Feature extraction¶

Binary morphological operations¶

This application performs binary morphological operations (dilation, erosion, opening and closing) on a mono band image with a specific structuring element (a ball or a cross) having one radius along X and another one along Y. NB: the cross shaped structuring element has a fixed radius equal to 1 pixel in both X and Y directions.

The BinaryMorphologicalOperation application has the following input parameters:

--in the input image to be filtered

--channel the selected channel index in the input image to be

processed (default value is 1)

--structype the choice of the structuring element type

(ball/cross) (default value is ball)

-(-structype.ball.xradius) the ball structuring element X Radius

(only if structype==ball) (default value is 5 pixels)

-(-structype.ball.yradius) the ball structuring element Y Radius

(only if structype==ball) (default value is 5 pixels)

--filter the choice of the morphological operation

(dilate/erode/opening/closing) (default value is dilate)

-(-filter.dilate.foreval) the foreground value for the dilation

(idem for filter.erode/opening/closing) (default value is 1)

-(-filter.dilate.backval) the background value for the dilation

(idem for filter.erode/opening/closing) (default value is 0)

--out the output filtered image

The application can be used like this:

otbcli_BinaryMorphologicalOperation  -in                     InputImage
                                     -channel                1
                                     -structype              ball
                                     -structype.ball.xradius 10
                                     -structype.ball.yradius 5
                                     -filter                 opening
                                     -filter.opening.foreval 1.0
                                     -filter.opening.backval 0.0
                                     -out                    OutputImage

Gray scale morphological operations¶

This application performs morphological operations (dilation, erosion, opening and closing) on a gray scale mono band image with a specific structuring element (a ball or a cross) having one radius along X and another one along Y. NB: the cross shaped structuring element has a fixed radius equal to 1 pixel in both X and Y directions.

The GrayScaleMorphologicalOperation application has the following input parameters:

--in the input image to be filtered

--channel the selected channel index in the input image to be

processed (default value is 1)

--structype the choice of the structuring element type

(ball/cross) (default value is ball)

-(-structype.ball.xradius) the ball structuring element X Radius

(only if structype==ball) (default value is 5 pixels)

-(-structype.ball.yradius) the ball structuring element Y Radius

(only if structype==ball) (default value is 5 pixels)

--filter the choice of the morphological operation

(dilate/erode/opening/closing) (default value is dilate)

--out the output filtered image

The application can be used like this:

otbcli_GrayScaleMorphologicalOperation  -in                     InputImage
                                        -channel                1
                                        -structype              ball
                                        -structype.ball.xradius 10
                                        -structype.ball.yradius 5
                                        -filter                 opening
                                        -out                    OutputImage

Haralick texture features¶

This application computes Haralick, advanced and higher order texture features on every pixel in the selected channel of the input image. The output image is multi band with a feature per band.

The HaralickTextureExtraction application has the following input parameters:

--in the input image to compute the features on

--channel the selected channel index in the input image to be

processed (default value is 1)

--texture the texture set selection [simple/advanced/higher]

(default value is simple)

--parameters.min the input image minimum (default value is 0)

--parameters.max the input image maximum (default value is 255)

--parameters.xrad the X Radius of the processing neighborhood

(default value is 2 pixels)

--parameters.yrad the Y Radius of the processing neighborhood

(default value is 2 pixels)

--parameters.xoff the \(\Delta\)X Offset for the

co-occurrence computation (default value is 1 pixel)

--parameters.yoff the \(\Delta\)Y Offset for the

co-occurrence computation (default value is 1 pixel)

--parameters.nbbin the number of bin per axis for histogram

generation (default value is 8)

--out the output multi band image containing the selected texture

features (one feature per band)

The available values for -texture with their relevant features are:

--texture=simple: In this case, 8 local Haralick textures features

will be processed. The 8 output image channels are: Energy, Entropy, Correlation, Inverse Difference Moment, Inertia, Cluster Shade, Cluster Prominence and Haralick Correlation. They are provided in this exact order in the output image. Thus, this application computes the following Haralick textures over a neighborhood with user defined radius. To improve the speed of computation, a variant of Grey Level Co-occurrence Matrix(GLCM) called Grey Level Co-occurrence Indexed List (GLCIL) is used. Given below is the mathematical explanation on the computation of each textures. Here \(g(i, j)\) is the frequency of element in the GLCIL whose index is i, j. GLCIL stores a pair of frequency of two pixels taken from the given offset and the cell index (i, j) of the pixel in the neighborhood window. :(where each element in GLCIL is a pair of pixel index and it’s frequency, \(g(i, j)\) is the frequency value of the pair having index is i, j).

