Characterization of Clouds, Fires and Smoke Plumes in Hyperspectral Images

Hyperspectral imaging (HSI) sensors have been used for more than a decade to aid in the detection and identification of diverse surface targets, topographical and geological features. Techniques for scene characterization can utilize individual or combinations of spectral bands to identify specific features in an image. This paper deals primarily with the problem of characterization of a partially smokeor cloud-filled atmosphere. Proper analysis of the scene allows further sensing of underlying surface features such as actively burning and burn scarred regions. Both a physics-based and a semi-automated feature extraction (principal components analysis, PCA) approach are used for identifying and characterizing features in a set of AVIRIS scenes dominated by areas of smoke plumes, clouds and burning grassland as well as burnt vegetation. A combination of the two approaches is used to both discriminate (PCA) and classify (physics based) various features in a smoke/cloud filled scene. An AVIRIS scene chosen for initial testing of the two algorithms was collected on 20 August 1992 in the foothills east of Linden, CA[1]. A typical AVIRIS scene covers a 10km x 10km area at 20m pixel resolution and 224 contiguous spectral bands of data over the range 400 to 2500 nm. The scene consists of a grass fire producing a thick plume of smoke extending toward the east (see Fig. 1a). A cloud produced by the thermal properties of the fire overlies the smoke plume. Northwest of the main fire, two smoldering fires produce a thin veil of smoke that covers much of the upper half of the scene. The southwest portion of the scene is cloud and smoke free. This scene (identified as the Linden scene in this paper) provides a variety of atmospheric and surface features from which to orient and characterize. A plot of the apparent reflectance (ratio of reflected to incoming solar radiance) of various identified features in the scene is shown in Fig. 1b. The legend displays the user identified features. The cloud is significantly brighter than the smoke over the entire spectral region. The hot area is brightest in the spectral region 2000 to 2500 nm, while the fire pixels are bright for wavelengths greater then 1150 nm.