Spatial-Temporal Burn Mapping Using Remote Sensing

Burn scars are readily observable on post-burn false color imagery (Figure 4, see MODIS Images link), suggesting that burning alters the fundamental spectral reflectance properties of imagery in a way that should be amenable to mapping with remote sensing. The challenge is to develop a technique that discriminates burned from non-burned areas, against a complex background of variable land cover. An effective burn mapping technique must also be temporally stable, that is, it must remain accurate and reliable throughout the burning season. This last requirement presents a particular challenge, because of the seasonal timing of burning in the Flint Hills. In order for the grassland canopy to ignite, it must be relatively dry. In addition, biomass production during the subsequent growing season is maximized if the canopy is burned near to the time when phenological green up would normally begin. Thus, the majority of burning in the Flint Hills is done in March or April. During this time, the grassland canopy is undergoing rapid phenological development as the winter senescence period ends. To adequately meet the needs of this application, any decision rule used to map burned areas must therefore continue to work under variable and rapidly changing canopy conditions. From a remote sensing perspective, this means that automatically detecting a burn requires an algorithm that can discriminate between soil/burn residue against a variety of backgrounds ranging from senescent (brown) canopy to actively photosynthesizing (green) canopy (see Figure 3). This discrimination is further complicated by the presence of water bodies, roads, and other land cover types which might spectrally resemble burn residue.

Our proposed method for mapping burned area is based on a techniques developed through a pilot study of Flint Hills burn mapping undertaken by the KSU Remote Sensing Research Laboratory, combined with a method developed for burned-area mapping of African grasslands. Our approach exploits both the temporal and spectral dimensions of remotely sensed data to map burns in the Flint Hills study area, an approach similar to that used in previous burn detection studies. We will map burns from the 2001-2007 growing seasons using imagery from the MODIS MOD09GQ data set, acquired at no charge from the Eros Data Center (EDC) Land Process Distributed Active Archive Center (LP-DAAC). Results from the pilot study showed that bands 1 and 2 of this data set offers the optimal combination of spatial, temporal, and spectral properties for this application. The MOD09GQ data set is available on a daily basis, allowing us to detect and assess burn areas soon after they occur. Bands 1 and 2 of this data set are sensitive in the Red and Near Infrared spectral regions, respectively. The spectral response of these two bands encompasses the 'red edge' region of the spectral reflectance curve for vegetation, a region closely associated with biophysical properties of plant canopies. Bands 1 and 2 of the MOD09GQ data set are also available at a spatial resolution of 250m, the finest resolution available using the MODIS sensor. Although this spatial resolution is coarse compared to other earth resource remote sensing satellites (e.g. SPOT-HRV, Landsat TM/ETM+), the scale at which burning occurs in the Flint Hills is such that a typical burn should be represented by multiple pixels, (see Figure 4) enabling estimates of burn area to be made.


Burns will be mapped by first acquiring all acceptably clear and cloud-free MODIS imagery for the burn period in each year of the study from LP-DAAC. Cloud condition will be assessed via the image metadata as well as through visual examination of the preview tiles. Images whose QA metadata indicate an excessively large number (>20%) of low-quality pixels (due to cloud contamination or other error-source) within the study area will be rejected. As earlier noted, the burning in the Flint Hills is generally carried out over a period of four to six weeks, thus application of our algorithm could require acquisition of as many as 50 MODIS images per year. However, the likelihood of cloudy conditions during this season suggests that less than half of the days will yield clear imagery Bands 1 and 2 will be extracted from each clear image, georectified, and stacked into a 2-band image using the MODIS Reprojection Tool software, yielding a series of 2-band, TIFF formatted daily images of the Flint Hills study area. These images will then be used to map burn area by automated detection of the burn scars.

Although a number of difference- or ratio-based indices for burn area mapping using MODIS bands 1 and 2 have been suggested, our experience shows that these are not as effective as in-band spectral reflectance for discriminating burns against a variety of canopy backgrounds that may include senescent vegetation. We will therefore use in-band reflectance from MODIS Bands 1 and 2 together in an object-based classification algorithm. O bject-based classification is appropriate for this application because it can account for individual or small groups of pixels that are often misclassified using conventional, pixel-based approaches. Classification problems often arise when single pixels representing small water bodies or cloud shadows are misclassified as burns, or when burned areas are classified as unburned due to low pre-burn biomass in the burned area (lack of sufficient post-burn spectral contrast). Misclassification can also occur when sub-pixel patches of woody vegetation or rocky terrain affect the reflectance of the entire pixel, causing it to appear unchanged even though burning has occurred in the pixel. Object-based techniques consider smaller pixels to be part of a larger object (a burn), thus preventing the ‘peppered' effect within otherwise homogenous classes common to pixel-based classification techniques. Additionally, user-defined inputs prevent the segmentation from smoothing over patches within burns that may have actually gone unburned.