“Energy” \(= f_1 = \sum_{i, j}g(i, j)^2\)

“Entropy” \(= f_2 = -\sum_{i, j}g(i, j) \log_2 g(i, j)\), or 0 if \(g(i, j) = 0\)

“Correlation” \(= f_3 = \sum_{i, j}\frac{(i - \mu)(j - \mu)g(i, j)}{\sigma^2}\)

“Inverse Difference Moment” \(= f_4 = \sum_{i, j}\frac{1}{1 + (i - j)^2}g(i, j)\)

“Inertia” \(= f_5 = \sum_{i, j}(i - j)^2g(i, j)\) (sometimes called “contrast”)

“Cluster Shade” \(= f_6 = \sum_{i, j}((i - \mu) + (j - \mu))^3 g(i, j)\)

“Cluster Prominence” \(= f_7 = \sum_{i, j}((i - \mu) + (j - \mu))^4 g(i, j)\)

“Haralick’s Correlation” \(= f_8 = \frac{\sum_{i, j}(i, j) g(i, j) -\mu_t^2}{\sigma_t^2}\) where \(\mu_t\) and \(\sigma_t\) are the mean and standard deviation of the row (or column, due to symmetry) sums. Above, \(\mu =\) (weighted pixel average) \(= \sum_{i, j}i \cdot g(i, j) = \sum_{i, j}j \cdot g(i, j)\) (due to matrix symmetry), and \(\sigma =\) (weighted pixel variance) \(= \sum_{i, j}(i - \mu)^2 \cdot g(i, j) = \sum_{i, j}(j - \mu)^2 \cdot g(i, j)\) (due to matrix symmetry).

--texture=advanced: In this case, 10 advanced texture features

will be processed. The 10 output image channels are: Mean, Variance, Dissimilarity, Sum Average, Sum Variance, Sum Entropy, Difference of Entropies, Difference of Variances, IC1 and IC2. They are provided in this exact order in the output image. The textures are computed over a sliding window with user defined radius.

To improve the speed of computation, a variant of Grey Level Co-occurrence Matrix(GLCM) called Grey Level Co-occurrence Indexed List (GLCIL) is used. Given below is the mathematical explanation on the computation of each textures. Here \(g(i, j)\) is the frequency of element in the GLCIL whose index is i, j. GLCIL stores a pair of frequency of two pixels taken from the given offset and the cell index (i, j) of the pixel in the neighborhood window. :(where each element in GLCIL is a pair of pixel index and it’s frequency, \(g(i, j)\) is the frequency value of the pair having index is i, j).

“Mean” \(= \sum_{i, j}i g(i, j)\)

“Sum of squares: Variance” \(= f_4 = \sum_{i, j}(i - \mu)^2 g(i, j)\)

“Dissimilarity” \(= f_5 = \sum_{i, j}(i - j) g(i, j)^2\)

“Sum average” \(= f_6 = -\sum_{i}i g_{x+y}(i)\)

“Sum Variance” \(= f_7 = \sum_{i}(i - f_8)^2 g_{x+y}(i)\)

“Sum Entropy” \(= f_8 = -\sum_{i}g_{x+y}(i) log (g_{x+y}(i))\)

“Difference variance” \(= f_10 = variance of g_{x-y}(i)\)

“Difference entropy” \(= f_11 = -\sum_{i}g_{x-y}(i) log (g_{x-y}(i))\)

“Information Measures of Correlation IC1” \(= f_12 = \frac{f_9 - HXY1}{H}\)

“Information Measures of Correlation IC2” \(= f_13 = \sqrt{1 - \exp{-2}|HXY2 - f_9|}\)

Above, \(\mu =\) (weighted pixel average) \(= \sum_{i, j}i \cdot g(i, j) = \sum_{i, j}j \cdot g(i, j)\) (due to matrix summetry), and

\(g_{x+y}(k) = \sum_{i}\sum_{j}g(i)\) where \(i+j=k\) and \(k = 2, 3, .., 2N_{g}\) and

\(g_{x-y}(k) = \sum_{i}\sum_{j}g(i)\) where \(i-j=k\) and \(k = 0, 1, .., N_{g}-1\)

