive to the radiant emittance from biomass burning because the maximum radiant emittance from typical vegetation fires occurs at wavelengths between 3-5m m (e.g. Langaas and Muirhead, 1988). Furthermore, it is possible to detect subpixel fires because the total radiant emittance from a fire is disproportionally greater than that from a cooler background. Kennedy et al. (1994) have calculated that a background environment at 300° K with a fire covering 400m² (0.04 per cent of an AVHRR pixel) and a temperature of 800° K would be sufficient to saturate the channel 3 sensor. Signal increases in channel 3 can also be the result of reflected solar radiation from hot, sparsely vegetated surfaces or from cloud tops (e.g. Gregoire et al. 1993). To distinguish between signal increases caused by fires and those caused by warm surfaces the difference between channel 3 and channel 4 (11.0m m) temperatures can be used. Fires will result in a much higher channel 3 temperatures than those for channel 4; whereas there will be a much smaller difference in channel temperatures over warm surfaces. To remove the effect of cloud reflectances, masking is normally performed before applying fire detection techniques. When channel 3 and 4 temperatures of a fire pixel are known for an unsaturated fire pixel, it is possible to determine the size and temperature of subpixel fires using the method of Dozier (1981). However, the technique is concerned with fire characterisation rather than detection and is not considered here.

2.1 AVHRR fire detection algorithms

The simplest methods of fire detection use visual image interpretation or empirically derived thresholds for channel 3 temperatures and/or the temperature differences between channel 3 and channel 4. There are numerous examples of such methods which have been used for detecting a range of fire types, including: deforestation fires (e.g. Malingreau et al. 1985; Flannigan and Vonder Haar 1986; Matson and Holben 1987; Matson et al 1987; Malingreau and Tucker 1988; Kaufman et al. 1990); savanna fires (e.g. Gregoire 1993; Kennedy et al. 1994) and straw burning (Muirhead and Cracknell, 1985; Cracknell and Saradjian 1996). The main disadvantage of these algorithms is that fixed thresholds are ecosystem and season specific (Justice et al. 1993; Langaas, 1995). As a result, they are not suited to mixed landscapes like those found in Bolivia.

2.2 Contextual algorithms

Contextual approaches use variable thresholds based on the immediate environment of the pixel being tested. This overcomes most of the problems of fixed limits and means that the algorithm can be applied consistently over large areas. Two separate but similar approaches are reported in the literature. Harris and Rothery (1995) developed a contextual method to consistently identify volcanic thermal anomalies over time. They compared the channel 3 and 4 temperature difference, of each pixel in the study area, with the mean difference of its immediate background. Then the target pixel and background difference was compared to the temperature difference for all pixels in the region immediately outside of the study area. If the temperature differences were greater in all cases the pixel was flagged as hot. The second method has been developed by Flasse and Ceccato (1996) within the Local Applications of Remote Sensing Techniques group (LARST) at the Natural Resources Institute (NRI) for operational fire detection and for use in the Global Vegetation Fire Product (Stuttard et al. 1995). This algorithm has been adopted for use with ATSR-2 data. It is specially designed for the detection of vegetation fires regardless of ecosystem, season or type of fire. The algorithm consists of five main steps which are repeated here with the notation given in Flasse and Ceccato.

[1] Masking of cloud, water and desert areas to remove non-fire pixels.

[2] Selection of a pixel as a potential fire pixels if,

and

where T₃ and T₄ are AVHRR channel 3 and channel 4 brightness temperatures.

[3] Masking of pixels with a very high reflectance that might contaminate the channel 3 signal. Pixels are removed if,

r ₂ ³ 20%

where r ₂ is the top-of-atmosphere bi-directional reflectance for the AVHRR near-infrared channel.

[4] Computation of background statistics for potential fire pixels using a variable spatial window starting with 3 x 3 pixels. Only pixels that have not been masked and that are not potential fires can be used in the computation. The window operates until a minimum of three or at least 25 per cent of the window pixels can be used. If insufficient pixels are available the window size is increased until a maximum of 15 x 15 pixels. The neighbouring pixels are used to calculate:

T_3b = T₃ mean for the neighbouring pixels

s _T3b = T₃ standard deviation of the neighbouring pixels

T_34b = [T₃ - T₄] mean of the neighbouring pixels

s _T34b = [T₃ - T₄] standard deviation of the neighbouring pixels

[5] Testing the potential fire against its background statistics. The potential fire is confirmed if,

and

where _pf is a potential fire.

