Export to Numpy¶

pyotb objects can be exported to numpy array.

import pyotb
import numpy as np

calibrated = pyotb.OpticalCalibration('image.tif', level='toa')  # this is a pyotb object
arr = np.asarray(calibrated)  # same as calibrated.to_numpy()

Interaction with Numpy¶

pyotb objects can be transparently used in numpy functions.

For example:

import pyotb
import numpy as np

inp = pyotb.Input('image.tif')  # this is a pyotb object

# Creating a numpy array of noise
white_noise = np.random.normal(0, 50, size=inp.shape)  # this is a numpy object

# Adding the noise to the image
noisy_image = inp + white_noise  # magic: this is a pyotb object that has the same georeference as input. 
                                 # `np.add(inp, white_noise)` would have worked the same
noisy_image.write('image_plus_noise.tif')

Limitations :

• The whole image is loaded into memory
• The georeference can not be modified. Thus, numpy operations can not change the image or pixel size

Export to rasterio¶

pyotb objects can also be exported in a format that is usable by rasterio.

For example:

import pyotb
import rasterio
from scipy import ndimage

# Pansharpening + NDVI + creating bare soils mask
pxs = pyotb.BundleToPerfectSensor(inp='panchromatic.tif', inxs='multispectral.tif')
ndvi = pyotb.RadiometricIndices({'in': pxs, 'channels.red': 3, 'channels.nir': 4, 'list': 'Vegetation:NDVI'})
bare_soil_mask = (ndvi < 0.3)

# Exporting the result as array & profile usable by rasterio
mask_array, profile = bare_soil_mask.to_rasterio()

# Doing something in Python that is not possible with OTB, e.g. gathering the contiguous groups of pixels
# with an integer index
labeled_mask_array, nb_groups = ndimage.label(mask_array)

# Writing the result to disk
with rasterio.open('labeled_bare_soil.tif', 'w', **profile) as f:
    f.write(labeled_mask_array)

This way of exporting pyotb objects is more flexible that exporting to numpy, as the user gets the profile dictionary. If the georeference or pixel size is modified, the user can update the profile accordingly.

Interaction with Tensorflow¶

We saw that numpy operations had some limitations. To bypass those limitations, it is possible to use some Tensorflow operations on pyotb objects.

You need a working installation of OTBTF >=3.0 for this and then the code is like this:

import pyotb

def scalar_product(x1, x2):
    """This is a function composed of tensorflow operations."""
    import tensorflow as tf
    return tf.reduce_sum(tf.multiply(x1, x2), axis=-1)

# Compute the scalar product
res = pyotb.run_tf_function(scalar_product)('image1.tif', 'image2.tif')  # magic: this is a pyotb object
res.write('scalar_product.tif')

For some easy syntax, one can use pyotb.run_tf_function as a function decorator, such as:

import pyotb

@pyotb.run_tf_function  # The decorator enables the use of pyotb objects as inputs/output of the function
def scalar_product(x1, x2):
    import tensorflow as tf
    return tf.reduce_sum(tf.multiply(x1, x2), axis=-1)

res = scalar_product('image1.tif', 'image2.tif')  # magic: this is a pyotb object

Advantages :

• The process supports streaming, hence the whole image is not loaded into memory
• Can be integrated in OTB pipelines

Limitations :

• It is not possible to use the tensorflow python API inside a script where OTBTF is used because of compilation issues between Tensorflow and OTBTF, i.e. import tensorflow doesn't work in a script where OTBTF apps have been initialized