dl: Deep learning operators: Noise2Fringe (n2f, n2fs3d) and Noise2Fringe Transformer (n2ft) filtering

moraine.api.dl (source). Every name is also moraine.<name> (import moraine as mr).

download_dl_model

download_dl_model(models: str | list = None, path: str = None)

Parameters:

Name	Type	Description	Default
models	str | list	deep learning models to be downloaded, all available models by default	None
path	str	directory to save these models, inside installed Moraine package by default	None

n2f

n2f(intf: ndarray, chunks: tuple = None, depths: tuple = (0, 0), model: str = None)

Noise2Fringe filtering of an interferogram; on the GPU (cupy input) the network runs in half precision.

Parameters:

Name	Type	Description	Default
intf	ndarray	interferogram, 2d np.complex64 or cp.complex64	required
chunks	tuple	chunksize, intf.shape by default	None
depths	tuple	width of the boundary	(0, 0)
model	str	path to the model weights (.pth), use the model comes with this package by default	None

n2fs3d

n2fs3d(adi: ndarray, intf: ndarray, chunks: tuple = None, depths: tuple = (0, 0), model: str = None)

Parameters:

Name	Type	Description	Default
adi	ndarray	amplitude dispersion index, 2d np.float32 or cp.float32	required
intf	ndarray	interferogram, 2d np.complex64 or cp.complex64	required
chunks	tuple	chunksize, intf.shape by default	None
depths	tuple	width of the boundary	(0, 0)
model	str	path to the model weights (.pth), use the model comes with this package by default	None

n2ft

n2ft(x: ndarray, y: ndarray, intf: ndarray, chunks: int = None, k: int = 128, model: str = None, cuda: bool = False, compile: bool = None)

Parameters:

Name	Type	Description	Default
x	ndarray	x coordinate, e.g., longitude, shape (n,) np.floating	required
y	ndarray	y coordinate, e.g., latitude, shape (n,) np.floating	required
intf	ndarray	interferogram, shape(n,) or shape(n,m) np.complex64	required
chunks	int	chunksize, intf.shape[0] by default	None
k	int	number of nearest neighbours of every point of a chunk that are filtered with it (halo)	128
model	str	path to the model weights (.pth), use the model comes with this package by default	None
cuda	bool	use gpu for inference	False
compile	bool	compile the model with torch.compile: the model then runs about 3 times faster on a GPU, but the compilation takes 15-40 s once per process (torch caches its result on disk). Default: when points times interferograms is at least 1e8	None

Examples

Run when this manual is built; the output follows each snippet.

The filters need PyTorch and the trained models (moraine.download_dl_model()), so this example is not run when the manual is built.

import numpy as np
import moraine as mr

mr.download_dl_model()                                   # once; the .pth files go to moraine/dl_model/

# rasters: Noise2Fringe on an interferogram stack (nlines, width, n_pairs), chunked with halos (depths)
intf_filtered = mr.n2f(intf, chunks=(256, 256), depths=(32, 32))

# with the amplitude dispersion index as a second input
intf_filtered = mr.n2fs3d(adi, intf, chunks=(256, 256), depths=(32, 32))

# point clouds: the Noise2Fringe Transformer on (n_points, n_pairs) with the coordinates of the points
intf_filtered = mr.n2ft(lon, lat, intf, chunks=20000, k=128, cuda=True)

The outputs have unit amplitude; the temporal coherence between the filtered and the raw interferograms (temp_coh) is the quality of a pixel or point.