ENVI includes two unsupervised classification methods, ISODATA and K-Mean.
ISODATA (Iterative Self-Orgnizing Data Analysis Technique) is a repetitive self-organizing data analysis technique that calculates the uniformly distributed class mean value in the data space, and then uses the minimum distance technique to iteratively aggregate the remaining pixels, and recalculates the mean value for each iteration , and according to the obtained new mean, the pixels are reclassified.
K-Means uses the cluster analysis method to randomly find clusters with similar cluster similarities, that is, the center position, which is calculated by using a "central object" (center of gravity) obtained by using the mean value of the objects in each cluster , and then iteratively reconfigure them to complete the classification process.
Unsupervised classification can generally be divided into 4 processes in general: performing unsupervised classification, class definition, merging subclasses, and evaluating classification results.
1
1) Perform unsupervised classification
The following takes TM images as an example to introduce the unsupervised classification operation process.
Due to the different parameter settings of ISODATA and K-Mean, the two classification methods are introduced separately.
-
ISODATA
In the Toolbox toolbox, double-click the Classification/Unsupervised Classification/IsoData Classification tool. In the Classification Input File dialog box, select the classified TM image file, click the OK button, and open the ISODATA Parameters dialog box. Set the following in the ISODATA Parameters dialog box. parameter.
(1) Range of categories (Number of Classes: Min, Max): Generally, the minimum number of input cannot be less than the final classification number, and the maximum number is 2~3 times of the final classification number. Min: 5, Max: 15.
(2) Maximum Iterations (Maximum Iterations): 15. The larger the number of iterations, the more accurate the results and the longer the calculation time.
(3) Change Threshold (Change Threshold): 5. When the number of changed pixels of each category is less than the threshold, the iterative process ends. The smaller this value is, the more accurate the result is and the greater the amount of computation is.
(4) Minimun#Pixel in Class: Type in the minimum number of pixels required to form a class. If the number of cells in a class is less than the minimum number of cells, the class will be deleted and the cells in it will be grouped into the closest class.
(5) Maximum classification standard deviation (Maximum Class Stdv): 1. In units of pixel values, if the standard deviation of a class is greater than the threshold, the class will be split into two classes.
(6) Minimum Class Distance between category means (Minimum Class Distance): 5. In pixels, if the distance between class means is less than the input minimum value, the categories will be merged.
(7) Maximum number of merge categories (Maximum#Merge Pairs): 2.
(8) The maximum standard deviation from the category mean
(Maximum Stdev From Mean): It is optional. Filter the pixels smaller than this standard deviation to participate in the classification.
(9) Maximum Distance Error: It is optional. Screen the pixels smaller than the maximum distance error to participate in the classification.
(10) Select the output path and file name, and click the OK button to perform unsupervised classification.
2.K-Means
In the Toolbox toolbox, double-click the Classification/Unsupervised Classification/K-Means Classifi-cation tool, in the Classification Input File dialog box, select the classified TM image file, click OK; open the K-Means Parameters dialog box, set K below -Parameters in the Means Parameters dialog.
(1) Number of Classes (Number of Classes): 15. Generally, it is 2 to 3 times the number of final classifications to be output.
(2) Maximum Iterations: 10.
The larger the number of iterations, the more accurate the result and the longer the operation time.
(3) The maximum standard deviation from the category mean (Maximum Stdev From Mean): is optional. Filter the pixels smaller than this standard deviation to participate in the classification.
(4) Maximum Distance Error: It is optional. Screen the pixels smaller than the maximum distance error to participate in the classification.
(5) Select the output path and file name, and click the OK button to perform unsupervised classification.
Unsupervised classification results:
The background value can be removed by clipping;
After performing unsupervised classification, a preliminary classification result is obtained, which requires category definition and subcategory merging.
First step category definition
Class definitions can be based on visual interpretation on higher resolution imagery, or based on field survey data.
(1) Open the visual interpretation basemap and display it in the window (if any).
(2) Open the TM unsupervised classification result image and display it in the window.
(3) In the Layer Manager (Layer Manager), right-click on Classes to select the Hide All Classes menu, and then check Classl, only one classification category will be displayed, and the name of the class can be identified visually.
(4) In the Toolbox toolbox, double-click the Raster Management/Edit ENVI Header tool, and select TM unsupervised classification results in the file input dialog box.
(5) In the Header Info panel, select Edit Attributes-→Classification Info, and click the OK button by default to open the Class Color Map Editing panel.
(6) In the Class Color Map Editing panel, select the corresponding category, enter the redefined category name in Class Name, and modify the display color at the same time.
(7) Repeat steps (3)~(6) to define other categories.
Tip: Since the NVI5.1 version automatically closes the image every time the image header file information is modified, it is best to modify the classification name and color in the Class Color Map Editing panel after all classifications have been interpreted in step (3).
The second step merges subclasses
When selecting the number of unsupervised classification categories, generally choose 2 to 3 times the number of final results, so after defining the categories, the same categories need to be merged.
(1) In the Toolbox toolbox, double-click the Classification/Post Classification/Combine Classes tool.
In the Combine Classes Input File dialog box, select the defined classification result, click the OK button, and open the Combine Classes Parameters panel.
(2) In the Combine Classes Parameters panel, select the combined category from the Select Input Class, select the combined category from the Select Out Class, and click the Add Combination button to add to the combined scheme. The combined scheme is displayed in the Combined Classes list. Click an item in the Combined Classes list to remove it from the scheme.
(3) After the merge plan is established, click the OK button to open the Combine Classes Output dialog box, and select "Yes" in the Re-move Empty Classes item to remove the blank class.
(4) Select the path and file name of the output merged result, and click the OK button to execute the merge.
The process of evaluating the classification results is the same as the method of supervised classification. You can refer to the content of the previous supervised classification evaluation, so I won’t go into details here.