v2.0.4

Sentinel-1 (SAR) workflow update

Sentinel-1 support has been overhauled. CoastSeg now downloads both the VV and VH polarizations
instead of VH alone, and maps shorelines with a new 3-band SAR segmentation model instead of an
Otsu threshold.

Highlights

• Sentinel-1 downloads now fetch VV and VH

  • Previously CoastSeg requested VH only, so each scene was a single band.
  • The new model needs both polarizations, so both the VV and VH are downloaded by default.
  • Action required: ROIs downloaded before this release are VH-only. Re-run the download for
    those sites and the missing band is fetched, overwriting the existing scenes in place.

• New 3-band SAR segmentation model

  • Each scene is stacked into a [VV, VH, VV-VH] dB composite, so the model sees a much more
    detailed version of the same acquisition than a single grey band when only the VH was used.
  • VV, VH and VV-VH respond differently to water, wet sand, vegetation and urban
    surfaces.
  • The new SAR model is hosted on Hugging Face at
    2320sharon/SAR_3_band_model.

• The new SAR 3 band model is now the default for SAR in both workflows

  • It runs in the CoastSat workflow (extract_all_shorelines) and the Zoo workflow
    (4_zoo_workflow_SAR.py and the SAR notebook) with no settings change.
  • Nothing downstream changed: the shoreline is still the land/water boundary inside the reference
    shoreline buffer, so transects, tide correction and sessions behave as before.
  • The SAR model runs in-process through onnxruntime, so unlike the other Zoo models it does
    not need the segmentation_workflow environment.

• Otsu thresholding is now only a fallback method for SAR

  • Set sar_segmentation = "otsu" to force it, or let CoastSeg use it automatically when the model
    cannot be loaded (no onnxruntime, no internet on first use, a bad sar_model_path).
  • Otsu needs a cleanly bimodal backscatter histogram, so it drifts on wind-roughened water, wet
    sand and radar shadow, and the drift changes from date to date. Below is the same scene with the
    same settings — only the segmentation differs.

Getting the model

• By default, do nothing. The model (~130 MB) is not shipped with CoastSeg. On the first SAR run
  it is downloaded from Hugging Face into a per-user cache folder outside the repository, so both
  workflows share one copy and later runs reuse it.
• To download it manually (offline machines, restricted networks, or your own retrained model),
  get SAR_3_band_model.onnx from
  Hugging Face and save it to
  CoastSeg/models/SAR_segmentation_model/. A model in that folder always wins over the download and
  both workflows find it with no settings change.
• To load one from anywhere else, point the workflow at it:

# CoastSat workflow: the .onnx file itself
coastsegmap.set_settings(sar_model_path=r"models/SAR_segmentation_model/SAR_3_band_model.onnx")

# Zoo workflow: the folder holding the .onnx
model_setting = {"img_type": "SAR", "local_model_path": r"models/SAR_segmentation_model"}

New settings

setting	workflow	default	what it does
sar_segmentation	CoastSat	"model"	"model" runs the 3-band model; "otsu" forces the legacy threshold for the whole run
sar_water_threshold	both	0.5	P(water) cutoff. Raise it to classify less of the scene as water, lower it to catch more
sar_model_path	CoastSat	""	Your own .onnx. Stored relative to the CoastSeg folder so sessions stay portable
local_model_path	Zoo	""	Folder holding your own .onnx; "" uses/downloads models/SAR_segmentation_model
sar_mask_nodata	Zoo	True	Excludes the empty parts of a SAR swath so they are not traced as land
sar_overwrite_segmentation	Zoo	True	Re-segments scenes whose outputs already exist

sar_segmentation and sar_water_threshold have widgets on the advanced settings tab; the model
paths are set through set_settings or config.json.

Other updates

• Every shoreline now records how it was segmented in output["segmentation_method"], and
  extracted_shoreline.report_sar_segmentation_methods(output) tallies a finished run — useful
  because individual scenes can fall back to Otsu even when the rest of the run used the model.
• When comparing outputs, note that Otsu shorelines carry a dB threshold in MNDWI_threshold while
  model shorelines carry NaN. Filter with extracted_shoreline.threshold_filter_applies(output)
  rather than reading that column directly, or the model rows are silently dropped.
• No speckle filtering is applied any more. The old workflow ran a size=15 median filter over
  the imagery before thresholding it. The SAR imagery is now left untouched — the model reads the raw
  dB bands and was trained that way, so smoothing them would only blur the shoreline.

What's Changed

• Release v2.0.4 by @2320sharon in #323
  Full Changelog: v2.0.3...v2.0.4