Biomass burning occurs throughout the world and has a significant impact on the EarthÛªs climate system . Therefore, frequent and accurate fire-scar mapping is important for fire prevention and management. While satellite remote sensing provides a critical means for monitoring biomass burning at both regional and global scales, there is a need to develop effective and automated techniques to map burned areas.
Figure 1: Example of the Terra MODIS satellite images over the burned area: (a) the dark region corresponds to the burned area; (b) the burned area is mostly coveredåÊ by clouds at the next time moment, which complicates automated fire mapping. Image Credit: MODIS data courtesy of NASA.
Some tools already exist for this purpose; most of them aim at detecting and classifying persistent changes in a daily vegetation index time series [2, 3]. However, few take advantage of spatial and temporal coherence simultaneously. Indeed, as seen in Figure 1, important noise from various sources is present in satellite observations. For instance, clouds may cover areas of interest, and lines of missing data may appear due to the incomplete coverage of the terrestrial surface. This noise, as well as low contrast in satellite images, prevents automatic tools from processing each image independently. Fortunately, most of the noise varies rapidly between successive frames. We can thus exploit temporal coherence to analyze such time series.
In this study, we analyzed 40 days of Terra Moderate Resolution Imaging Spectroradiometer (MODIS) atmospherically-corrected Level 2G daily surface reflectance measurements over the tropical savannas in the Northern Australia (ÛÏMOD09GAÛ product, tile h31v10), acquired in September-October 2011. Figure 2 shows a map of the study area. Wildfires in this region of Australia are frequent and extensive. Because of the vast areas affected and low population density, remote-sensed mapping is a critical tool for dealing with fires effectively. We used MODIS band 5 (1.240 ë_m) 500-m land surface reflectance data, as these measurements provide the highest burned-unburned separability and are largely insensitive to smoke aerosols .
Figure 2: Map of the study area over the tropical savannas in the Northern Australia. The map was generated using the Earth Observation and Modeling website of the University of Oklahoma.
Our objective is to map growing burned areas in a time series of MODIS satellite observations. We developed an approach which processes all images simultaneously and enforces the constraint that the burned areas can only grow in time. This allows the exploitation of temporal information not available for a single image.
Our method is based on graph cuts, an optimization tool that optimally segments an image into object of interest (burned area, in our case) and background. This tool comes from Graph Theory; it rewrites a segmentation problem as a search for the minimal cut in a weighted graph . As can be seen in Figure 3, each image can be mapped to the connected graph with two additional nodes: source and sink, which correspond to background and foreground, respectively. The minimal cut problem consists in finding the set of edges with the minimal sum of weights, which separates the source from the sink, yielding the partition of the image into background and foreground. The weights of edges between image pixels and sink/source express the probability of belonging to foreground/background areas, and are computed from histograms of burned/unburned reliable regions. The weights between neighboring pixels of the image grid express spatial coherence and are based on intensity difference.
Figure 3: Enforcing shape growth in an image sequence.
It turns out that the fire growth constraint can be easily enforced in graph cuts . Fire growth in a sequence of images can be expressed as the property that if a pixel belongs to the burned area at a certain time t, then it also belongs to the burned area at the next time moment t+1. This property can be introduced in the graph by connecting successive images in time series with directed infinite links, as shown in Figure 3. Such links precisely prevent each pixel from switching from foreground to background in time. Thus, we transform the problem of segmenting each image independently into a joint segmentation of all images together, subject to the fire growth constraint. By minimizing a criterion expressed on the resulting spatio-temporal graph, the proposed method yields a globally optimal segmentation.
Results and discussion
We applied the proposed method to the set of MODIS images described above. Burned and unburned reliable regions used for computation of foreground and background histograms, respectively, can be selected from the active fire observations or derived products . Figure 4 depicts the contours of detected burned areas for four time moments. These experimental results show that, in spite of the amount of noise and big portions of missing data, the proposed approach succeeded in segmenting growing burned areas, which would not be possible with the segmentation of independent frames. The computational time for the processing of the considered 40-image data set is 9 s on a 2.7 GHz Intel Core i7 processor with 16 GB 1600 MHz DDR3, and in practice it grows linearly with the number of images.
Figure 4: Contours of detected burned areas over the MODIS band 5 images over Northern Australia, acquired in September-October 2011.
In conclusion, the proposed graph-cut-based method with object growth enforcement deals effectively with significant noise, missing data and low contrasts, yielding good performance for the segmentation of growing burned areas. As both spatial and temporal resolutions of modern satellite sensors increase, this raises the importance of applying spatio-temporal methods for analyzing data acquired by new sensors. In the future, we are interested in extending our work for the segmentation of long time series of satellite data, covering the dry season with frequent fires and the wet season with biomass recovery. The ultimate goal would be to develop a complete tool for global fire monitoring, which would help us to better understand the evolution of the EarthÛªs climate and biosphere.
Yuliya Tarabalka and Guillaume Charpiat, INRIA Sophia-Antipolis MÌ©diterranÌ©e, AYIN and STARS teams, 2004 route des Lucioles, 06902 Sophia Antipolis, France. E-mail: email@example.com, firstname.lastname@example.org.
References L. B. Lentile, Z. A. Holden, A. M. S. Smith, M. J. Falkowski, A. T. Hudak, P. Morgan, S. A. Lewis, P. E. Gessler, N. C. Benson, ÛÏRemote sensing techniques to assess active fire characteristics and post-fire effects,Û International Journal of Wildland Fire, no. 15, pp. 319-345, 2006.  J. B. Lewis, ÛÏFire mapping for managers in North Australian savanna; an object orientated classification model for MODIS imagery,Û in Proc. of ARSPCÛª12, Fremantle, Western Australia, 2004, pp. 1ÛÒ12.  L. Giglio, T. Loboda, D. P. Roy, B. Quayle, and C. O. Justice, ÛÏAn active-fire based burned area mapping algorithm for the MODIS sensor,Û Remote Sensing of Environment, no. 113, pp. 408ÛÒ420, 2009.  D. P. Roy, P. E. Lewis, and C. O. Justice, ÛÏBurned area mapping using multi-temporal moderate spatial resolution data – a bi-directional reflectance model-based expectation approach,Û Remote Sensing of Environment, no. 83, pp. 263ÛÒ286, 2002.  Y. Boykov and V. Kolmogorov, ÛÏAn experimental comparison of min-cut/max-flow algorithms for energy minimization in vision,Û TPAMI, no. 26(9), pp. 1124ÛÒ1137, 2004.  Y. Tarabalka and G. Charpiat, ÛÏA graph-cut-based method for spatio-temporal segmentation of fire from satellite observations,Û in Proc. of IGARSS, Melbourne, Australia