2.3. Technical Description
2.3.1. Atmospheric Modeling
The contribution to image content from the atmosphere is probably the most important aspect of the imaging chain for remote sensing scientists to understand and characterize. One of the main focuses in the algorithm community is the development of algorithms to correctly characterize and remove the contributions of the atmosphere so that subsequent algorithms can operate on surface leaving radiances. The atmosphere introduces both multiplicative terms from absorption features and additive terms in the form of scattered and emitted energy. Additionally, the atmosphere also serves as a source of illumination. In the thermal region, reflected atmospheric emission and scattered radiance terms contribute to the target signature. The ability to perform high precision temperature emissivity separation relies on the ability of atmospheric removal algorithms to remove spectral features from the reflected target signature.
The most difficult challenge in characterizing the atmosphere is accounting for the variability in the atmosphere's contribution. Spectral transmission, emission and scattering by the atmosphere are dependent upon the atmospheric conditions (air temperature, wind speed, etc.), content (gas amounts, aerosol types and concentrations) and the path geometry. For high altitude and space-based imagers that acquire a large land area, it is not uncommon to find localized atmospheric effects that make the atmospheric removal process even more difficult. For scene modelers to provide realistic image data to algorithm developers, the simulated imagery should contain atmospheric contributions that vary as a function of scene (geographic location, time of day, atmosphere type), geometry and path length. For our simulation efforts at RIT, we find that the Air Force Geophysics Laboratory's (AFGL) atmospheric modeling codes are capable of predicting all of these required elements. MODTRAN's current spectral resolution of 2 wavenumbers can meet the spectral resolution requirements for some thermal hyper-spectral scenarios. For higher spectral resolution simulations, the user must exchange the speed of the MODTRAN band model for the slower line model algorithm utilized by FASCODE.
Rather than interactively call the atmospheric modeling code during the image generation phase, many SIG codes precompute a set of look-up tables and interpolate when necessary. DIRSIG utilizes a MODTRAN based look-up table generator which produces a series of entries for the critical atmospheric paths of interest. The angular resolution of most path types can be controlled by the user. The spectral resolution is controlled by the user on a channel by channel basis.
The DIRSIG atmospheric look-up tables contain a full hemisphere sky map so that the correct sky radiance can be selected when either directly viewing the sky or when incorporating the downwelled sky radiance into the full BRDF determination of the radiance reflected by a scene element. This rigorous treatment of the sky hemisphere allows DIRSIG to more accurately reproduce the appearance of atmospheric absorption features in reflected surface radiances. This level of model complexity is important for users in the hyperspectral algorithm development community which may develop or improve algorithms that must attempt to correctly separate spectral features of the surface emissivity from the masking atmosphere.
2.3.2. Temperature Prediction
In the thermal region of the spectrum, the temperature of targets and their backgrounds must be well predicted since the blackbody emission of these surfaces dominates the target's signature. Surface temperatures are driven by material specific thermodynamic parameters (specific heat, thermal conductivity, etc.) and by environmental parameters including meteorology (air temperature, wind speed, etc.), absorption of solar radiation, radiational exchange with backgrounds, active heat sources, etc. The spatial variation observed within a material can be the result of variations in thermodynamic parameters (due to impurities), thickness, orientation, solar insolation (shadowing), and wind loading.
For temperature prediction, DIRSIG utilizes a passive thermodynamic model called THERM [DCS 1990] that was developed by DCS Corporation for the Air Force in the late 80's. The model utilizes a meteorological history that includes air temperature, wind speed, direct (solar) insolation, diffuse (sky) insolation, humidity, etc. The meteorological history is used to drive a passive slab model that incorporates conventional thermodynamic parameters such as thermal conductivity, density, heat capacity, etc. to produce surface temperatures. THERM also has an option to include an internal heat source. The temperatures are calculated separately for each surface facet as a function of time based on first principles models. Each facet is assumed to be thermally independent of each other's and exhibit an isothermal surface behavior.
The use of the passive slab modeling approach used by THERM reflects RIT's focus on passive environmental simulation. The THERM model is not applicable for objects that are actively heated by internal sources (for example, engines, etc.). For these applications, there are much more robust and complex nodal temperature prediction models such as PRISM, MuSES, and SIRIM that can produce very accurate results for targets containing active heat sources, however, at the expense of the additional time required to construct the specific model (see Figure 2-1). Results from these active models can imported into DIRSIG simulations for targets of interest using a temperature attributed geometry model.
Figure 2-1. An active vehicle temperature model produced with PRISM (image courtesy of Signature Research, Inc and the National Ground Intelligence Center (NGIC)).
The following sections describe some mechanisms used by DIRSIG to modify the thermal model inputs to reflect the meteorological and environmental conditions, orientation and thermodynamic properties for a given hit point. This interaction with the thermal model results in improved thermal phenomenology and the ability to produce sub-facet temperature variation.
2.3.2.1. Insolation History
Most natural backgrounds and many man-made targets have significant thermal inertia which places increasing importance on the meteorological and environmental conditions in the recent past as well as the current acquisition time. For instance, in order to accurately predict temperatures in shadow regions, a solar insolation history is desired. DIRSIG computes a 24-hour solar insolation history for each rendered pixel. Figure 2-2 illustrates how the ray tracer is utilized to determine the shadowing condition for each time in the broad-band solar insolation history used by the temperature prediction model. To account for transmissive objects in the scene, broad-band transmissions are precomputed for each material and are used to attenuate the solar insolation when necessary.
The thermal image in Figure 2-3 depicts a vehicle casting a shadow on the transition between concrete and grass. Both materials exhibit the "soft" shadows indicative of surfaces observed in real imagery. The softer edged shadow on the concrete reflects the higher thermal inertia of the slab, as opposed to the sharper shadow on the grass which has a much lower thermal inertia.
Figure 2-3. An LWIR simulation of a simple scene of a vehicle casting a shadow onto an concrete slab surrounded by grass.
This approach is also used to reproduced thermal "scaring" or shadows left by objects that have been removed from a scene prior to image acquisition (see Figure 2-4).
2.3.2.2. Radiational Loading
So far this discussion of temperature prediction has focused on positive radiational loads. However, for night time and pre-dawn tasking, the cold sky is of equal importance as a negative radiational load (or sink). The importance of accurate sky exposure modeling arises from the significant contrast difference between the sky (with apparent temperatures of approximately 220 K) and backgrounds (with apparent temperatures of approximately 300 K). Using a similar mechanism that modifies the direct solar insolation on a surface based on obstruction by scene background elements, we modify the diffuse sky insolation to reflect the fraction of the hemisphere above the surface that is sky (see Figure 2-5). The sky fraction is computed simultaneously by the DIRSIG bi-directional reflectance computation which casts a series of rays into the hemisphere above the target to identify incident sky or background contributions.
The image in Figure 2-6 illustrates the effects of sky exposure on surface temperatures. The real imagery on top was acquired with RIT's thermal line scanner over Rochester, NY. The same scene and sensor has been simulated by DIRSIG on the bottom. The warmer regions that appear between the homes are due to decreased sky exposure and increased radiational exchange with warmer surfaces. Although the actual imagery was acquired on a calm night, the same phenomenology can be observed in other acquisitions for areas sheltered from the wind.
2.3.3. Spatial-Spectral Variation and Texture
The DIRSIG model attempts to reproduce texture in image targets using an approach that reflects the actual origin of spatial variation. The appearance of texture in observed imagery results from spatial variations in reflectance (arising from inhomogeneities in the material), orientation, surface structure, shading or from a combination of such factors. To simulate texture in targets, DIRSIG utilizes a large database of reflectance curves for a given material (presumed to represent the variations from inhomogeneities) [Schott 1995] and a simple bi-directional reflectance model to introduce variances due to orientation and surface structure [Brown 1997]. A texture image is assigned to the material class which represents the spatial variation of reflectance for a specified wavelength region. During the rendering process, a mapping mechanism identifies a pixel that is then used to drive the selection of a spectral reflectance curve from the large database of spectral measurements using a statistical mechanism that relates the variation in the texture image to variations in the database (see Figure 2-7). The selected spectral reflectance curve is then utilized in all the spectral computations involving that surface for the pixel currently being rendered. This approach introduces both the spatial and spectral correlation required for rigorous algorithm development applications. This approach is described in more detail in Section 6.2.2.
In the thermal region of the spectrum, texture can also arise from spatial variations in surface temperature resulting from differences in solar insolation, orientation, thickness, thermal conductivity, solar absorption and thermal emissivity. Although the spectral emissivity and reflectance are changed to reproduce the optical surface texture, the thermal model must be made aware of these perturbations to predict temperatures that reflect these optical properties. To do this, each reflectance/emissivity curve in the database is convolved with a 5800 K and 300 K blackbody distribution to produce the respective broad-band solar absorption and thermal emissivity. In operation, after the texture mechanism selects a given curve from the database, the corresponding broad-band optical properties are passed to the thermal model. The result is a spatial variation in surface temperature across a facet that reflects the spatial variations in solar absorption and radiational exchange with backgrounds.
The reflectance of a material's surface has been shown to be a function of wavelength, illumination angle, and view angle. The bi-directional reflectance distribution function (BRDF) describes these geometry specific reflectance values for all combinations of illumination and observation angles as a function of wavelength. In the thermal region of the spectrum, many materials take on more specular characteristics (especially at low view angles), therefore correct determination of the background in the specular direction is necessary. Additionally, the contrast between the backgrounds and the sky is much higher in the thermal region than in the reflective region, placing additional importance on accurate modeling.
DIRSIG allows the user to include a unique BRDF database with the spectral data set for each material in a scene. To correctly incorporate the specular and diffuse background contributions, DIRSIG casts approximately 100 rays into the hemisphere above the hit point to include radiances from sky and background sources. The radiances identified by each of these rays is then weighted by their geometry specific reflectances.
2.3.4. Sensor Modeling
DIRSIG also features a geometric sensor model that can simulate a variety of user configurable sensor geometries (i.e. line scanners, pushbrooms, whiskbrooms, etc.). The sensor model also allows the user to spatially oversample and integrate pixels to reproduce the mixed pixel challenges of actual imagery. Although this document contains "ideal" images, a post processing environment has also been developed that applies the appropriate MTF and noise contributions based on the sensor being modeled.
2.3.5. Spectral Resolution
With the growing availability of hyperspectral imagery, we find that the finer absorption and reflectance features that were suppressed in broad-band acquisitions reveal new phenomenology. Actual thermal hyperspectral imaging platforms have spectral resolutions on the order of 20 - 40 nm or 2 - 6 wavenumbers. For accurate simulations, however, it is desirable to oversample each spectral channel and spectrally integrate to produce the per channel result. This means that the desired spectral resolution for input data is approximately 5 - 10 nm or 0.5 - 3 wavenumber. Although data at this spectral resolution can be acquired with Fourier transform spectrometers (FTS), databases with a large number of field collected measurements are limited. It is even more difficult to acquire a data set with multiple measurements per sample to characterize the spectral variability of a given sample, or measurements that attempt to characterize the bi-directional qualities of the sample. At the moment, modelers and analysts must rely on the scarce data available while encouraging the development of more robust data sets.