Object-based mapping consists of two steps, 1) image segmentation, and 2) classification. In the segmentation step, spatial objects {defined as groups of adjacent pixels treated as a single entity} will be segmented using the Fractal Net Evolution Approach (FNEA) as implemented in the eCognition software, v4.0. FNEA defines objects using a bottom-up technique, in which single pixels are merged into objects until they meet a user-specified threshold defined in both spectral and geometric terms. These user-defined thresholds allow control over the size of the objects (scale parameter & compactness/smoothness), the weights of each spectral band in the imagery, and the shape of the objects (shape factor & compactness/smoothness). Experience with mapping burns in the Flint Hills has shown that the optimal parameters for image segmentation were a scale parameter of 65, a shape factor of 0.25, and a compactness/smoothness ratio of 50/50. Because both bands are useful for detecting grassland burns, both are given equal weight in the segmentation.

The second (classification) step will utilize a supervised decision-tree algorithm. Training objects for the algorithm will be selected for each of five classes: water, clouds (if present), burned areas, crops, and grassland. These data will be used to train a decision tree classifier algorithm. Use of a decision tree is indicated for two reasons. First, the number of available training objects for some classes is likely to be small, and decision trees are robust in the presence of sparse or non-normal data. Second (and related) the decision tree method allows factors such as object shape, size, and texture to be included as classification variables, along with the object's spectral information. These object geometric properties often do not conform to a parametric distribution, making them incompatible with parametric decision rules. A new decision tree will be developed for each image, to accommodate phenological difference between dates.

Once each image is classified, all objects will be imported into the ArcMap GIS. Because only burned areas are of interest here, objects belonging to other classes will be recoded as ‘unburned'. All mapped burns near reservoirs and large rivers will also be eliminated, since water is likely to be confused with burned area. Two types of burn information will be maintained within the ArcMap spatial data base, one consisting of burns coded by the date of first detection, and the other consisting of a cumulative map of all burned areas. Note that date of first detection may not be the actual burn date, since the burning may have occurred after the time of satellite overpass or on a day when cloud conditions prevented clear image acquisition .

Clearly, the most crucial step in development of the classification algorithm is determination of a set of rules for discriminating between burned and unburned areas. Although conceptually simple, the decision-tree classifier is subject to error, the most significant of which are likely to be caused by the continual phenological development of the canopy during the green up period. To circumvent this problem, we will calibrate decision trees using field observations of burned and non-burned grasslands at the Konza Prairie Biological Station (KPBS) and Rannells Ranch experimental range, both prairie research sites administered by KSU. In-situ spectral reflectance will be measured using an Analytical Spectral Device (ASD) portable hyperspectral radiometer. The spectral response of burned and non-burned sites will be measured immediately prior to burning and for several days afterward using the ASD instrument. The spectral response of MODIS bands 1 and 2 will then be synthesized from the hyperspectral data and used to improve the accuracy of the classification decision tree.

Accuracy of the burn maps will be assessed through field survey. Obviously, it is not possible to check the accuracy of burn area calculation for the retrospective (2001-2007) data set, although results from the pilot study showed that the burn mapping technique described above achieved detection accuracies exceeding 90%. We will therefore begin our field analysis during the 2008 burning season and continue through all three years of the project. Burn locations for 2008-2010, will be determined by periodic field reconnaissance in the Flint Hills during the active fire season. As earlier noted, burns tend to be concentrated in time on prime burning days, thus our field surveys will be conducted soon after these burning days are observed to have occurred. Burn sites will be located by visual survey, and their locations determined using a portable GPS receiver. When access to the land can be obtained (i.e. where owner permission is granted), GPS will be used to delineate the perimeter of burns, from which ground-based estimates of the burned area can be obtained. The area of these sites will be compared to satellite-derived estimates as a further check on the accuracy of the mapping algorithm. A spatial data base of burn reference data will be constructed within a GIS framework to facilitate comparison with image data.

In areas where access to burns is not possible (either due to terrain restrictions or access problems), the location of burned areas will still be mapped using triangulation techniques based on sitings made from accessible locations proximal to the burn site. Proximal locations will be located via GPS. This technique will not provide detailed information on burn area, but it will allow us to assess the accuracy with which burn sites can be detected. Because retrospective accuracy assessment is not possible, the accuracy of the 2008-10 burn maps will be used as a surrogates for the accuracy of mapped data from the retrospective years of the study.

Burned Area Mapping in the Flint Hills Using Object-Based Classification of MODIS Imagery Poster