Python API

The applications can also be accessed from Python, through a module named otbApplication. However, there are technical requirements to use it. If you use OTB through standalone packages, you should use the supplied environment script otbenv to properly setup variables such as PYTHONPATH and OTB_APPLICATION_PATH (on Unix systems, don’t forget to source the script). In other cases, you should set these variables depending on your configuration.

On Unix systems, it is typically available in the /usr/lib/otb/python directory. Depending on how you installed OTB, you may need to configure the environment variable PYTHONPATH to include this directory so that the module becomes available from Python.

On Windows, you can install the otb-python package, and the module will be automatically available from an OSGeo4W shell.

As for the command line and GUI launchers, the path to the application modules needs to be properly set with the OTB_APPLICATION_PATH environment variable. The standard location on Unix systems is /usr/lib/otb/applications. On Windows, the applications are available in the otb-bin OSGeo4W package, and the environment is configured automatically so OTB_APPLICATION_PATH doesn’t need to be modified OTB_APPLICATION_PATH.

Once your environment is set, you can use OTB applications from Python, just like this small example:

#  Example on the use of the Smoothing application

# The python module providing access to OTB applications is otbApplication
import otbApplication as otb

# Let's create the application with codename "Smoothing"
app = otb.Registry.CreateApplication("Smoothing")

# We set its parameters
app.SetParameterString("in", "my_input_image.tif")
app.SetParameterString("type", "mean")
app.SetParameterString("out", "my_output_image.tif")

# This will execute the application and save the output file
app.ExecuteAndWriteOutput()

Basics

In the otbApplication module, two main classes can be manipulated :

  • Registry, which provides access to the list of available applications, and can create applications.
  • Application, the base class for all applications. This allows one to interact with an application instance created by the Registry.

Here is one example of how to use Python to run the Smoothing application, changing the algorithm at each iteration.

#  Example on the use of the Smoothing application
#

# We will use sys.argv to retrieve arguments from the command line.
# Here, the script will accept an image file as first argument,
# and the basename of the output files, without extension.
from sys import argv

# The python module providing access to OTB applications is otbApplication
import otbApplication

# otbApplication.Registry can tell you what application are available
print('Available applications: ')
print (str( otbApplication.Registry.GetAvailableApplications()))

# Let's create the application  "Smoothing"
app = otbApplication.Registry.CreateApplication("Smoothing")

# We print the keys of all its parameters
print (app.GetParametersKeys())

# First, we set the input image filename
app.SetParameterString("in", argv[1])

# The smoothing algorithm can be set with the "type" parameter key
# and can take 3 values: 'mean', 'gaussian', 'anidif'
for type in ['mean', 'gaussian', 'anidif']:

  print('Running with ' + type + ' smoothing type')

  # Now we configure the smoothing algorithm
  app.SetParameterString("type", type)

  # Set the output filename, using the algorithm type to differentiate the outputs
  app.SetParameterString("out", argv[2] + type + ".tif")

  # This will execute the application and save the output to argv[2]
  app.ExecuteAndWriteOutput()

If you want to handle the parameters from a Python dictionary, you can use the functions SetParameters() and GetParameters().

params = {"in":"myInput.tif", "type.mean.radius":4}
app.SetParameters(params)
params2 = app.GetParameters()

Numpy array processing

Input and output images to any OTB application in the form of NumPy array is now possible in OTB Python wrapping. The Python wrapping only exposes OTB Application engine module (called ApplicationEngine) which allows one to access existing C++ applications. Due to blissful nature of ApplicationEngine’s loading mechanism no specific wrapping is required for each application.

NumPy extension to Python wrapping allows data exchange to application as an array rather than a disk file. Of course, it is possible to load an image from file and then convert it to NumPy array or just provide a file as explained in the previous section via Application.SetParameterString(…).

The bridge between NumPy and OTB makes it easy to plug OTB into any image processing chain via Python code that uses GIS/Image processing tools such as GDAL, GRASS GIS, OSSIM that can deal with NumPy.

Below code reads an input image using Python Pillow library (fork of PIL) and convert it to NumPy array. The NumPy array is used as an input to the application via SetImageFromNumpyArray(…) method. The application used in this example is ExtractROI. After extracting a small area the output image is taken as NumPy array with GetImageFromNumpyArray(…) method thus avoid writing output to a temporary file.

import sys
import os
import numpy as np
import otbApplication
from PIL import Image as PILImage

pilimage = PILImage.open('poupees.jpg')
npimage = np.asarray(pilimage)
inshow(pilimage)

ExtractROI = otbApplication.Registry.CreateApplication('ExtractROI')
ExtractROI.SetImageFromNumpyArray('in', npimage)
ExtractROI.SetParameterInt('startx', 140)
ExtractROI.SetParameterInt('starty', 120)
ExtractROI.SetParameterInt('sizex', 150)
ExtractROI.SetParameterInt('sizey', 150)
ExtractROI.Execute()

ExtractOutput = ExtractROI.GetImageAsNumpyArray('out')
output_pil_image = PILImage.fromarray(np.uint8(ExtractOutput))
imshow(output_pil_image)

Mix OTB with other libraries

It’s now possible to mix Python libraries (rasterio to open images, scikit-image, scikit-learn for processings) with OTB. In order to build efficient code, here are some tips illustrated by a small example.