Although the algorithm has not been thoroughly tested, the initial results seem promising with 90% of visually identifiable fires being detected and commission errors limited to 15% (Stuttard et al. 1995; Flasse and Ceccato 1996). However, there are several weaknesses in the algorithm that have been recognised by the authors. The most important of these are the use of fixed values in the high reflectance test in step 3 and the first context test in step 5a; and the arbitrary limits set for the minimum acceptable sample size and maximum allowable window size. The validity of these values is investigated as part of this study. See Justice and Dowty (1993) for a detailed review of fire detection algorithms.

3. METHODOLOGY

The algorithm described above was firstly amended with respect only to the identification of potential fires. The ATSR-2 3.7m m band saturates at 312° K so a pixel was flagged as a potential fire if,

T_1a = saturation

and

T₂< 304° K

where T_1b and T₂ here denote ATSR-2 3.7m m and 11.0m m channels respectively. The saturation criteria ensures that the minimum T_1b temperature of a potential fire is 312° K. Although this is lower than for AVHRR (saturation = 321° K), the effect of relected solar radiation in channel T_1a will be less marked because of the earlier overpass time.

For this study a pixel was classified as cloud if,

T₃ < 285° K and V₁ + V₂ > 40%

and

classified as water if,

NDVI < -12 and T_1b < 5%

where T₃ is the ATSR-2 12m m band, V₁ and V₂ are the red and green bands and T_1b is the short wave infrared channel. The cloud screen is a simple test to remove cold and very reflective pixels from the analysis. Use of less rigorous thresholds did not remove a sufficiently large proportion of pixels that were identifiable as cloud on the images. There was no requirement for a desert mask with the Bolivia data. The context tests were left unchanged. In order to test the response of fire detection rates to changes in the fixed thresholds, the algorithm was applied using all combinations of: high reflectance values from 15 to 25%; minimum acceptable sample sizes of 3, 6, 9, 12 and 15, and maximum allowable window sizes from 3 x 3 to 17 x 17.

3.1 ATSR-2 data processing

ATSR-2 data were obtained from the Rutherford Appleton Laboratory (RAL) as gridded brightness temperature (GBT) products. Each product consists of 512 x 512 geolocated, collocated nadir and forward-view images at a 1km resolution (Bailey 1994). For this study only nadir view data were used. The spectral bandwidths are given in table 1. Thermal data is supplied as calibrated brightness temperatures and the reflectance data was converted to top-of-atmosphere reflectance using calibration coefficients supplied by RAL. All the available images of lowland Bolivia in 1995 were examined and only those with a full compliment of nadir channel data were selected for further analysis. The images were visually inspected to remove those with approximately 80-100 per cent cloud cover. The remaining 18 images were used in the study and provide a reasonably even coverage of the 1995 fire season (May to December) .

4. RESULTS AND DISCUSSION

4.1 Scattered Cloud

Visual inspection of the algorithm results showed that a significant number of fire pixels were occurring in areas of scattered cloud and at the edges of large cloud banks. In both cases it is reasonable to assume that subpixel clouds are flagged as potential fires because the cloud is sufficient to cause T_1a saturation and a reduction in T₄ temperatures. These pixels are not rejected by the contextual tests because the immediate background is cloud free and T₃ temperature and T₃, T₄ differences are lower. To overcome this problem a simple three pixel buffer added around cloud pixels classified which has the effect of ‘infilling’ areas of scattered cloud and removing false detections (see figure 1). The infilling should also improve the validity of background statistics in general.

4.2 Minimum sample size and fire counts

Figure 1 shows a plot of fire counts against the minimum sample size for three different window sizes. The high reflectance test threshold has been set at 20%. There are two important trends. First, fire counts are significantly larger for a minimum sample size of three than for other sample sizes, and second, fire counts for sample sizes between 6 and 15 are relatively stable but show a gradual decline. Interpreting these results is difficult because increasing the minimum sample size criteria has a direct effect on the mean and standard deviation estimates for the background, by increasing the number of samples used in the statistics; and an indirect effect, by forcing the statistics to be calculated on different pixels with a different proximity and relationship to the potential fire pixel. For example, a minimum sample size of 9 or over precludes the use of a 3 x 3 window. However, it is clear that sample size influences the algorithm results and that there is a trade off between getting statistics from the immediate fire background (i.e. the 3 x 3 or 5 x 5 windows) and getting a good mean and standard deviation estimate (i.e. a larger sample size). The limit was changed to a minimum sample size of 6 and appears to provide a reasonable balance between the two requirements. The minimum proportion of pixels needed for any given window size was also increased to 33% so that more representative sample of the background was used in the computation.

