A simple usage example¶
Check this Jupyter notebook for a more complete example.
Suppose we have two images of about the same portion of the sky, and we would like to transform one of them to fit on top of the other one. Suppose we do not have WCS information, but we are confident that we could do it by eye, by matching some obvious asterisms on the two images.
In this particular use case, astroalign can be of great help to automatize the process.
After we load our images into numpy arrays, we simple choose one to be the source image to be transformed and the other to be the target.
astroalign will also accept as input, data objects with data and mask properties, like NDData, CCDData and Numpy masked arrays. For more information, see Dealing with Data Objects with data and mask properties (NDData, CCDData, Numpy masked arrays)
The usage for this simple most common case would be as follows:
>>> import astroalign as aa >>> registered_image, footprint = aa.register(source, target)
registered_image is now a transformed (numpy array) image of
source that will match pixel to pixel to
footprint is a boolean numpy array, True for masked pixels with no information.
Flux may not be conserved after the transformation.
Mask Fill Value¶
If you need to mask the aligned image with a special value over the region where transformation had no pixel information, you can use the footprint mask to do so:
>>> registered_image, footprint = aa.register(source, target) >>> registered_image[footprint] = -99999.99
Or you can pass the value to the fill_value argument:
>>> registered_image, footprint = aa.register(source, target, fill_value=-99999.99)
Both will yield the same result.
Finding the transformation¶
In some cases it may be necessary to inspect first the transformation parameters before applying it,
or we may be interested only in a star to star correspondence between the images.
For those cases, we can use
find_transform will return a scikit-image SimilarityTransform object that encapsulates the matrix transformation,
and the transformation parameters.
It will also return a tuple with two lists of star positions of
source and its corresponding ordered star postions on
>>> transf, (source_list, target_list) = aa.find_transform(source, target)
source and target here can be either numpy arrays of the image pixels, or any iterable (x, y) pair, corresponding to a star position.
The transformation parameters can be found in
and the transformation matrix in
If the transformation is satisfactory, we can apply it to the image with
Continuing our example:
>>> if transf.rotation > MIN_ROT: ... registered_image = aa.apply_transform(transf, source, target)
If you know the star-to-star correspondence¶
estimate_transform from scikit-image is imported into astroalign as a convenience.
If for any reason you know which star corresponds to which other, you can call
Let us suppose we know the correspondence:
- (127.03, 85.98) in source –> (175.13, 111.36) in target
- (23.11, 31.87) in source –> (0.58, 119.04) in target
- (98.84, 142.99) in source –> (181.55, 206.49) in target
- (150.93, 85.02) in source –> (205.60, 91.89) in target
- (137.99, 12.88) in source –> (134.61, 7.94) in target
Then we can estimate the transform:
>>> src = np.array([(127.03, 85.98), (23.11, 31.87), (98.84, 142.99), ... (150.93, 85.02), (137.99, 12.88)]) >>> dst = np.array([(175.13, 111.36), (0.58, 119.04), (181.55, 206.49), ... (205.60, 91.89), (134.61, 7.94)]) >>> tform = aa.estimate_transform('affine', src, dst)
And apply it to an image with
apply_transform or to a set of points with
Applying a transformation to a set of points¶
matrix_transform from scikit-image is imported into astroalign as a convenience.
To apply a known transform to a set of points, we use matrix_transform. Following the example in the previous section:
>>> dst_calc = aa.matrix_transform(src, tform.params)
dst_calc should be a 5 by 2 array similar to the
Dealing with Data Objects with data and mask properties (NDData, CCDData, Numpy masked arrays)¶
In general, for objects with data and mask properties, it is convenient to transform their masks along with the data and to add the footprint onto the mask.
Astroalign provides this functionality with the propagate_mask argument to register and apply_transform.
>>> from astropy.nddata import NDData >>> nd = NDData([[0, 1], [2, 3]], [[True, False], [False, False]])
and we want to apply a clockwise 90 degree rotation:
>>> import numpy as np >>> from skimage.transform import SimilarityTransform >>> transf = SimilarityTransform(rotation=np.pi/2., translation=(1, 0))
Then we can call astroalign as usual, but with the propagate_mask set to True:
>>> aligned_image, footprint = aa.apply_transform(transf, nd, nd, propagate_mask=True)
This will transform nd.data and nd.mask simultaneously and add the footprint mask from the transformation onto nd.mask:
>>> aligned_image array([[2., 0.], [3., 1.]]) >>> footprint array([[False, True], [False, False]])
Creating a new object of the same input type is now easier:
>>> new_nd = NDData(aligned_image, mask=footprint)
The same will apply for CCDData objects and Numpy masked arrays.
Dealing with hot pixels¶
Hot pixels always appear on the same position of the CCD. If your image is dominated by hot pixels, the source detection algorithm may pick those up and output the identity tranformation.
To avoid this, you can use CCDProc’s cosmicray_lacosmic to clean the image before trying registration:
from ccdproc import cosmicray_lacosmic as lacosmic clean_image = lacosmic(myimage)
See Module API for the API specification.