Make the coregistration approaches more modular.

[erikmannerfelt]

There is no consensus on the optimal coregistration pipeline, so adding modularity would help explore this.

A structure suggestion

It could be done in scikit-learn's fashion for regression models:

from sklearn.linear_model import LinearRegressor, RANSACRegressor
import numpy as np

X = np.linspace(0, 1, num=50).reshape(-1, 1)
y = np.linspace(0, 2, num=50)

for regressor in LinearRegressor, RANSACRegressor:
    model = regressor()
    model.fit(X, y)
    print(model.score(X, y))

The point of the above example is to show how interchangeable the different regression models are.

The function sklearn.pipeline.make_pipeline() allows for custom regression pipelines to be created. If for example the samples should be normalized before, this could easily be done:

from sklearn.linear_model import LinearRegressor
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline

model = make_pipeline([StandardScaler(), LinearRegressor(fit_intercept=True)])

model.fit(X,y)
# etc...

A similar approach could be done for coreg.py:

from xdem import coreg

reference_dem = ....
dem_to_be_aligned = ....
other_dem = ....

model = coreg.AmauryCoreg()
model.fit(reference_dem, dem_to_be_aligned)

print(model.nmad)

model.transform(other_dem)

... or something similar.
If a custom pipeline is sought:

model = coreg.make_pipeline([coreg.BiasCorr(), coreg.OutlierRemoval(), coreg.ICPCoreg()])

model.fit(reference_dem, dem_to_be_aligned)

Basically, all coreg approaches should be subclasses of a BaseCoreg class or similar, and all should have the methods .fit() and .transform() (or similar terminology). The make_pipeline() could be adapted from the equivalent in scikit-learn.

[adehecq]

I like it!

[rhugonnet]

Yes, this looks really great!! I'd be curious of what @fmaussion thinks about this. How is the modular approach of OGGM conceived? I'm not familiar with it myself.

[adehecq]

Assuming we have the worse DEM possible, with gaps, noise and with a very large transformation, it would be good to have the following steps:

• remove a vertical shift, or possibly a ramp
• calculate residuals to a reference elevation and remove large outliers
• calculate a rigid (or possibly non rigid) transformation
• refine with Nuth & Kaab which is always more accurate on subpixel shifts