Segmentation

Performs segmentation of an image, and output either a raster or a vector file. In vector mode, large input datasets are supported.

Description

This application allows one to perform various segmentation algorithms on a multispectral image. Available segmentation algorithms are two different versions of Mean-Shift segmentation algorithm (one being multi-threaded), simple pixel based connected components according to a user-defined criterion, and watershed from the gradient of the intensity (norm of spectral bands vector). The application has two different modes that affects the nature of its output.

In raster mode, the output of the application is a classical image of unique labels identifying the segmented regions. The labeled output can be passed to the ColorMapping application to render regions with contrasted colours. Please note that this mode loads the whole input image into memory, and as such can not handle large images.

To segment large data, one can use the vector mode. In this case, the output of the application is a vector file or database. The input image is split into tiles (whose size can be set using the tilesize parameter), and each tile is loaded, segmented with the chosen algorithm, vectorized, and written into the output file or database. This piece-wise behavior ensure that memory will never get overloaded, and that images of any size can be processed. There are few more options in the vector mode. The simplify option allows simplifying the geometry (i.e. remove nodes in polygons) according to a user-defined tolerance. The stitch option tries to stitch together the polygons corresponding to segmented region that may have been split by the tiling scheme.

Parameters

Input Image -in image Mandatory
The input image to segment

Segmentation algorithm -filter [meanshift|cc|watershed|mprofiles] Default value: meanshift
Choice of segmentation algorithm (mean-shift by default)

• Mean-Shift
  OTB implementation of the Mean-Shift algorithm (multi-threaded).
• Connected components
  Simple pixel-based connected-components algorithm with a user-defined connection condition.
• Watershed
  The traditional watershed algorithm. The height function is the gradient magnitude of the amplitude (square root of the sum of squared bands).
• Morphological profiles based segmentation
  Segmentation based on morphological profiles, as described in Martino Pesaresi and Jon Alti Benediktsson, Member, IEEE: A new approach for the morphological segmentation of high resolution satellite imagery. IEEE Transactions on geoscience and remote sensing, vol. 39, NO. 2, February 2001, p. 309-320.

Mean-Shift options

Spatial radius -filter.meanshift.spatialr int Default value: 5
Spatial radius of the neighborhood.

Range radius -filter.meanshift.ranger float Default value: 15
Range radius defining the radius (expressed in radiometry unit) in the multispectral space.

Mode convergence threshold -filter.meanshift.thres float Default value: 0.1
Algorithm iterative scheme will stop if mean-shift vector is below this threshold or if iteration number reached maximum number of iterations.

Maximum number of iterations -filter.meanshift.maxiter int Default value: 100
Algorithm iterative scheme will stop if convergence hasn’t been reached after the maximum number of iterations.

Minimum region size -filter.meanshift.minsize int Default value: 100
Minimum size of a region (in pixel unit) in segmentation. Smaller clusters will be merged to the neighboring cluster with the closest radiometry. If set to 0 no pruning is done.

Connected components options

Condition -filter.cc.expr string Mandatory
User defined connection condition, written as a mathematical expression. Available variables are p(i)b(i), intensity_p(i) and distance (example of expression: distance < 10 )

Watershed options

Depth Threshold -filter.watershed.threshold float Default value: 0.01
Depth threshold Units in percentage of the maximum depth in the image.

Flood Level -filter.watershed.level float Default value: 0.1
flood level for generating the merge tree from the initial segmentation (between 0 and 1)

Morphological profiles based segmentation options

Profile Size -filter.mprofiles.size int Default value: 5
Size of the profiles

Initial radius -filter.mprofiles.start int Default value: 1
Initial radius of the structuring element (in pixels)

Radius step -filter.mprofiles.step int Default value: 1
Radius step along the profile (in pixels)

Threshold of the final decision rule -filter.mprofiles.sigma float Default value: 1
Profiles values under the threshold will be ignored.

---

Processing mode -mode [vector|raster] Default value: vector
Choice of processing mode, either raster or large-scale.

• Tile-based large-scale segmentation with vector output
  In this mode, the application will output a vector file or database, and process the input image piecewise. This allows performing segmentation of very large images.
• Standard segmentation with labeled raster output
  In this mode, the application will output a standard labeled raster. This mode can not handle large data.

