Extending Isofit to handle multisurface classifications by evan-greenbrg · Pull Request #617 · isofit/isofit

evan-greenbrg · 2024-12-20T01:34:18Z

This is includes changes to the code base to handle an external classification file that is mapped to specific configurations of surface models and executions varying per-pixel or per-superpixel.

The goals of the changes:

Utilize external or internal classifications of surface configurations via apply_oe, directly in a manually configured config, and able to run both in per-pixel mode and in superpixel modes.
Remain efficient both in terms of computation time and memory requirements.

Things to highlight:

New backwards compatible "Surface" section of the isofit config. Based on the input or generated classification file, the surface config is populated.

"surface": {
    "Surfaces": {
        "glint_model_surface": {
            "surface_category": "glint_model_surface",
            "surface_file": "/home/bgreenbe/store/Projects/Biosscape/prism/study_area/prm20231110t071521/Multisurface/027/data/surface_glint_model_surface.mat",
            "surface_int": 0
         },
         "multicomponent_surface": {
            "surface_category": "multicomponent_surface",
            "surface_file": "/home/bgreenbe/store/Projects/Biosscape/prism/study_area/prm20231110t071521/Multisurface/027/data/surface_multicomponent_surface.mat",
                    "surface_int": 1
                }
            },
            "base_surface_class_file": "/home/bgreenbe/store/Projects/Biosscape/prism/study_area/prm20231110t071521/Multisurface/027/input/prm20231110t071521_surface_class",
            "multi_surface_flag": true,
            "surface_class_file": "/home/bgreenbe/store/Projects/Biosscape/prism/study_area/prm20231110t071521/Multisurface/027/input/prm20231110t071521_subs_surface_class"
        }
}

The "Surfaces" key is only used when running multiple surfaces.

Now surface specific surface.mat files are possible, and encouraged. When generating from .json, the surface model will use a "surface type" key to generate shared surface groupings into respective .mat files.

The cleanest implementation was to first make this a required key for multi-surface runs from .json. This could be made optional. Passed .mat files will use the same surface prior file across types.

      {
        "input_spectrum_files":
          [
            "/home/bgreenbe/store/Data/SurfaceLibraries/filtered_other"
          ],
        "n_components": 1,
        "windows": [
          {"interval":[300,740], "regularizer":10, "correlation":"decorrelated"},
          {"interval":[740,1250], "regularizer":1e-6, "correlation":"EM", "name":"shallow-water"},
          {"interval":[1250,1325], "regularizer":1e-8, "correlation":"EM", "name": "osf"},
          {"interval":[1325,1960], "regularizer":10, "correlation": "decorrelated" },
          {"interval":[1960,2070], "regularizer":1e-6, "correlation":"EM","name": "co2" },
          {"interval":[2070,2300], "regularizer":10, "correlation":"decorrelated" },
          {"interval":[2300,2500], "regularizer":1e-3, "correlation":"EM",  "isolated": 1,"name": "noise" }
        ],
        "surface_type": "multicomponent_surface"
      },

Isofit is now run sequentially across surface types.

surface_index = index_spectra_by_surface(input_config, index_pairs)
for surface_class_str, class_idx_pairs in surface_index.items():
    logging.info(f"Running surfaces: {surface_class_str}")
    ...
    # If multisurface, update config to reflect surface.
    # Otherwise, returns itself
    config = update_config_for_surface(
        deepcopy(input_config), surface_class_str
    )
    self.fm = fm = ForwardModel(config)
    ...

On the spectrum write, the statevector is matched against a full statevector by name (not index). A fill value populates elements that are present in the full statevector, but not a pixel statevector.
Analytical line processing is handled similarly in sequence across present surface types

A couple of other changes to note:

Refactored the analytical line. I simplified the file initialization, which reduced clutter updating output metadata. Would like a second pair of eyes on this to make sure the read/write is similarly memory intensive as dev branch.
Tests for multisurface runs (Would need to hook up to github actions)
Built in classifier based on the surface.json file. So I didn't have to build an external classification pipeline I incorporated a classifier based on the surface prior selection.
Generalized extractions.py to use "reducers." Extractions can now take any generic reducer to move between full pixel and super pixel resolution. Necessary for aggregating classified pixels.
Misc cleaning throughout.

There are a couple glint-related updates that I'm going to include in a separate PR.

brentwilder · 2025-08-05T21:38:24Z

This worked for me! Testing the case of a single, multi-component surface model for an EMIT scene I have locally (emit20250327t212148).

pgbrodrick · 2025-08-08T16:20:30Z

+    surface_category: str,
+    multisurface: bool = False,


Add args strings to documentation for these two arguments below.

evan-greenbrg · 2025-08-11T23:35:27Z

+        meshgrid = np.meshgrid(self.rows, self.cols)
+        index_pairs[:, 0] = meshgrid[0].flatten(order="f")
+        index_pairs[:, 1] = meshgrid[1].flatten(order="f")
+        del meshgrid


I tested the indexing function using numpy where and actually found that numpy's performance is actually better than the view-based indexing.

