Image processing and information extraction

Simple calculus with channels

The BandMath application provides a simple and efficient way to perform band operations. The command line application and the corresponding Monteverdi module (shown in the section [Band:sub:math module]) are based on the same standards. It computes a band wise operation according to a user defined mathematical expression. The following code computes the absolute difference between first bands of two images.

otbcli_BandMath -il input_image_1 input_image_2
                -exp "abs(im1b1 - im2b1)"
                -out output_image

The naming convention “im[x]b[y]” designates the yth band of the xth input image.

The BandMath application embeds built-in operators and functions listed in muparser documentation thus allowing a vast choice of possible operations.

Images with no-data values

Image files can contain a no-data value in their metadata. It represents a special pixel value that should be treated as “no data available for this pixel”. For instance, SRTM tiles use a particular no-data value of -32768 (usually found on sea areas).

On multiband images, the no-data values are handled independently for each band. The case of an image with no-data values defined only for a subset of its bands is supported.

This metadata is now handled by OTB image readers and writer (using the GDAL driver). The no-data value can be read from an image files and stored in the image metadata dictionary. It can also be exported by image writers. The OTB filters that produce a no-data value are able to export this value so that the output file will store it.

An application has been created to manage the no-data value. The application has the following features:

  • Build a mask corresponding to the no-data pixels in the input image: it gives you a binary image of the no-data pixels in your input image.
  • Change the no-data value of the input image: it will change all pixels that carry the old no-data value to the new one and update the metadata
  • Apply an external mask to the input image as no-data: all the pixels that corresponds have a null mask value are flagged as no-data in the output image.

For instance, the following command converts the no-data value of the input image to the default value for DEM (which is -32768):

otbcli_ManageNoData -in input_image.tif
                    -out output_image.tif
                    -mode changevalue
                    -mode.changevalue.newv -32768

The third mode “apply” can be useful if you apply a formula to the entire image. This will likely change the values of pixels flagged as no-data, but the no-data value in the image metadata does not change. If you want to fix all no-data pixels to their original value, you can extract the mask of the original image and apply it on the output image. For instance:

otbcli_ManageNoData -in input_image.tif
                    -out mask.tif
                    -mode buildmask

otbcli_BandMath -il input_image.tif
                -out filtered_image.tif
                -exp "2*im1b1-4"

otbcli_ManageNoData -in filtered_image.tif
                    -out output_image.tif
                    -mode apply
                    -mode.apply.mask mask.tif

You can also use this “apply” mode with an additional parameter “mode.apply.ndval”. This parameter sets the output nodata value of the input mask.

Segmentation

Segmenting objects across a very high resolution scene and with a controlled quality is a difficult task for which no method has reached a sufficient level of performance to be considered as operational.

Even if we leave aside the question of segmentation quality and consider that we have a method performing reasonably well on our data and objects of interest, the task of scaling up segmentation to real very high resolution data is itself challenging. First, we can not load the whole data into memory, and there is a need for on the flow processing which does not cope well with traditional segmentation algorithms. Second, the result of the segmentation process itself is difficult to represent and manipulate efficiently.

The experience of segmenting large remote sensing images is packed into a single Segmentation in OTB Applications .

You can find more information about this application here.

Large-Scale Mean-Shift (LSMS) segmentation

LSMS is a segmentation workflow which allows to perform tile-wise segmentation of very large image with theoretical guarantees of getting identical results to those without tiling.

It has been developed by David Youssefi and Julien Michel during David internship at CNES.

For more a complete description of the LSMS method, please refer to the following publication, J. Michel, D. Youssefi and M. Grizonnet, “Stable Mean-Shift Algorithm and Its Application to the Segmentation of Arbitrarily Large Remote Sensing Images,” in IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 2, pp. 952-964, Feb. 2015. The workflow consists in chaining 3 or 4 dedicated applications and produces a GIS vector file with artifact-free polygons corresponding to the segmented image, as well as mean and variance of the radiometry of each band for each polygon.