Tile-based large-scale segmentation with vector output options

Output vector file -mode.vector.out filename [dtype] Mandatory
The output vector file or database (name can be anything understood by OGR)

Writing mode for the output vector file -mode.vector.outmode [ulco|ovw|ulovw|ulu] Default value: ulco
This allows one to set the writing behaviour for the output vector file. Please note that the actual behaviour depends on the file format.

• Update output vector file, only allow creating new layers
  The output vector file is opened in update mode if existing. If the output layer already exists, the application stops, leaving it untouched.
• Overwrite output vector file if existing.
  If the output vector file already exists, it is completely destroyed (including all its layers) and recreated from scratch.
• Update output vector file, overwrite existing layer
  The output vector file is opened in update mode if existing. If the output layer already exists, it si completely destroyed and recreated from scratch.
• Update output vector file, update existing layer
  The output vector file is opened in update mode if existing. If the output layer already exists, the new geometries are appended to the layer.

Mask Image -mode.vector.inmask image
Only pixels whose mask value is strictly positive will be segmented.

8-neighbor connectivity -mode.vector.neighbor bool Default value: false
Activate 8-Neighborhood connectivity (default is 4).

Stitch polygons -mode.vector.stitch bool Default value: true
Scan polygons on each side of tiles and stitch polygons which connect by more than one pixel.

Minimum object size -mode.vector.minsize int Default value: 1
Objects whose size is below the minimum object size (area in pixels) will be ignored during vectorization.

Simplify polygons -mode.vector.simplify float Default value: 0.1
Simplify polygons according to a given tolerance (in pixel). This option allows reducing the size of the output file or database.

Layer name -mode.vector.layername string Default value: layer
Name of the layer in the vector file or database (default is Layer).

Geometry index field name -mode.vector.fieldname string Default value: DN
Name of the field holding the geometry index in the output vector file or database.

Tiles size -mode.vector.tilesize int Default value: 1024
User defined tiles size for tile-based segmentation. Optimal tile size is selected according to available RAM if null.

Starting geometry index -mode.vector.startlabel int Default value: 1
Starting value of the geometry index field

OGR options for layer creation -mode.vector.ogroptions string1 string2...
A list of layer creation options in the form KEY=VALUE that will be passed directly to OGR without any validity checking. Options may depend on the file format, and can be found in OGR documentation.

Standard segmentation with labeled raster output options

Output labeled image -mode.raster.out image [dtype] Mandatory
The output labeled image.

Examples

From the command-line:

# Example of use with vector mode and watershed segmentation
otbcli_Segmentation -in QB_Toulouse_Ortho_PAN.tif -mode vector -mode.vector.out SegmentationVector.sqlite -filter watershed

# Example of use with raster mode and mean-shift segmentation
otbcli_Segmentation -in QB_Toulouse_Ortho_PAN.tif -mode raster -mode.raster.out SegmentationRaster.tif uint16 -filter meanshift

From Python:

# Example of use with vector mode and watershed segmentation
import otbApplication

app = otbApplication.Registry.CreateApplication("Segmentation")

app.SetParameterString("in", "QB_Toulouse_Ortho_PAN.tif")
app.SetParameterString("mode","vector")
app.SetParameterString("mode.vector.out", "SegmentationVector.sqlite")
app.SetParameterString("filter","watershed")

app.ExecuteAndWriteOutput()

# Example of use with raster mode and mean-shift segmentation
import otbApplication

app = otbApplication.Registry.CreateApplication("Segmentation")

app.SetParameterString("in", "QB_Toulouse_Ortho_PAN.tif")
app.SetParameterString("mode","raster")
app.SetParameterString("mode.raster.out", "SegmentationRaster.tif")
app.SetParameterOutputImagePixelType("mode.raster.out", 3)
app.SetParameterString("filter","meanshift")

app.ExecuteAndWriteOutput()

Limitations

In raster mode, the application can not handle large input images. Stitching step of vector mode might become slow with very large input images. MeanShift filter results depends on the number of threads used. Watershed and multiscale geodesic morphology segmentation will be performed on the amplitude of the input image. This application does not handle no data values. No data pixels will be treated as regular pixels, This may lead to unexpected segmentation results and crashes.

See also

LargeScaleMeanShift