--texture=higher: In this case, 11 local higher order statistics

texture coefficients based on the grey level run-length matrix will be processed. The 11 output image channels are: Short Run Emphasis, Long Run Emphasis, Grey-Level Nonuniformity, Run Length Nonuniformity, Run Percentage, Low Grey-Level Run Emphasis, High Grey-Level Run Emphasis, Short Run Low Grey-Level Emphasis, Short Run High Grey-Level Emphasis, Long Run Low Grey-Level Emphasis and Long Run High Grey-Level Emphasis. They are provided in this exact order in the output image. Thus, this application computes the following Haralick textures over a sliding window with user defined radius: (where \(p(i, j)\) is the element in cell i, j of a normalized Run Length Matrix, \(n_r\) is the total number of runs and \(n_p\) is the total number of pixels):

“Short Run Emphasis” \(= SRE = \frac{1}{n_r} \sum_{i, j}\frac{p(i, j)}{j^2}\)

“Long Run Emphasis” \(= LRE = \frac{1}{n_r} \sum_{i, j}p(i, j) * j^2\)

“Grey-Level Nonuniformity” \(= GLN = \frac{1}{n_r} \sum_{i} \left( \sum_{j}{p(i, j)} \right)^2\)

“Run Length Nonuniformity” \(= RLN = \frac{1}{n_r} \sum_{j} \left( \sum_{i}{p(i, j)} \right)^2\)

“Run Percentage” \(= RP = \frac{n_r}{n_p}\)

“Low Grey-Level Run Emphasis” \(= LGRE = \frac{1}{n_r} \sum_{i, j}\frac{p(i, j)}{i^2}\)

“High Grey-Level Run Emphasis” \(= HGRE = \frac{1}{n_r} \sum_{i, j}p(i, j) * i^2\)

“Short Run Low Grey-Level Emphasis” \(= SRLGE = \frac{1}{n_r} \sum_{i, j}\frac{p(i, j)}{i^2 j^2}\)

“Short Run High Grey-Level Emphasis” \(= SRHGE = \frac{1}{n_r} \sum_{i, j}\frac{p(i, j) * i^2}{j^2}\)

“Long Run Low Grey-Level Emphasis” \(= LRLGE = \frac{1}{n_r} \sum_{i, j}\frac{p(i, j) * j^2}{i^2}\)

“Long Run High Grey-Level Emphasis” \(= LRHGE = \frac{1}{n_r} \sum_{i, j} p(i, j) i^2 j^2\)

The application can be used like this:

otbcli_HaralickTextureExtraction  -in             InputImage
                                  -channel        1
                                  -texture        simple
                                  -parameters.min 0
                                  -parameters.max 255
                                  -out            OutputImage

SFS texture extraction¶

This application computes Structural Feature Set textures on every pixel in the selected channel of the input image. The output image is multi band with a feature per band. The 6 output texture features are SFS’Length, SFS’Width, SFS’PSI, SFS’W-Mean, SFS’Ratio and SFS’SD. They are provided in this exact order in the output image.

It is based on line direction estimation and described in the following publication. Please refer to Xin Huang, Liangpei Zhang and Pingxiang Li publication, Classification and Extraction of Spatial Features in Urban Areas Using High-Resolution Multispectral Imagery. IEEE Geoscience and Remote Sensing Letters, vol. 4, n. 2, 2007, pp 260-264.

The texture is computed for each pixel using its neighborhood. User can set the spatial threshold that is the max line length, the spectral threshold that is the max difference authorized between a pixel of the line and the center pixel of the current neighborhood. The adjustement constant alpha and the ratio Maximum Consideration Number, which describes the shape contour around the central pixel, are used to compute the \(w - mean\) value.

The SFSTextureExtraction application has the following input parameters:

--in the input image to compute the features on

--channel the selected channel index in the input image to be

processed (default value is 1)

--parameters.spethre the spectral threshold (default value is 50)

--parameters.spathre the spatial threshold (default value is 100

pixels)

--parameters.nbdir the number of directions (default value is 20)

--parameters.alpha the alpha value (default value is 1)

--parameters.maxcons the ratio Maximum Consideration Number

(default value is 5)

--out the output multi band image containing the selected texture

features (one feature per band)

The application can be used like this:

otbcli_SFSTextureExtraction -in             InputImage
                            -channel        1
                            -out            OutputImage