Step 1: Mean-Shift Smoothing

The first step of the workflow is to perform Mean-Shift smoothing with the MeanShiftSmoothing application:

otbcli_MeanShiftSmoothing -in input_image.tif
                          -fout filtered_range.tif
                          -foutpos filtered_spatial.tif
                          -ranger 30
                          -spatialr 5
                          -maxiter 10
                          -modesearch 0

Note that the modesearch option should be disabled, and that the foutpos parameter is optional: it can be activated if you want to perform the segmentation based on both spatial and range modes.

This application will smooth large images by streaming them, and deactivating the modesearch will guarantee that the results will not depend on the streaming scheme. Please also note that the maxiter is used to set the margin to ensure these identical results, and as such increasing the maxiter may have an additional impact on processing time.

Step 2: Segmentation

The next step is to produce an initial segmentation based on the smoothed images produced by the MeanShiftSmoothing application. To do so, the LSMSSegmentation will process them by tiles whose dimensions are defined by the tilesizex and tilesizey parameters, and by writing intermediate images to disk, thus keeping the memory consumption very low throughout the process. The segmentation will group together neighboring pixels whose range distance is below the ranger parameter and (optionally) spatial distance is below the spatialr parameter.

otbcli_LSMSSegmentation -in filtered_range.tif
                        -inpos filtered_spatial.tif
                        -out  segmentation.tif uint32
                        -ranger 30
                        -spatialr 5
                        -minsize 0
                        -tilesizex 256
                        -tilesizey 256

Note that the final segmentation image may contains a very large number of segments, and the uint32 image type should therefore be used to ensure that there will be enough labels to index those segments. The minsize parameter will filter segments whose size in pixels is below its value, and their labels will be set to 0 (nodata).

Please note that the output segmented image may look patchy, as if there were tiling artifacts: this is because segments are numbered sequentially with respect to the order in which tiles are processed. You will see after the result of the vectorization step that there are no artifacts in the results.

The LSMSSegmentation application will write as many intermediate files as tiles needed during processing. As such, it may require twice as free disk space as the final size of the final image. The cleanup option (active by default) will clear the intermediate files during the processing as soon as they are not needed anymore. By default, files will be written to the current directory. The tmpdir option allows to specify a different directory for these intermediate files.

Step 3 (optional): Merging small regions

The LSMSSegmentation application allows to filter out small segments. In the output segmented image, those segments will be removed and replaced by the background label (0). Another solution to deal with the small regions is to merge them with the closest big enough adjacent region in terms of radiometry. This is handled by the LSMSSmallRegionsMerging application, which will output a segmented image where small regions have been merged. Again, the uint32 image type is advised for this output image.

otbcli_LSMSSmallRegionsMerging -in filtered_range.tif
                               -inseg segmentation.tif
                               -out segmentation_merged.tif uint32
                               -minsize 10
                               -tilesizex 256
                               -tilesizey 256

The minsize parameter allows to specify the threshold on the size of the regions to be merged. Like the LSMSSegmentation application, this application will process the input images tile-wise to keep resources usage low, with the guarantee of identical results. You can set the tile size using the tilesizex and tilesizey parameters. However unlike the LSMSSegmentation application, it does not require to write any temporary file to disk.

Step 4: Vectorization

The last step of the LSMS workflow consists in the vectorization of the segmented image into a GIS vector file. This vector file will contain one polygon per segment, and each of these polygons will hold additional attributes denoting the label of the original segment, the size of the segment in pixels, and the mean and variance of each band over the segment. The projection of the output GIS vector file will be the same as the projection from the input image (if input image has no projection, so does the output GIS file).

otbcli_LSMSVectorization -in input_image.tif
                         -inseg segmentation_merged.tif
                         -out segmentation_merged.shp
                         -tilesizex 256
                         -tilesizey 256

This application will process the input images tile-wise to keep resources usage low, with the guarantee of identical results. You can set the tile size using the tilesizex and tilesizey parameters. However unlike the LSMSSegmentation application, it does not require to write any temporary file to disk.

All-in-one

The LargeScaleMeanShift application is a composite application that chains all the previous steps:

  • Mean-Shift Smoothing
  • Segmentation
  • Small region merging
  • Vectorization

Most of the settings from the previous applications are also exposed in this composite application. The range and spatial radius used for the segmentation step are half the values used for Mean-Shift smoothing, which are obtained from LargeScaleMeanShift parameters. There are two output modes: vector (default) and raster. When the raster output is chosen, last step (vectorization) is skipped.

otbcli_LargeScaleMeanShift -in input_image.tif
                           -spatialr 5
                           -ranger 30
                           -minsize 10
                           -mode.vector.out segmentation_merged.shp

There is a cleanup option that can be disabled in order to check intermediate outputs of this composite application.