View-based indexing: ~0.03s
Numpy-based indexing: ~0.005s

I'm going to swap the view-based indexing back to numpy. However, there is a speed difference between the memory contiguous index_pairs and the vstack (not contiguous) index_pairs.

Non-contiguous index_pairs: ~0.025s
Contiguous index_pairs: ~0.005s

This function is only called 2 * n_worker times throughout apply_oe so the differences here are so minor to be essentially irrelevant.

evan-greenbrg · 2025-08-12T16:12:37Z

evan-greenbrg · 2025-08-12T16:43:58Z

+            )
+        else:
+            self.svf = []
+


@brentwilder I ended making the skyview file consistent with the rest of the input files in that they are held as memory objects at the Worker level.

I also edited the case of no sky-view file to get around holding the ones array in memory. Let me know if you have any reservations at all.

It'd be super useful for you to run through a sky-view case on this PR whenever you have the bandwidth to do so! (that is once I get the checks passing...)

unbohn · 2025-08-27T18:13:55Z

Passing a .mat instead of a .json file as --surface_path to apply_oe causes a KeyError in line 2002 of template_construction.py if classify_multisurface is set to True as the multisurface classification expects one .mat file per surface category. We should decide on the default behavior here, whether we want to allow both input options, or clearly document that only a .json file is accepted. At least, we should raise a warning in case the user accidentally provides a single prebuilt .mat file to apply_oe.

unbohn · 2025-08-27T21:50:57Z

Ok, the infrastructure runs without errors for me finally. Will inspect results qualitatively now...

unbohn · 2025-08-27T23:03:13Z

Qualitative results are interesting. Here are reflectance RGBs from running a classic single surface inversion on EMIT (upper image) vs. the multisurface approach with two surface categories (lower image):

unbohn · 2025-08-27T23:06:09Z

When looking closer at the RGBs, you notice that the multi surface-based result is going more into the greenish/blueish color. Both images are using the exact same stretching. This becomes more obvious when looking at randomly selected spectra:

There is a magnitude offset and in some cases even a shape difference.

unbohn · 2025-08-28T16:18:51Z

Magnitude and shape differences are very likely introduced by using the 6c emulator for the multi surface version, as opposed to the 1c sRTMnet for the classic L2A. Therefore, I think this PR is ready to merge, based on comprehensive testing of various scenarios.

evan-greenbrg · 2025-08-28T17:38:13Z

Action items:

Resolve question about splitting .mat inputs for multisurface runs
- Long term solution should be to condense to a single output .mat file. surface.component should then be able to only use the priors with the matching categories.
- Short term solution is to only allow .json surface inputs for the multisurface runs. Add checks for this case.
Test processing loads with new resource tracker

…or fit_params Updates to AL and sky glint prior Edits to sky glint prior Editing glint fit parameters and adding indexing Simplifying fit_params wl choice Fixing typoes in surface_glint_model Rebase with dev

Initializing output matrices in analytical line with numpy not envi Fixed classification cleaning function Fixed index shift in multicomponent classifier Analytical Line bug fixes Adding coupled total transm to hueristic atmosphere Removing duplciate analytical_model function from rebase Fixed typo in template_constructions Changing simple to algebraic Fixed logging Changed state order to alphabetical Bad object reference Rolling back transm change Fixing bugs in examples test Fixing rebase issues Fixed examples Fixed bad invert_algebraic return Typo in template Fixed bad classification filtering

…put init func. Missed incorrect function argument

Loose bugs to get running through AOE Removing multi-scattering from AOE. Edits to AOE init to remove discontinuities Glint updates to heuristic atm Fix surface path in config and correct surface creation on non multisurface runs

Fixing classification cli. Tweaking glint est on init Fixed staging and config for single surface json Moving multistate to core Moved multistate.py to core. Fixed surface config for input .mat files. Fixing direct .mat input for surface Bad indent Missed surface_mapping var

Fixing bugs after rebase isofit run gets n_rows from SpectrumFile Moving tests into test_examples editing working directory in apply oe test Edits to test_cli SpectrumFile errors out if trying to build output that already exists Bug fix for tests Added initialization method to IO class Removing multisurface test label updaing flush dirs to reflect correct working dir Update Glint model. Fix surface copying Surface matrices size checks for 1-component surfaces Fixed typos Fixed logging type

evan-greenbrg · 2025-09-09T15:50:01Z

Resource logs for multisurface run:

evan-greenbrg added the enhancement New feature or request label Dec 20, 2024

evan-greenbrg requested review from jammont, pgbrodrick and unbohn December 20, 2024 01:34

unbohn reviewed Dec 20, 2024

View reviewed changes

Comment thread isofit/radiative_transfer/radiative_transfer.py Outdated

unbohn reviewed Dec 20, 2024

View reviewed changes

Comment thread isofit/utils/template_construction.py Outdated

This comment was marked as outdated.