4.3 Maximum allowable window size

Changes in the maximum allowable window size do not have a significant effect on fire count, except that larger windows result in gradually larger fire counts because they allow more potential fires to be tested. Ideally, the choice of window size should be made with reference to the scale of the landscape. For example, the landscape of Eastern Bolivia is a mosaic of savannas and forests. Depending on which window size is used potential savanna fires could be tested against background statistics calculated only from savanna pixels or from a mixture of savanna and forest pixels. For a contextual approach to work best, it is preferable that the a potential fire is tested against statistics calculated from a similar land cover type. One solution could be to incorporate a vegetation mask within the algorithm which would adjust the maximum allowable window size (or other parameters) automatically; or select only pixels of a similar vegetation type to the potential fire, for the background statistics.

4.4 High reflectance thresholds and fire counts.

Figure 2 shows a plot of fire counts against different thresholds for the high reflectance test with three separate curves for the periods May-August, September and October. It can be seen that fire counts generally decrease as the threshold is increased. This is explained by fewer highly reflective pixels being masked as the threshold rises. Mean T3 background temperature and T3, T4 differences are increased by the inclusion of these pixels in the statistics causing more context test failures. There is no clear point at which the threshold is more or less influential and the relationship between the threshold and fire counts is highly variable between time. This suggests that the threshold decision has a marked and arbitrary affect on algorithm performance and that this affect varies over time. In order to reduce the impacts of this test, the threshold was increased to 25% were there is a decreased rate of change in the fire count.

4.5 Context Test Adjustment

A final adjustment was made to the algorithm by changing the first context tests to:

T_3pf > T_3b + [3s _T34b]

This removes the 3° k fixed threshold and gives more weight to the local variance in temperature difference.

4.6 Visual Interpretation

The new algorithm was re-applied to the 18 test images. Results from the re-amended algorithm have not yet been fully tested. Preliminary work indicates that the algorithm works well for the Bolivian dataset. Figure 3 shows fires identified by the algorithm for a part scene in August. The two active fires visible from their smoke plumes have both been identified by the algorithm. The fire front of the larger fire is partially described by the pattern of fire pixels. It is also worth noting that these fires are occurring in different ecosystems and that there are no obvious incorrect detections. From all fires identified by the algorithm, 30 were randomly selected and tested visually in the imagery. All were located within areas able, and likely, to be affected by biomass burning (i.e. savanna, agricultural land and at the savanna-forest boundary). Seven had associated smoke plumes or visible reflectances much lower relative to their neighbours and consistent with charred or scarred vegetation. The rest could neither be confirmed or rejected visually which is not surprising considering that the technique is detecting subpixel events. Field data listing field locations and dates of active fires during 1995 will be used to further test the algorithm.

5. CONCLUSION

Initial results suggest that it is possible to automatically detect mid-morning fire activity using ATSR-2 data with the amended contextual algorithm. Also, the contextual approach can be used to successfully discriminate vegetation fires in different environments for the duration of an entire dry season. Further work will be aimed at validating the algorithm using field data.

6. ACKNOWLEDGEMENTS

M. C. Perrin is supported by NERC CASE studentship GT4/94/356/L with Earth Observation Science Ltd. The ATSR-2 data was supplied by RAL and NERC.

7. REFERENCES

Brunette, J. M., Vices, J. B., Fountain, J., Manissadjan, K., Podaire, A. and Lavenu, F., 1991, Remote sensing of biomass burning in west Africa with NOAA-AVHRR, in Levine, J. S. (editor), 1991, Global Biomass Burning, MIT Press, Cambridge, MA.

Cracknell, A. P and Saradjian, M. R., 1996, Monitoring of straw burning in the U. K. using AVHRR data - summer 1995, International Journal of Remote Sensing, 17, 12, 2463-2466.

Crutzen, P. J. and Goldammer, J. G, 1993, Fire in the Environment: the Ecological, Atmospheric, and Climatic Importance of Vegetation Fires, Wiley & Sons, New York.

