Methods

OVERVIEW

DIRSIG stands for Digital Imaging and Remote Sensing Image Generation Model; it is a synthetic image generation system. The goal of this project is to provide an assessment of DIRSIG’s texture characterization techniques for a grass target³. DIRSIG will be used to generate a synthetic image of grass, which will then be compared to a truth image of grass. In an attempt to improve the DIRSIG output, an investigation will be done into different techniques that may be employed to change DIRSIG’s texture characterization routines.
      Distinct methods that will be examined include expanding DIRSIG’s database of spectra for grass targets. The results of these methods will be tested against the truth data and compared with DIRSIG’s current method for the texture characterization of a grass target. Research will concentrate in part on developing an algorithm that examines two additional areas of the reflectance spectra. Z-scores in three different areas will be compared to determine which spectra should be mapped into a particular grass pixel.
            The goal of this research is to assess the current techniques used by DIRSIG and to investigate the effects that several alternate or additional techniques have on the quality of DIRSIG’s outputs. A significant amount of research will be done on texture characterization techniques for synthetic images in an attempt to determine a better technique to be implemented into DIRSIG.

ASSESMENT

The Target

The target examined in this research is a grass target. A grass target was chosen because of the high variance that exists within images of grass; there are also many real world applications that apply to the imaging of grass as part of a scene.

Acquisition of Images

    In order to make an assessment of DIRSIG’s texture characterization techniques³, there must be a truth image to compare with DIRSIG’s synthetic imagery. Two scenes were chosen as test scenes to be simulated in DIRSIG in order to incorporate two different methods for building a spectral database. The first was an image taken by the airborne spectrometer MISI. A subsection of the image was taken; it was selected to include only a grass target. An image taken by the airborne sensor HYDICE was the second image examined, once again a subset of the image containing only a grass target was used. Both images were obtained from the Digital Image and Remote Sensing group at Rochester Institute of Technology.

Simulation of MISI in DIRSIG

A specific MISI image was simulated in DIRSIG; spectral data for the database to be used with this scene was collected using a hand held spectrometer. A model of the MISI airborne hyperspectral imaging spectrometer was incorporated into DIRSIG in order to simulate real world data. The generated DIRSIG images of grass were than tested utilizing ENVI’s Principle Components analysis in order to determine the quality of the DIRSIG image, results for the DIRSIG image were than compared with results found for a MISI image taken under the same conditions of the same spectra.

The generated DIRSIG image of grass was tested utilizing ENVI’s Principle Components analysis in order to determine the quality of the DIRSIG image. The DIRSIG image was than compared with those found for the MISI image of real grass. Once the quality of the DIRSIG image was quantized, distinct methods were employed to change DIRSIG’s texture characterization routines in an attempt to improve the model’s output.

Simulation of HYDICE in DIRSIG

    The second scene that was used was that of a HYDICE image. In this case the spectra for DIRSIG’s spectral database were extracted directly from the scene data. This created an optimal set up for testing DIRSIG’s routines, since we knew that an exact fit for the spectra at each pixel existed within the spectral database. In this manor we were able to test DIRSIG’s ability to map these curves to the appropriate pixel.  The spectral database was incorporated into a DIRSIG run with available models for the HYDICE sensor. The DIRSIG output image was than compared to the actual HYDICE image using ENVI’s Principle Components analysis.

METHODS FOR IMPROVEMENT

Once the quality of the DIRSIG image has been assessed, distinct methods were employed to change DIRSIG’s texture characterization routines in an attempt to improve the model’s output. The improved methods that were employed included doubling the size of DIRSIG's spectral database for grass targets and creating an improved z-score algorithm that was implemented in DIRSIG. This algorithm checks z-score statistics in two additional regions of the reflectance spectra in addition to the bandpass that DIRSIG was previously using, in order to determine which spectra in the database will be used to characterize a pixel of grass.

Expanding DIRSIG’s Spectral Database

In the Simulation of MISI

  In the simulation of the MISI image, the spectral database was made up of the original set of thirty spectra that was collected for the grass target. In order to examine what effect the size of the spectral database would have on the quality of the DIRSIG image, the spectral database was expanded.  A supporting utility for DIRSIG, expand_emissivity_file², was used to accomplish this. This procedure uses the input emissivity curves to generate another set of curves, the number of curves the user determines, that have the same multi-spectral statistics as the original set. The spectral database for MISI was expanded from a set of thirty curves to a set of two hundred.

In the Simulation of HYDICE

In the simulation of the HYDICE image, spectra for the database were extracted directly from the scene. In order to assess DIRSIG’s ability to create a set of spectra that correspond accordingly, the expand_emissivity_file² utility was used again in order to create two separate sets of curves to be tested in DIRSIG. The original extracted set of reflectance curves consisted of two hundred and fifty six curves. The first set of spectra that was constructed using the expand_emissivity_file² utility was composed of ten spectra. The second set was constructed to have the same number of curves as the set that was originally extracted.

Incorporation of Additional Bandpass Regions

DIRSIG applies texture to each pixel spectrally.  Current techniques² use a texture image taken from one bandpass to select a reflectance curve from a large database of reflectance curves in order to represent the spectral variations within the given material. Means and standard deviations are used to calculate z-scores² for both the grayscale texture image and for each curve in the database over the same bandpass. Z-scores are compared to select a curve. Once a curve is selected, that reflectance curve is utilized in the computations for the given pixel in any spectral region being modeled.

    First, a quantitative measurement is made of the texture image that the user has imputed. This quantitative measurement is made through statistical means. The average mean and standard deviation are computed for the texture image. For each pixel in the grayscale texture image, the brightness is used in the form of the pixel's digital count, in order to calculate a z-score for that pixel in the texture image, as shown in Equation 1.

Equation (1) ²

    DIRSIG relates the spectral database of curves to the texute image by examining each of the curves in the same bandpass region that the texture image was taken. For the reflectance database, a method for comparison was derived that ranked each curve according to their relationship to the mean of the family of curves. The mean reflectance value of each curve was computed as described in Equation 2:

Equation (2) ²

Note: N is the number of curves;  i = 1, N where s_avg,i  is the average reflectance over the bandpass from l_min to l_max curve i in the set of N curves, n_i is the spectral reflectance for the i^th curve at l, and n_i  is the number of points across the bandpass for the i^th curve.

The mean and standard deviation for the bandpass averages are then computed as described in Equation 3 and Equation 4.

Equation (3) ²

Equation (4) ²

the Z-score for a curve, i, is than calculated as described in Equation 5:

Equation (5) ²

The incorporation of additional bandpass regions utilized the use of multiple bandpass regions with this current technique. For each bandpass to be examined, a texture image was created in DIRSIG. These bands were than specified in DIRSIG using the current technique for the input of one texture band. This resulted in a list of z-scores to be examined for each spectrum in the database, DIRSIG than selects the spectra that fits all of the bandpass regions most closely. This results in a list of parameters for DIRSIG to fit for a spectrum instead of the one parameter that it was examining before. In this way a number of regions could be examined to ensure that the accuracy in one region was not disregarded due to the need for accuracy in another. The use of one, two, and three bandpasses in DIRSIG's texture characterization routines were examined for comparison.

Table of Contents