The DIRSIG model has been developed to aid the remote sensing community in several different application areas.
The process of designing and building a new imaging instrument merges personnel from a broad range of disciplines including optical sciences, mechanical engineering and electrical engineering. The performance evaluation of a potential imaging system is commonly performed on a subsystem basis and evaluated using performance metrics with respect to system performance specifications. Once the engineers are satisfied with subsystem performance, system components are built and ultimately integrated for final system testing.
The potential role of a synthetic image generation tool in this environment is as a virtual instrument test bed. A robust imaging instrument simulation tool would allow instrument builders to input the detailed system specifications and produce synthetic image data from the virtual instrument. This data can be used to evaluate system performance using conventional image quality metrics in a similar manner that is used today. However, in addition, the synthetic image data allows the user to test the performance of algorithms on the instrument data. This allows the designers to evaluate system performance in terms of output data products. For example, if a system is being designed to produce "crop health maps", the output of the crop health algorithms can be compared to the model inputs to evaluate the accuracy of the produced maps and how the system can be tuned to improve the output products.
The recent trend in multi- and hyper-spectral algorithms is to utilize physics based models to "train" or "constrain" these algorithms. For example, modern atmospheric compensation algorithms utilize atmospheric radiative transfer codes (such as MODTRAN) to predict the correlation between absorption features in the atmospheric transmission and spectral absorption features observed in hyper-spectral imagery to remove the atmospheric contribution.
An extension to this idea is to model the possible observation space of a specific target to improve a target detection algorithm. For these applications, the spectral reflectance of the target of interest is usually pulled from a database, and the challenge is to determine what the expected radiance spectrum will be as observed at the sensor from this target? That will depend on a wide variety of factors. What is the atmosphere between the target and sensor? Is the target sub-pixel? If so, how much of a pixel does the target fill? What is the background that fills the rest of the pixel? Is the target near other objects? Do these objects "color" the target by reflecting light onto the target? What will the temperature of the target be? Use of a rigorous scene simulation tool allows the algorithm developer to develop an expected signature space that the target is likely to occupy. If the algorithm is constrained to identify targets within this predicted space, then false alarms may be eliminated.
Once an algorithm has been designed and implemented, the difficult task of testing and evaluating the performance of the algorithm must be addressed. Historically, this is performed using sets of test images that have been extensively ground truthed so that the results of the algorithm can be evaluated. The limitation in this approach is the availability of well characterized test data sets. For example, what if a data set is not available that contains the target of interest? Another limitation is the diversity in available data sets. Is the algorithm robust if the performance can be verified on only a single scene type (forested, urban, etc.)? Or under a single set of atmospheric conditions? Or for a single time of day?
When a synthetic image generation model is being utilized, the user can create a large set of test images that feature variations in scene type, time of data, atmospheric conditions, etc. so that the algorithm can be evaluated under more conditions. The use of synthetic data sets provides the user with per-pixel ground truth which allows the algorithm performance to be evaluated at every pixel rather than at a few selected sites that were ground truthed. Synthetic data can also be generated at a significant cost savings over extensive field collection campaigns.
At no point in time, however, should we expect the state-of-art image generation models (including DIRSIG) to be robust enough to be utilized exclusively for testing. However, synthetic imagery provides a valuable complement to real imagery or as a possible surrogate when real data are not available.
Human analysts are trained to look for specific targets, activities or phenomenology by studying large sets of imagery that have been reviewed by experienced analysts or for which the ground conditions were documented by some other means. The image sets for use in analyst training can be creating from archived imagery or from specific data collection efforts. However, how does one train an analyst to look for a specific target or event when stock imagery is not available to train with? In this instance, synthetic data may provide a solution as a tool with which the manifestation of the event may be explored.
In addition, a robust synthetic image generation tool may play a roll as a hypothesis testing tool for human analysts. For example, the analyst may identify a specific element in real image data that they believe to be the target or event of interest. To build confidence in their assessment, they may choose to model the element of interest in the context of the scene around it to verify the appearance or manifestation of the element that they have observed is actually the element of interest.