| A thermal infrared remote sensing project was implemented to develop methods for identifying, classifying, and mapping thermal features. This study is directed at geothermal features, with the expectation that new protocols developed here will apply to the wildland fire thermal environment. Airborne multi-spectral digital imagery was acquired over the geothermally active Norris Basin region of Yellowstone National Park, USA. Two image acquisitions were flown, with one near solar noon and the other at night. The five-band image data included thermal infrared (TIR), near-infrared (NIR), and three visible bandpasses. While focused on TIR, the study relied on the multi-spectral visible and NIR data as well as on an ancillary hyperspectral data set. The raw, five-band data were uncalibrated, requiring implementation of two calibration protocols. First, a vicarious calibration procedure was developed to compute reflectance for the visible and NIR bands using an independently calibrated hyperspectral dataset. Second, a dynamic, in-scene calibration procedure was developed for the thermal sensor that exploited natural, pseudo-invariant thermal reference targets instrumented with kinetic temperature recorders. A suite of thermal attributes was derived, including daytime and nighttime radiant temperatures, a temperature difference (DeltaT), albedo, one minus albedo, and apparent thermal inertia (ATI). The albedo terms were computed using a published weighed-average albedo algorithm based on ratios of the narrowband red and NIR reflectances to total solar irradiance for the respective red and NIR bandpasses. In the absence of verifiable "truth," a step-wise chain of unsupervised classification and multivariate analysis exercises was performed, drawing heavily on "fuzzy truth." A final classification synthesizes a "thermal phenomenology" comprised of four components: spectral, statistical, geographical/contextual, and feature space. In situ measurements paired with image data provided effective calibrations, although results were strongly influenced by detrimental effects of requisite pre-processing routines. The classification gradient utilizing cluster-busting provided more discriminating information than a 'hard' classification. The fuzzy classifications were consistently based on a smaller number of radiometric variables than expected. The four-component phenomenology model developed here to identify geothermal features can be applied to other thermodynamic systems, particularly those related to wildland fire such as fire behavior, energy deposition, and thermal effects. |
