Add parametric nowcasting and calibrated stochastic rainfall generation for v2 by alanseed · Pull Request #604 · pySTEPS/pysteps

alanseed · 2026-11-04T23:39:31Z

Adds parametric ensemble nowcasting from three consecutive rain-rate observations, together with calibrated stochastic rainfall generation and marginal, spectral, and temporal estimators. This ports the contribution in #578 to the v2 main branch and its xarray dataset interface.

The forecast derives grid spacing and cadence from coordinates, accepts supplied motion or estimates Lucas–Kanade motion internally, and returns rain rates with ensemble identifiers, valid times, forecast reference time, and geographic metadata. It preserves the NumPy numerical engine, independent member random streams, temporal-only fallback, and the distinction between missing and dry input. Callbacks remain arrays and support generation without retaining the full ensemble.

The synthetic and Swiss examples use the dataset interface. The Swiss example includes lossless native NetCDF export and a GIF of three QPE observations followed by the first member's six forecasts.

Validation:

Full suite on Python 3.13: 987 passed, 1 skipped. The run used a writable temporary working directory and a test configuration with the packaged FMI GeoTIFF filename/importer settings.
The 253 focused parametric, ingestion, conversion, spatial-processing, and exporter tests also passed.
Saved reference forecast arrays and existing numerical tolerances are unchanged.
Three saved paper fields passed estimator comparison against the source implementation.
Three paper histories, each with two seeds, two members, and three leads, passed source comparisons. Fixed-parameter evolution was identical; maximum coupled forecast difference was below 0.000085 mm/h.
The actual Swiss run produced 20 members by 6 leads on a 150 by 150 grid at 2 km. NetCDF reopening preserved values, coordinates, timestamps, and attributes exactly.
The nine GIF frames were checked against their source fields, and the forecast was identical to the prior validated export.

Scope: the first interface supports complete regular grids with square pixels and integer lead counts. These checks establish implementation and serialization behaviour; they do not establish Swiss calibration or forecast skill. CF support for additional projections, including UTM and Swiss somerc, remains separate. Existing projection text and available latitude/longitude coordinates are preserved.

Cross-platform and Python 3.11 validation will run in upstream CI.

RubenImhoff

Hi @alanseed, super great work! Still working on fully understanding all the steps that take place in this parametric approach (also to understand how to use and adapt this in the future). I have not given it a full review yet, but added some notes and suggestions based on my own quick reading through it to understand the method.
It may be good if @ladc, @dnerini and perhaps @pulkkins also have a look at it. :)

RubenImhoff · 2026-11-05T10:30:55Z

I am very happy with all the documentation attached! This kind of documentation is not present / not the same for the other pysteps modules. How should we proceed? Does it make sense to merge this documentation into one or two docs in a sub folder under either pysteps_reference or user_guide? @dnerini and @ladc, what is your take on this?

I see that there is quite some documentation already - so then the question remains what we want to do with the documentation under proposals.

RubenImhoff · 2026-11-05T10:31:38Z

@@ -7,5 +7,6 @@ Implementation of deterministic and ensemble nowcasting methods.

 .. automodule:: pysteps.noise.interface
 .. automodule:: pysteps.noise.fftgenerators


Suggested change

.. automodule:: pysteps.noise.fftgenerators

RubenImhoff · 2026-11-05T10:31:58Z

@@ -7,5 +7,6 @@ Implementation of deterministic and ensemble nowcasting methods.

 .. automodule:: pysteps.noise.interface


Suggested change

.. automodule:: pysteps.noise.interface

.. automodule:: pysteps.noise.fftgenerators

.. automodule:: pysteps.noise.interface

Good reason to directly make this alphabetical. :)

RubenImhoff · 2026-11-05T10:32:11Z


 .. automodule:: pysteps.noise.interface
 .. automodule:: pysteps.noise.fftgenerators
+.. automodule:: pysteps.noise.parametric


Suggested change

.. automodule:: pysteps.noise.parametric

RubenImhoff · 2026-11-05T10:32:19Z

 .. automodule:: pysteps.noise.fftgenerators
+.. automodule:: pysteps.noise.parametric
 .. automodule:: pysteps.noise.motion
 .. automodule:: pysteps.noise.utils


Suggested change

.. automodule:: pysteps.noise.utils

.. automodule:: pysteps.noise.parametric

.. automodule:: pysteps.noise.utils

RubenImhoff · 2026-11-05T14:20:49Z

Looks good! It might be good to add some docstrings to different parts in the functions to ensure all steps are clear for future developers and maintainers.

This also holds for the code in the functions under ./nowcasts

RubenImhoff · 2026-11-05T14:25:21Z

-The time step of the output is taken from the inputs.
+The ``parametric`` method requires exactly three complete, untransformed
+rain-rate observations in mm/h on a regular grid with square pixels. Pixel
+size and cadence are derived from coordinates; the time stepsize is in seconds.


Suggested change

size and cadence are derived from coordinates; the time stepsize is in seconds.

size and time step are derived from coordinates; the time step is in seconds.

RubenImhoff · 2026-11-05T14:25:57Z

+or (ens_number, time, y, x) for ensemble methods. Method-specific forecast
+functions document their input history and optional settings.

-The output depends on the type of the method.


Good to still replace this with what the output is (an xarray DataSet)

RubenImhoff · 2026-11-05T14:28:06Z

+    return selected, audit
+
+
+def _advance(


Good to add some docstrings to the function.

RubenImhoff · 2026-11-05T14:30:09Z

+from ._cascade_advection import advect_field, prepare_advection_coordinates
+
+# Evaluated nowcast_v2 settings, not the later plotting-only timescale fit.
+_ESTIMATION_DEFAULTS = dict(


Are these all default settings that the user can change? And if so, does this happen in the main calling function? If that is the case, I would prefer having the defaults set in the function (as default) and then these dictionaries can be filled with the variable value (either the default or the user-assigned value).

RubenImhoff

Hi @alanseed, super great work! Still working on fully understanding all the steps that take place in this parametric approach (also to understand how to use and adapt this in the future). I have not given it a full review yet, but added some notes and suggestions based on my own quick reading through it to understand the method.
It may be good if @ladc, @dnerini and perhaps @pulkkins also have a look at it. :)

alanseed added 9 commits November 5, 2026 08:55

parametric engine done

2e6b2b7

examples added

523e702

Update estimator guide for implemented parametric engine

4dd2c03

Fix parametric documentation headings

e3d1a02

Allow platform variation in fitted marginal statistics

52bf915

parametric temporal fallback added

c1c4da5

tests for parametric temporal fallback added

de22a03

black reformatting completed

f9e9bfe

pysteps_param xarray model

51b8945

alanseed requested review from RubenImhoff and ladc November 4, 2026 23:39

RubenImhoff reviewed Nov 5, 2026

View reviewed changes

		@@ -7,5 +7,6 @@ Implementation of deterministic and ensemble nowcasting methods.

		.. automodule:: pysteps.noise.interface
		.. automodule:: pysteps.noise.fftgenerators

	.. automodule:: pysteps.noise.utils
	.. automodule:: pysteps.noise.parametric
	.. automodule:: pysteps.noise.utils

	size and cadence are derived from coordinates; the time stepsize is in seconds.
	size and time step are derived from coordinates; the time step is in seconds.

Conversation

alanseed commented Nov 4, 2026

Uh oh!

RubenImhoff left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

RubenImhoff left a comment

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants