MODELLING THE IMPACT OF WILDFIRE ON SPECTRAL REFLECTANCE P. Lewis 1 , T. Quaife 1 , J. Gomez-Dans 1,2 , M. Disney 1 , M. Wooster 2 , D. Roy 3 , B. Pinty 4 1. NCEO/Dept. Geography, University College London, Gower St., London WC1E 6BT, UK 2. NCEO/Dept. Geography, King's College London, Strand, London WC2R 2LS, UK 3. Geographic Information Science Center of Excellence, South Dakota State University, Wecota Hall, Box 506B, Brookings, SD 57007-3510, USA 4. Institute for Environment and Sustainability (IES), EC Joint Research Centre, Via E. Fermi 1, TP 440, 21020 Ispra (VA), Italy Abstract: This paper presents a method to extract information on the impact of wildfire on vegetation canopies. A simple linear model is proposed with a ‘generic’ spectral model of the impacts of wildfire on vegetation canopies. This allows a term related to the projected proportion of a pixel affected by fire (fcc) to be estimated from measurements of pre- and post-fire spectral reflectance. The properties of fcc are investigated using a hybrid radiative transfer model to simulate the impacts of wildfire in a multi-layered canopy. Spectral sampling from MODIS is assumed (7 bands). The fcc is confirmed to relate to the fractional area of a pixel affected by fire, although in multi-layer canopies it is modulated by a term dependent on the contrast between the pre-fire reflectance of areas affected by fire and those unaffected. Index Terms — Wildfire, radiative transfer modelling, MODIS. 1. INTRODUCTION There are many methods proposed to detect evidence of wildfires from optical Earth Observation data. These generally report in a binary manner: fire or no-fire, so that derived estimates, e.g. of carbon emissions, are calculated based on total fire-affected area and supposed average fire emission characteristics. When data of moderate spatial resolution are used in detection (e.g. 500 m for MODIS), it is likely that only some proportion of the pixel area will have been affected by fire, so if all fires were detected, the sum would be an over-estimate of the tru area affected. It is most likely that the proportion of a pixel affected impacts the detection algorithms, in that there will be some ‘detection threshold’ below which evidence for a fire is unreliable, and this will in turn tend to provide an under-estimate of area. The trade-off of these factors should keep estimates of burned area within reasonable bounds, but also will account for many of the discrepancies found between different satellite products. It is preferable then, if only for the purpose of area calculation, to attempt to estimate the proportion of a pixel affected by fire. This has been attempted using regressions between sub- pixel proportions and MODIS reflectance data [1] and linear mixture modelling of post-fire scene components for Landsat ETM+ [2] and hyperspectral data [3]. An alternative approach, coming from a more fire ecology-based motivation, has been to attempt to directly map burn ‘severity’, for example relating ground measures of a composite burn metric to various forms of normalized burn indices [4]. 2. LINEAR MODEL Following [5] we consider a linear mixture model of fire impacts: + = fcc a b ( ) (1) where and + are the pre- and post-fire spectral directional reflectance, fcc is the product of the fractional area affected by fire and the degree of which affected materials are converted to the burn signal, a is the reflectance of the materials affected by fire, and b is the ‘burn signal’, the reflectance that affected areas become after the fire. A model of b is presented: b = a 0 + 2a 1 m 0 0 ( ) 0 ( ) 2 2 m 0 ( ) (2) which is a constrained quadratic in wavelength , having a maximum at m (assumed 2000 nm), and offset wavelength of 0 (assumed 400 nm) and two empirical parameters a 0 and a 1 that should be non-negative. This function is capable of describing the spectral reflectance of char, ash, and exposed (dry) soil over the solar reflective range, i.e. we assume that the impact of fire is the conversion of pre-fire material to these. This means that our main parameter of interest, fcc defined by equation 1 is in essence the fraction of fire-affected material that is converted from a to char, ash and/or exposed soil. If we have measures of , + and a then in three or more wavebands, we can estimate fcc and the two associated burn signal parameters from equation 1. This is a powerful concept because: (i) it is generic because unlike other mixture model approaches, it requires no direct estimate of the spectrum of the different materials in the pre-fire signal; (ii) rather than attempting to map some index with no physical meaning (such as a burn index), we have a model parameter with a clear physical definition. Of course, if the impact of fire were other than those described, the model would be unlikely to perform well. In this sense, we can note that this is essentially a method for the mapping of fcc from observations a short time post-fire. The model should be tolerant to some factors such as the dissipation of ash and char (leaving exposed soil) that may happen hours, days or weeks after the fire, but is unlikely to perform well a long time afterwards if e.g. there is significant green vegetation growth. IV - 1019 978-1-4244-3395-7/09/$25.00 ©2009 IEEE IGARSS 2009