Comparison of Three AVHRR-Based Fire Detection Algorithms for Interior Alaska

Three satellite fire detection models (threshold, contexburned over 2.7 million ha within Alaska (Table 1). Most of the area burned in Alaska is from wildfires caused by tual, and fuel mask) were compared and evaluated using lightning in remote regions. Because of the remoteness National Oceanographic and Atmospheric Administraand large area of Alaska, a timely and cost-effective tion (NOAA)-11, NOAA-12, and NOAA-14 Advanced method of detecting wildfires is needed. Daily satellite Very High Resolution Radiometer sensor data from intedata may provide timely and accurate detection of burnrior Alaska. The fixed threshold model compared the raing fires throughout Alaska. The Advanced Very High diant temperature of each pixel to predetermined threshResolution Radiometer (AVHRR) sensor on the National old values. The contextual model compared the radiant Oceanographic and Atmospheric Administration (NOAA) temperature of each pixel to its surrounding (backsatellites is useful for fire detection studies for two reaground) pixels. The fuel mask model is similar to the consons. First, due to a wide scanning range and the polar textual model, but pixels were tested for fuel availability convergence of orbital paths, AVHRR provides at least according to pre-fire vegetation index values. Fire locaeight images of Alaska every day and thus satisfies the tion data from the Alaska Fire Service was used to assess temporal frequency that is needed for effective fire dethe accuracy of the fire detection models. Fire detection tection. Second, the spectral range of AVHRR encomaccuracy: (a) was highest using the fuel mask model; (b) passes several wavelengths important for fire detection was lowest using the fixed threshold model; (c) increased (Li et al., 1997). as fire size increased; (d) was considerably greater in afternoon images than morning or night images. Fire detection methods may be less accurate in taiga/tundra OBJECTIVES regions such as interior Alaska due to landscape heteroMany of the published AVHRR wildfire detection studgeneity and relatively low aboveground fuel. Elsevier ies have been conducted at lower latitudes (Matson and Science Inc., 2000 Holben, 1987; Kaufman et al., 1990; Langaas, 1993; Kennedy et al., 1994; Eva and Flasse, 1996; Pereira and Setzer, 1996; Pozo et al., 1997). There are substantial INTRODUCTION differences of climate and landscape between lower latitude ecosystems and the taiga/tundra landscape of inteDuring the 1990s there have been many large wildfires rior Alaska. For example, during much of the Alaskan in interior Alaska due to a relatively dry climate and high fire season daylight is nearly continuous, and therefore fuel loadings. For example, the Miller Reach fire of 1996 use of nighttime AVHRR imagery to avoid solar reflecwas the most costly fire ever recorded in Alaska in terms tion (Langaas, 1993; Lee and Tag, 1990) may not be feaof economic loss. Since 1990, over 4,000 fires have sible. The landcover of interior Alaska is a mosaic of broadleaf forest, coniferous forest, wetlands, tundra, and glaciers controlled by topography and fluvial/glacial/fire Department of Forest Sciences, School of Agriculture and Land Resource Management, University of Alaska Fairbanks disturbance history (Fig. 1). Because of the heterogeneity of the landscape, some fire detection methods develAddress correspondence to D. L. Verbyla, University of Alaska, School of Agriculture & Land Resource Management, Department of oped from other ecosystems may not be applicable in inForest Sciences, Fairbanks, AK 99975-7200. E-mail: dverbyla@merlin. terior Alaska. salrm.alaska.edu Received 14 December 1998; revised 31 March 1999. The objectives of this study were: