The Terrain Categorization (TERCAT) tool creates an output product in which pixels with similar spectral properties are clumped into classes. These classes may be either user-defined, or automatically generated by the classification algorithm. The TERCAT tool provides all of the standard ENVI classification algorithms, plus an additional algorithm called Winner Takes All.

  1. From the Toolbox, select SPEAR > SPEAR TERCAT. The ENVI TERCAT Wizard displays the File Selection panel.
  2. Click Select Input File, choose a file, then click OK. The input image should be a multispectral file in any format readable by ENVI.
  3. To optionally process only a portion of the scene, click Select Subset. A small Select Spatial Subset dialog appears.
  4. Click Spatial Subset. The standard Select Spatial Subset dialog appears. When finished, click OK to return to the File Selection panel.
  5. By default, output files are saved to the same directory and use the same rootname as the input file, minus any extension. Output files are appended with a unique suffix. To change the directory and/or root filename, click Select Output Root Name.
  6. Click Next. The Atmospheric Correction panel appears.
  7. Perform atmospheric correction.
  8. When atmospheric correction is complete, click Next. The Method Selection panel appears.
  9. Select the classification methods to use:
    • Unsupervised: These methods do not require training data to create a TERCAT, though the classes will not be labeled.
    • Supervised: These methods require you to train the algorithms by creating ROIs that include representative pixels of the desired classes. The Winner Takes All (TERCAT) classification method classifies each pixel to the most commonly occurring class for all supervised methods performed. You must select at least two other supervised methods to run Winner Takes All.
  10. When method selection is complete, click Next. The Select Training Pixels panel appears.
  11. If you selected any supervised classification methods, a natural color composite image is loaded into a display group, and the ROI dialog appears.
  12. If you selected only unsupervised classification methods, you do not need to select training pixels. Click Next to go to the TERCAT Parameters panel (step 13).
  13. For supervised classification, use the ROI Tools dialog to add pixels to the ROI. Create a new ROI for each class to map. Try to collect training pixels uniformly across the image. Give the ROIs descriptive names and representative colors to make them easy to identify.
  14. For supervised classification, click Next. The ROI Selection dialog appears. Select one or more ROIs to map, then click OK. The TERCAT Parameters panel appears.
  15. Each TERCAT method has its own set of advanced parameters that you can adjust, to change the way to run the algorithm. To view the advanced parameters, click Show Advanced Options, then click on the desired tab. In most cases, using the default values produces satisfactory results. For more information on the parameters, refer to See "Classification Tools" for the TERCAT classification method of interest.

    For Winner Takes All, you can apply a weighting to each selected supervised classification method g. By default, each method is set to 1.0 which means each method is treated equally. Optionally, adjust the weights so that favored algorithms have larger values.

  16. To save the rule images created during TERCAT processing, select the Output rule Images? check box available when you show the advanced parameters. The rule images are opened in the Available Bands Lists before processing, but are not used again by the Wizard.
  17. Click Next. The Examine Results panel appears.
  18. When you are finished examining results, click Next in the Examine Results panel, then click Finish to exit the Wizard.

Examining the TERCAT Results

After processing is complete, two dynamically linked display groups appear:

  • One display group contains the reference image.
  • One display group contains the TERCAT results.

Change the image loaded in each display group by selecting a new image from the Reference Display and Results Display #1 drop-down lists. You can also load a third dynamically linked image into a display group for comparison by selecting an image from the Results Display #2 drop-down list. Below shows a reference image (left) and two results display groups (center and right) for comparison of results (imagery courtesy of DigitalGlobe).

The Winner Takes All result is determined by choosing the class for each pixel by the majority vote of all supervised TERCAT methods you chose to run. For example, if three TERCATs determined a pixel was asphalt and two determined it was concrete, the Winner Takes All result is asphalt. In the case of a tie, the majority class of the neighboring pixels are used to classify the pixel in question.

In addition to the TERCAT product, ENVI creates a Winner Takes All (Probability) layer. This layer indicates the level of confidence for each pixel’s classification. In asphalt/concrete example, the probability would be:

The probability for a pixel in which all TERCATs agreed would be 1.0. Therefore, brighter pixels in the probability layer indicate high confidence, and dark pixels indicate low confidence. If many low confidence pixels remain, you could select more training pixels to better represent the low confidence land cover, and you could additionally add a new land cover class. Below shows a reference image (left), Winner Takes All (TERCAT) (center), and Winner Takes All (Probability) (right) (imagery courtesy of DigitalGlobe).

  1. Post-classification processing techniques in ENVI are available directly from the Examine Results panel.
  2. From the TERCATs to process list, select the TERCAT(s) on which to perform the desired processing.
  3. In the Classes to process list, select the class(es) on which to perform the desired processing.
  4. Select the Processing to perform from the drop-down list. Depending on the method you choose, parameters may appear in the area below the Processing to perform drop-down list. The following are available:
    • Majority Analysis: Set the Kernel Size parameter. Larger kernels result in “cleaner” results, though will remove small, legitimate classes. The majority results are appended to the lists in the Display Results section.
    • Clump Classes: Set the Operator Sizeparameter. The clumping results are appended to the lists in the Display Results section.
    • Sieve Classes: Set the Group Min Threshold value to the maximum sized cluster to sieve from the results. The Number of Neighbors value indicates the type of neighborhood examined for determining clusters. Sieved pixels are assigned to the Unclassified class. The sieving results are appended to the lists in the Display Results section.
    • Class Statistics: This method creates a report that shows general statistics about the selected TERCAT(s) and classes, including mean spectra, area covered, and so forth.
    • Class Overlay: The new image is loaded into a new display group.
  5. Click Go to begin processing. Below shows examples of different post-classification filtering options using default parameters

Following is the class statistics dialog showing mean spectrum for each class and general statistical information, such as area covered by each class.

Example class overlay image with trees, roads, and reservoir class overlaid on natural color reference image (imagery courtesy of DigitalGlobe).