  • OTB expects “(height, width, channels)” shaped arrays while rasterio and other libs usually use “(channels, height, width)” arrays. The Numpy np.transpose(x,y,z) helps you transpose axis
  • For single-band images, use otbapp.SetImageFromNumpyArray(..) method and otbapp.SetVectorImageFromNumpyArray(..) otherwise.
  • *OTB returns a reference on the numpy array output : depending on your use-case, you may copy the array (numpy.copy())
  • OTB expects C contiguous arrays : sometimes it’s not the case, for example if multiple process use a shared_memory array. In that specific case, you may use Numpy.ascontiguousarray method to make it work properly
ds = rio.open(image)
np_image_raw = ds.read()
print(">> shape rio.open().read() "+str(np_image_raw.shape))

nbchannels = np_image_raw.shape[0]
height = np_image_raw.shape[1]
width = np_image_raw.shape[2]
# use np.transpose to switch axis : OTB expects [height, width, nb channels] images
np_image=np_image_raw.transpose(1, 2, 0)
[ ... ]
app_rindices = otb.Registry.CreateApplication("RadiometricIndices")
app_rindices.SetVectorImageFromNumpyArray('in', np_image)
# OTB expects C contiguous arrays and in certain conditions (ex : shared_memory arrays),
# the ndarray should be transform as following
app_rindices.SetVectorImageFromNumpyArray('in', np.ascontiguousarray(np_image))
[ ... set parameters ..]
app_rindices.Execute() # we don't write output
# Get output result
res_indices = app_rindices.GetVectorImageAsNumpyArray("out")
# to write it with rasterio or pass it to another library, you may switch back axis
res_indices = res_indices.transpose(2,0,1)
# Be aware that as soon as OTB application is deleted from memory, the ndarray is deallocated...
# You should consider making a copy of the array
res_indices = app_rindices.GetVectorImageAsNumpyArray("out").copy()

In-memory connection

Applications are often used as part of larger processing workflows. Chaining applications currently requires to write/read back images between applications, resulting in heavy I/O operations and a significant amount of time dedicated to writing temporary files.

Since OTB 5.8, it is possible to connect an output image parameter from one application to the input image parameter of the next parameter. This results in the wiring of the internal ITK/OTB pipelines together, permitting image streaming between the applications. Consequently, this removes the need of writing temporary images and improves performance. Only the last application of the processing chain is responsible for writing the final result images.

In-memory connection between applications is available both at the C++ API level and using the Python bindings.

Here is a Python code sample which connects several applications together:

import otbApplication as otb

app1 = otb.Registry.CreateApplication("Smoothing")
app2 = otb.Registry.CreateApplication("Smoothing")
app3 = otb.Registry.CreateApplication("Smoothing")
app4 = otb.Registry.CreateApplication("ConcatenateImages")

app1.IN = argv[1]
app1.Execute()

# Connection between app1.out and app2.in
app2.SetParameterInputImage("in",app1.GetParameterOutputImage("out"))

# Execute call is mandatory to wire the pipeline and expose the
# application output. It does not write image
app2.Execute()

app3.IN = argv[1]

# Execute call is mandatory to wire the pipeline and expose the
# application output. It does not write image
app3.Execute()

# Connection between app2.out, app3.out and app4.il using images list
app4.AddImageToParameterInputImageList("il",app2.GetParameterOutputImage("out"));
app4.AddImageToParameterInputImageList("il",app3.GetParameterOutputImage("out"));

app4.OUT = argv[2]

# Call to ExecuteAndWriteOutput() both wires the pipeline and
# actually writes the output, only necessary for last application of
# the chain.
app4.ExecuteAndWriteOutput()

Note: Streaming will only work properly if the application internal implementation does not break it, for instance by using an internal writer to write intermediate data. In this case, execution should still be correct, but some intermediate data will be read or written.

Mixed in-memory / on-disk connection

As an extension to the connection of OTB Applications (described in previous section), a mixed mode is also available to easily switch between:

  • in-memory: if you want to avoid unnecessary I/O between applications
  • on-disk: if you want intermediate outputs on disk

This mixed mode is based on the Application::ConnectImage() function. This is how you use it:

# for in-memory mode
app1 = otb.Registry.CreateApplication("Smoothing")
app2 = otb.Registry.CreateApplication("Smoothing")

app1.IN = input_image

app2.ConnectImage("in", app1, "out")
app2.OUT = output_image
app2.ExecuteAndWriteOutput()

Comparing with the standard in-memory connection, you can notice that:

  • the syntax to do the connection is simpler
  • you don’t need to call Execute() on upstream applications anymore, this is done recursively by calling ExecuteAndWriteOutput() on the latest application

The on-disk version of this code is very similar:

# for on-disk mode
app1 = otb.Registry.CreateApplication("Smoothing")
app2 = otb.Registry.CreateApplication("Smoothing")

app1.IN = input_image
app1.OUT = temp_image

app2.ConnectImage("in", app1, "out")
app2.OUT = output_image
app2.PropagateConnectMode(False)
app2.ExecuteAndWriteOutput()

The function PropagateConnectMode() is applied recursively in the upstream applications. It allows to change between the 2 modes:

  • True : means in-memory mode (this is the default)
  • False : means on-disk mode

When you want to use the on-disk mode, you have to specify the path to the intermediate image in the output image parameter of the upstream application (app1.OUT in the previous example). If this path is empty, the in-memory mode is used as fallback.

This feature also works for InputImageList. Calling the function ConnectImage("il", app1, "out"), with il being an input image list, will append a new image to the list, and connect it to output parameter out.

Load and save parameters to XML

As with a the command line interface you can save application parameters to an xml file:

# Save application parameters to XML
app = otb.Registry.CreateApplication('BandMath')
app.SetParameterStringList("il", ["image1.tif", "image2.tif"], True)
app.SetParameterString("out", out, True)
app.SetParameterString("exp", "cos(im1b1)+im2b1*im1b1", True)
app.SaveParametersToXML("parameters.xml")

And load them later for execution:

# Load application parameters from XML
app = otb.Registry.CreateApplication("BandMath")
app.LoadParametersFromXML("parameters.xml")
app.ExecuteAndWriteOutput()

Interactions with OTB pipeline

The application framework has been extended in order to provide ways to interact with the pipelines inside each application. It applies only to applications that use input or output images. Let’s check which functions are available in the Application class. There are lots of getter functions:

Function name return value
GetImageOrigin(...) origin of the image (physical position of the first pixel center)
GetImageSpacing(...) signed spacing of the image
GetImageSize(...) size of the LargestPossibleRegion
GetImageNbBands(...) number of components per pixel
GetImageProjection(...) Projection WKT string
GetMetadataDictionary(...) the entire MetaDataDictionary
GetImageRequestedRegion(...) requested region
GetImageBasePixelType(...) pixel type of the underlying Image/VectorImage.
GetImageMetadata(...) the ImateMetadata object

All these getters functions use the following arguments:

  • key: a string containing the key of the image parameter
  • idx: an optional index (default is 0) that can be used to access ImageList parameters transparently

There is also a function to send orders to the pipeline:

PropagateRequestedRegion(key, region, idx=0): sets a given RequestedRegion on the image and propagate it, returns the memory print estimation. This function can be used to measure the requested portion of input images necessary to produce an extract of the full output.

Note: a requested region (like other regions in the C++ API of otb::Image) consists of an image index and a size, which defines a rectangular extract of the full image.

This set of functions has been used to enhance the bridge between OTB images and Numpy arrays. There are now import and export functions available in Python that preserve the metadata of the image during conversions to Numpy arrays:

  • ExportImage(self, key): exports an output image parameter into a Python dictionary.
  • ImportImage(self, key, dict, index=0): imports the image from a Python dictionary into an image parameter (as a monoband image).
  • ImportVectorImage(self, key, dict, index=0): imports the image from a Python dictionary into an image parameter (as a multiband image).

The Python dictionary used has the following entries:

  • 'array': the Numpy array containing the pixel buffer
  • 'origin': origin of the image
  • 'spacing': signed spacing of the image
  • 'size': full size of the image
  • 'region': region of the image present in the buffer
  • 'metadata': metadata dictionary (contains projection,…)

The metadata dictionary contains various type of data. Here are the available keys of the dictionary, ordered by type:

  • double:

    AbsoluteCalibrationConstant, AverageSceneHeight, BlueDisplayChannel, CalFactor, CalScale, CenterIncidenceAngle, DataType, GreenDisplayChannel, LineSpacing, NoData, NumberOfColumns, NumberOfLines, OrbitNumber, PRF, PhysicalBias, PhysicalGain, PixelSpacing, RSF, RadarFrequency, RangeTimeFirstPixel, RangeTimeLastPixel, RedDisplayChannel, RescalingFactor, SatAzimuth, SatElevation, SolarIrradiance, SpectralMax, SpectralMin, SpectralStep, SunAzimuth, SunElevation, TileHintX, TileHintY

  • string:

    AREA_OR_POINT, BandName, BeamMode, BeamSwath, EnhancedBandName, GeometricLevel, Instrument, InstrumentIndex, LAYER_TYPE, METADATATYPE, Mission, Mode, OTB_VERSION, OrbitDirection, Polarization, ProductType, RadiometricLevel, SensorID, Swath

  • LUT 1D:

    SpectralSensitivity

  • time object:

    AcquisitionDate, AcquisitionStartTime, AcquisitionStopTime, ProductionDate

This dictionary also contains metadata related to projection and sensor model. The corresponding keys are not accessible at the moment. But the dictionary offers a few extra methods:

  • GetProjectedGeometry() returns a string representing the projection. It can be a WKN, an EPSG or a PROJ string.
  • GetProjectionWKT() returns a string representing the projection as a WKT.
  • GetProjectionProj() returns a string representing the projection as a PROJ string.

Now some basic Q&A about this interface:

  • What portion of the image is exported to Numpy array?

    By default, the whole image is exported. If you had a non-empty requested region (the result of calling PropagateRequestedRegion()), then this region is exported.

  • What is the difference between ImportImage and ImportVectorImage?

    The first one is here for Applications that expect a monoband otb::Image. In most cases, you will use the second one: ImportVectorImage.

  • What kind of objects are there in this dictionary export?

    The array is a numpy.ndarray. The other fields are wrapped objects from the OTB library but you can interact with them in a Python way: they support len() and str() operator, as well as bracket operator []. Some of them also have a keys() function just like dictionaries.

This interface allows you to export OTB images (or extracts) to Numpy array, process them by other means, and re-import them with preserved metadata. Please note that this is different from an in-memory connection.

Here is a small example of what can be done:

import otbApplication as otb
from sys import argv

# Create a smoothing application
app = otb.Registry.CreateApplication("Smoothing")
app.SetParameterString("in",argv[1])

# only call Execute() to setup the pipeline, not ExecuteAndWriteOutput() which would
# run it and write the output image
app.Execute()

# Setup a special requested region
myRegion = otb.itkRegion()
myRegion['size'][0] = 20
myRegion['size'][1] = 25
myRegion['index'][0] = 10
myRegion['index'][1] = 10
ram = app.PropagateRequestedRegion("out",myRegion)

# Check the requested region on the input image
print(app.GetImageRequestedRegion("in"))

# Create a ReadImageInfo application
app2 = otb.Registry.CreateApplication("ReadImageInfo")

# export "out" from Smoothing and import it as "in" in ReadImageInfo
ex = app.ExportImage("out")
app2.ImportVectorImage("in", ex)
app2.Execute()

# Check the result of ReadImageInfo
someKeys = ['sizex', 'sizey', 'spacingx', 'spacingy', 'sensor', 'projectionref']
for key in someKeys:
  print(key + ' : ' + str(app2.GetParameterValue(key)))

# Only a portion of "out" was exported but ReadImageInfo is still able to detect the
# correct full size of the image

Corner cases

There are a few corner cases to be aware of when using Python wrappers. They are often limitations, that one day may be solved in future versions. If it happens, this documentation will report the OTB version that fixes the issue.

Multithreading problem when calling ZonalStatistics application

There is a known multithreading problem that occurs only in python when you call, with python >= 3.8, where you can get the error “free() : invalid next size”. It is due to a python GIL problem, that sometimes invalidates a part of the memory before a thread tries to access it. For more information, refer to the issue 2334 on the forge

A workaround for this problem is to use the ZonalStatistics application in C++ / via the command line or graphical tool.

Calling UpdateParameters()

UpdateParameters() is available to the Python API. But in normal use, it does not need to be called manually. From OTB 7.0.0 and later, it is called automatically after each call to SetParameter*() methods. With previous versions of OTB you may need to call it after setting a parameter.

No metadata in NumPy arrays

With the NumPy module, it is possible to convert images between OTB and NumPy arrays. For instance, when converting from OTB to NumPy array:

  • An Update() of the underlying otb::VectorImage is requested. Be aware that the full image is generated.
  • The pixel buffer is copied into a numpy.array

As you can see, there is no export of the metadata, such as origin, spacing, geographic projection. It means that if you want to re-import a NumPy array back into OTB, the image won’t have any of these metadata. This can pose problems for applications that relate to geometry, projections, and also calibration.

Future developments will probably offer a more adapted structure to import and export images between OTB and the Python world.