Dempster Shafer based Classifier Fusion

This framework is dedicated to perform cartographic validation starting from the result of a detection (for example a road extraction), enhance the results viability by using a classifier fusion algorithm. Using a set of descriptors, the processing chain validates or invalidates the input geometrical features.

Fuzzy Model (requisite)

The DSFuzzyModelEstimation application performs the fuzzy model estimation (once by use case: descriptor set / Belief support / Plausibility support). It has the following input parameters:

  • -psin a vector data of positive samples enriched according to the “Compute Descriptors” part
  • -nsin a vector data of negative samples enriched according to the “Compute Descriptors” part
  • -belsup a support for the Belief computation
  • -plasup a support for the Plausibility computation
  • -desclist an initialization model (xml file) or a descriptor name list (listing the descriptors to be included in the model)

The application can be used like this:

otbcli_DSFuzzyModelEstimation -psin     PosSamples.shp
                              -nsin     NegSamples.shp
                              -belsup   "ROADSA"
                              -plasup   "NONDVI" "ROADSA" "NOBUIL"
                              -desclist "NONDVI" "ROADSA" "NOBUIL"
                              -out      FuzzyModel.xml

The output file FuzzyModel.xml contains the optimal model to perform information fusion.

First Step: Compute Descriptors

The first step in the classifier fusion based validation is to compute, for each studied polyline, the chosen descriptors. In this context, the ComputePolylineFeatureFromImage application can be used for a large range of descriptors. It has the following inputs:

  • -in an image (of the studied scene) corresponding to the chosen descriptor (NDVI, building Mask…)
  • -vd a vector data containing polyline of interest
  • -expr a formula (“b1 >0.4”, “b1 == 0”) where b1 is the standard name of input image first band
  • -field a field name corresponding to the descriptor codename (NONDVI, ROADSA...)

The output is a vector data containing polylines with a new field containing the descriptor value. In order to add the “NONDVI” descriptor to an input vector data (“inVD.shp”) corresponding to the percentage of pixels along a polyline that verifies the formula “NDVI >0.4”:

otbcli_ComputePolylineFeatureFromImage -in   NDVI.TIF
                                       -vd  inVD.shp
                                       -expr  "b1 > 0.4"
                                       -field "NONDVI"
                                       -out   VD_NONDVI.shp

NDVI.TIF is the NDVI mono band image of the studied scene. This step must be repeated for each chosen descriptor:

otbcli_ComputePolylineFeatureFromImage -in   roadSpectralAngle.TIF
                                       -vd  VD_NONDVI.shp
                                       -expr  "b1 > 0.24"
                                       -field "ROADSA"
                                       -out   VD_NONDVI_ROADSA.shp

otbcli_ComputePolylineFeatureFromImage -in   Buildings.TIF
                                       -vd  VD_NONDVI_ROADSA.shp
                                       -expr  "b1 == 0"
                                       -field "NOBUILDING"
                                       -out   VD_NONDVI_ROADSA_NOBUIL.shp

Both NDVI.TIF and roadSpectralAngle.TIF can be produced using Monteverdi feature extraction capabilities, and Buildings.TIF can be generated using Monteverdi rasterization module. From now on, VD_NONDVI_ROADSA_NOBUIL.shp contains three descriptor fields. It will be used in the following part.

Second Step: Feature Validation

The final application (VectorDataDSValidation ) will validate or invalidate the studied samples using the Dempster-Shafer theory . Its inputs are:

  • -in an enriched vector data “VD_NONDVI_ROADSA_NOBUIL.shp”
  • -belsup a support for the Belief computation
  • -plasup a support for the Plausibility computation
  • -descmod a fuzzy model FuzzyModel.xml

The output is a vector data containing only the validated samples.

otbcli_VectorDataDSValidation -in      extractedRoads_enriched.shp
                              -descmod FuzzyModel.xml
                              -out     validatedSamples.shp