Dozier, J., 1981, A method for satellite identification of surface temperature fields of subpixel resolution, Remote Sensing of Environment, 11, 221-229.

Flannigan, M. D. and Vonder-Haar, T. H., 1986, Forest fire monitoring using NOAA satellite AVHRR, Canadian Journal of Forest Research, 16, 975-982.

Flasse, S. P. and Ceccato., 1996, A contextual algorithm for AVHRR fire detection, International Journal of Remote Sensing, 17, 2, 419-242.

Gregoire, J. -M., 1993, Use of AVHRR for the study of vegetation fires in Africa: Fire Management Perspectives, Euro courses: Advances in the use of AVHRR data for Land Applications, Ispra, Italy.

Harris, A. and Rothery, D., 1995, Thermal monitoring of volcanoes using data from the AVHRR, In Proceedings of the 21st Annual Conference of the Remote Sensing Society, 11-14 September 1995, Southampton, 528-535.

IGBP, 1994, International Global Atmospheric Chemistry (IGAC) Project - the Operational Plan, IGBP Report No.32.

Justice, C. O. and Dowty, P (eds) 1993, IGBP-DIS satellite fire detection algorithm workshop technical report. IGBP-DIS Working Paper No. 9, 88pp., February 1993, NASA/GSFC, Greenbelt, Maryland, USA.

Justice, C. O., Malingreau, J. P., and Setzer, A. W., 1993, Satellite remote sensing of fires: potential and limitations. In Fire in the Environment: the Ecological, Atmospheric, and Climatic Importance of Vegetation Fires, edited by Crutzen, P. J. and Goldammer, J. G, Wiley & Sons, pp77-88.

Kaufman, Y., Tucker, C. J., and Fung, I., 1990, Remote sensing of biomass burning in the tropics. Journal of Geophysical Research, 95, 9927-9939.

Kenndy, P. J., Belward, A. S., and Gregoire, J -M., 1994, An improved approach to fire monitoring in West Africa using AVHRR data. International Journal of Remote Sensing, 15, 2235-2255.

Langaas, S, 1993, Diurnal cycles in savanna fires, Nature, 363, p120.

Langaas, S., and Muirhead, K., 1988, 'Monitoring' bushfires in west Africa by weather satellite. In Proceedings of 22nd International Symposium on Remote Sensing of Environment, Abidjan, Cote d'Ivoire, 20-26 October 1988, (Ann Abor, MI: ERIM), 2, 253-268.

Langaas, S., 1995, A critical review of sub-resolution fire detection techniques and principles using thermal satellite data, submitted to Remote Sensing Reviews.

Levine, J. S. (editor), 1991, Global Biomass Burning, MIT Press, Cambridge, MA.

Malingreau, J. P. and Eva, H., 1995, Notes on the temporal sampling of remotely sensed data for active fire monitoring, IGBP-DIS Workshop on Global Fire Monitoring - Joint Research Centre, Ispra, Italy.

Malingreau, J. P., Stephens, G. and Fellows, L., 1985, Remote sensing of forest fires: Kalimantan and north Borneo in 1982-83, Ambio, 14, 6, 314-321.

Malingreau, J. P., and Tucker, C. J., 1988, Large scale deforestation in the south-eastern Amazon basin of Brazil. Ambio, 17, 49-55.

Matson, M. and Holben, B., 1987, Satellite detection of tropical burning in Brazil, International Journal of Remote Sensing, 8,3, 509-516.

Matson, M., Stephens, G., and Robinson, J., 1987, Fire detection using data from the NOAA-N satellites, International Journal of Remote Sensing, 8, 7, 961-970.

Muirhead, K. and Cracknell, A. P., 1985, Straw burning over Great Britain detected by AVHRR, International Journal of Remote Sensing, 6, 5, 827-833.

Robinson, J. M., 1991, Problems in global fire evaluation: is remote sensing the solution?, chapter 8, in Levine, J. S. (editor), 1991, Global Biomass Burning, MIT Press, Cambridge, MA.

Stuttard, M., Boardman, S., Ceccato, P., Downey, I., Flasse, S., Gooding, R., and Muirhead, K., 1995, Global Vegetation Fire Product Final Report for Joint Research Centre, Contract 10444-94-09-FIEP ISP GB, Ispra, Italy.

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry