The name DIRSIG is an acronym for "Digital Imaging and Remote Sensing Image Generation". The first part of the formal name comes from the Digital Imaging and Remote Sensing (DIRS) Laboratory at the Rochester Institute of Technology (RIT) where the model was created.
The DIRSIG model is a complex synthetic image generation application which produces simulated imagery in the visible through thermal infrared regions. The model is designed to produce broad-band, multi-spectral and hyper-spectral imagery through the integration of a suite of first principles based radiation propagation sub models. These sub models are responsible for tasks ranging from the bi-directional reflectance distribution function (BRDF) predictions of a surface to the dynamic scanning geometry of a line scanning imaging instrument. In addition to sub models that have been specifically created for the DIRSIG model, several of these components (MODTRAN and FASCODE) are the modeling workhorses for the multi- and hyper-spectral community. All modeled components are combined using a spectral representation and integrated radiance images can be simultaneously produced for an arbitrary number of user defined bandpasses.
The scene simulation work being conducted at the Rochester Institute of Technology (RIT) has been focused on developing a radiometrically rigorous scene simulation capability. The DIRSIG model is capable of producing imagery that features mixed pixels, complex in-situ illumination loadings and the spectral statistics observed in real world materials. We are confident that the physics implemented by the model is extensive enough to create realistic imagery when robust input databases are utilized and the proper sensor modeling is employed.
One of the primary goals of this modeling effort is to produce imagery that can be used to test the performance of spatial and spectral image exploitation algorithms. If the modeling tools can successfully reproduce imagery with spatial and spectral "clutter" comparable to real-world imagery, then candidate algorithms can be extensively tested over a wider range of environmental conditions at a significant cost savings over field collections. In addition to the convenience of creating synthetic data over a wide variety of environmental conditions and image acquisition approaches, a synthetic data set can provide "perfect" ground truth without the technical and logistical challenges of running a successful field campaign. As we will discuss later, the synthetic image ground truth can provide the algorithm tester with very detailed insights into the performance of the algorithm that cannot be matched by field collections. It should be noted, however, that the authors do not believe that synthetic data should ever wholly replace real-world image data for algorithm testing. Instead, synthetic data should be considered a powerful tool to assist in the testing of algorithms and potentially as a surrogate when real data are not available.
First principles based approaches imply that fundamental physics, chemistry and mathematical theories are used to predict higher level phenomenologies. For example, the interaction between light and matter can be described using the work of Fresnel and others. These theories can be used to predict whether a photon with a certain wavelength will be absorbed or reflected by a material with a specific chemical composition. At a much higher level, the same interaction might be summarized as the "color" of the material.
Another example of a first principles approach would include the prediction of a surface temperature using fundamental properties including thermal conductivity, density, radiational absorption factors, radiational and convective loadings, etc. These parameters can be used with a set of fundamental governing equations that describe the flow of energy in and out of the surface to predict the steady-state temperature.
The DIRSIG model produces imagery using a predictive engine that is built around a collection of first principles based sub models.
Conventionally, spectroscopy has referred to the use of high spectral resolution absorption and reflectance features to identify material composition and/or concentration in a given sample. Although multiple measurements may be acquired to limit noise contributions, spectroscopy usually denotes the use of a single spectral sample for study. Imaging spectrometers combine the high spectral resolution aspects of spectroscopy with the availability of multiple spatial samples found in an image. The collected data from such instruments are images with spectral data associated with each pixel. In the past, image exploitation algorithms such as atmospheric characterization/removal and land-cover classification were performed on images with limited numbers of spectral bands. The first imaging spectrometers were designed and constructed to operate in the reflective region of the spectrum. The image products generated by these platforms contain signatures at spectral resolutions high enough to utilize conventional spectroscopy exploitation methods on a pixel-by-pixel basis. As a result, new suites of exploitation algorithms have emerged that specifically focus on known absorption and reflectance signatures in the atmosphere and targets of interest. Advances in thermal imaging technologies over the past decade have resulted in a new generation of platforms with hyper-spectral image acquisition capabilities that were once restricted to the reflective region of the spectrum. The thermal infrared region contains unique "fingerprint" absorption and reflectance signatures of materials that have been exploited by laboratory infrared spectroscopy for decades. The utilization of these features should allow analysts and algorithms to more accurately identify both the materials and the atmosphere being imaged through.
The DIRSIG model finds its roots in a simple image rendering program that was started by some students at RIT in the mid to late 1980's. The model would render simplistic scenes into thermal radiance images using basic material thermodynamic properties and broad-band emissivity values applied to 2D scene elements. The AIRSIM THERM temperature prediction model would predict surface temperatures using a supplied weather history. This temperature would be used to compute a blackbody radiance which would be weighted by the surface emissivity. The predicted surface leaving radiances could then be propagated through the atmosphere using path radiances and transmissions from LOWTRAN. This resulting model could be used to perform basic target and background thermal contrast studies. However, since the scenes were 2D in nature, the target and background interactions that result in phenomenology including shadows, etc. where nonexistent.
In the early 1990's, this basic modeling capability was advanced to make use of 3D scene geometry. This 3D rendering model is the first piece of software referred to by the name DIRSIG. A ray-traced approach was utilized which allowed the virtual camera to be placed anywhere within the scene. The ray tracer was also used to introduce shadows and specular reflections. Under the hood, the radiative transfer engine began to perform the radiometry calculations on a spectral basis. With this new feature, the user could supply spectral response curve data to predict radiances for specific sensors. However, the model was only applicable in the long-wave thermal infrared region of the spectrum where the solar contributions where minimal.
The model was then expanded to include photons directly transmitted and scattered by the atmosphere from the Sun. This included the initial approaches to estimating the diffuse illumination loading onto surface by sampling the hemisphere above the target using the ray tracer.
In the mid 1990's, operational hyper-spectral imaging systems started to appear in the form of the AVIRIS, HYDICE and SEBASS sensors. The adaptation of traditional spectroscopy and signal processing algorithms for use with hyper-spectral data resulted in user demand for synthetic hyper-spectral data sets that could be used for algorithm testing. The inherent spectral nature of the DIRSIG model under the hood meant that only minor improvements were required for the model to produce hyper-spectral image cubes. However, to address the need for improved spectral clutter, new techniques were incorporated that introduced spatially and spectrally correlated reflectance variations that produce the texture variations observed within materials.
Transmissive materials where also added to the simulation capabilities during this period. This allowed the model to predict the solar load on a target of interest beneath scene elements including vegetation and camouflage netting. This also allowed the tool to model the absorption by transmissive volumes including factory gas plumes and clouds.
A geometric sensor model was introduced in 1995 which allowed the model to produce imagery that contained the geometric distortions that would be produced by scanning imaging systems like line and pushbroom scanners. The optical modulation transfer function (MTF) of the sensor needed to be modeled in post-processing of the sensor reaching radiance field.
In the late 1990's, the model was being used internally by graduate students and research staff to produce imagery test sets to evaluate the performance of land cover classification, spectral unmixing, sensor fusion, and target detection algorithms under different image acquisition conditions. Until this point, only a handful of sponsoring organizations outside of RIT were using the DIRSIG model. At this time, the model was an engineering level code with very little user documentation. In 1999, RIT was approached by the U.S. Government to release the model to all government organizations and contractors.
To meet this request, the model underwent a significant software rebuild to make the model more robust for a general user community. A new set of user documentation was created and a training course was designed to train new users in the physics behind the model and how to use the model to accomplish their goals. In January of 2000, the first open DIRSIG training class was held at RIT and the first group of DIRSIG users received copies of what was titled DIRSIG Release 3.0. (internally, DIRSIG had evolved to version 2.5 prior to the first general release).
The first additions to Release 3.0 included the incorporation of a factory stack plume model developed by EPA and JPL. Another effort resulted in the integration of a simple gas cloud model that was suitable for modeling battlefield releases of chemical weapons. These new modeling features allowed DIRSIG users to simulation simple gas plumes and clouds in order to evaluate gas plume detection, identification and quantification algorithms.
In 2002, we embarked upon two of the most technically challenging improvements to the DIRSIG model since its inception. The first was to update the underlying spectral radiative transport engine to handle polarized photon fluxes. The second was to introduce methods to model Light Detection and Ranging (LIDAR) systems. These two improvements to the model were found to be too fundamental to achieve using the existing source code as a starting point. At this point, a new version of the DIRSIG model was started that replaced the monolithic source code written in "C" with a new flexible infrastructure written in "C++" using modern software design practices. The resulting model features the same fundamental physics embodied in the legacy versions of DIRSIG but in a more flexible form that will allow the model to continue to evolve. This improved version of the model will be released in 2004 as DIRSIG Release 4.
This manual describes the features and behavior of DIRSIG Release 3 versions (3.0 through 3.N). When DIRSIG 4 is released, the manual will be updated to reflect the improved features.