Sign in to view

evan-greenbrg mentioned this pull request Dec 20, 2024

ApplyOE tutorial for the multisurface run isofit/isofit-tutorials#15

Merged

evan-greenbrg force-pushed the multisurface_fresh branch from ddcf117 to 324259e Compare January 13, 2025 16:13

evan-greenbrg force-pushed the multisurface_fresh branch 2 times, most recently from c15d282 to 4ff9896 Compare February 10, 2025 21:09

evan-greenbrg force-pushed the multisurface_fresh branch from 4ff9896 to 4070cff Compare March 24, 2025 16:52

evan-greenbrg force-pushed the multisurface_fresh branch from 386b9c8 to bc054f5 Compare April 7, 2025 17:03

evan-greenbrg marked this pull request as draft April 9, 2025 22:50

evan-greenbrg force-pushed the multisurface_fresh branch from 1ab1c41 to 3223bce Compare May 19, 2025 21:51

evan-greenbrg force-pushed the multisurface_fresh branch from 3223bce to cd08ced Compare May 27, 2025 19:12

evan-greenbrg force-pushed the multisurface_fresh branch 4 times, most recently from 108f56d to 6b72fe5 Compare July 9, 2025 22:44

evan-greenbrg force-pushed the multisurface_fresh branch from 48caa43 to e3a4fcf Compare July 16, 2025 21:59

evan-greenbrg marked this pull request as ready for review July 30, 2025 21:50

evan-greenbrg force-pushed the multisurface_fresh branch from 8fd189c to 1ef884f Compare August 5, 2025 16:40

pgbrodrick reviewed Aug 8, 2025

View reviewed changes

Comment thread isofit/utils/template_construction.py Outdated

evan-greenbrg commented Aug 8, 2025

View reviewed changes

Comment thread isofit/utils/template_construction.py

pgbrodrick reviewed Aug 8, 2025

View reviewed changes

Comment thread isofit/utils/apply_oe.py Outdated

evan-greenbrg commented Aug 11, 2025

View reviewed changes

evan-greenbrg force-pushed the multisurface_fresh branch from c04c10e to fbf0dc4 Compare August 12, 2025 16:00

evan-greenbrg commented Aug 12, 2025

View reviewed changes

This was referenced Aug 12, 2025

Edits to multisurface tests isofit/isofit-tutorials#27

Merged

Broken row_column argument in isofit.run #747

Open

unbohn reviewed Aug 26, 2025

View reviewed changes

Comment thread isofit/utils/apply_oe.py

unbohn reviewed Aug 27, 2025

View reviewed changes

Comment thread isofit/utils/multicomponent_classification.py

unbohn reviewed Aug 27, 2025

View reviewed changes

Comment thread isofit/utils/multicomponent_classification.py Outdated

unbohn reviewed Aug 27, 2025

View reviewed changes

Comment thread isofit/core/isofit.py Outdated

evan-greenbrg force-pushed the multisurface_fresh branch from 96bd26e to da12a71 Compare August 28, 2025 16:29

unbohn approved these changes Aug 28, 2025

View reviewed changes

evan-greenbrg force-pushed the multisurface_fresh branch from b077112 to a581a67 Compare September 8, 2025 22:17

evan-greenbrg added 8 commits September 8, 2025 16:04

Edited sky glint prior, added indexing in glint model, edited bands f…

e77a124

…or fit_params Updates to AL and sky glint prior Edits to sky glint prior Editing glint fit parameters and adding indexing Simplifying fit_params wl choice Fixing typoes in surface_glint_model Rebase with dev

Can now set surface.mat to a specific surface model. Added shared out…

91b75ab

…put init func. Missed incorrect function argument

Added unit tests for multisurface runs

4c23519

Loose bugs to get running through AOE Removing multi-scattering from AOE. Edits to AOE init to remove discontinuities Glint updates to heuristic atm Fix surface path in config and correct surface creation on non multisurface runs

Adding raise error for multisurface run with input .mat surface file

8b2cb9b

Fixing config checks

7fcd635

evan-greenbrg force-pushed the multisurface_fresh branch from a581a67 to 7fcd635 Compare September 8, 2025 23:34

pgbrodrick merged commit b8d4c67 into isofit:dev Sep 11, 2025
17 checks passed

Conversation

evan-greenbrg commented Dec 20, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

This comment was marked as outdated.

This comment was marked as outdated.

This comment was marked as outdated.

brentwilder commented Aug 5, 2025

Uh oh!

pgbrodrick Aug 8, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

evan-greenbrg Aug 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

evan-greenbrg commented Aug 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

evan-greenbrg Aug 12, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

unbohn commented Aug 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

unbohn commented Aug 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

unbohn commented Aug 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

unbohn commented Aug 27, 2025

Uh oh!

unbohn commented Aug 28, 2025

Uh oh!

evan-greenbrg commented Aug 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

evan-greenbrg commented Sep 9, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

evan-greenbrg commented Dec 20, 2024 •

edited

Loading

evan-greenbrg Aug 11, 2025 •

edited

Loading

evan-greenbrg commented Aug 12, 2025 •

edited

Loading

evan-greenbrg Aug 12, 2025 •

edited

Loading

unbohn commented Aug 27, 2025 •

edited

Loading

unbohn commented Aug 27, 2025 •

edited

Loading

unbohn commented Aug 27, 2025 •

edited

Loading

evan-greenbrg commented Aug 28, 2025 •

edited

Loading