Sentinel-1 SAR Backscatter -- VH / VV / VH:VV RGB Composite¶

This notebook is an end-to-end manual validation of the sar_backscatter process (ADR docs/adr/0001-sar-backscatter.md, Phase 1: ellipsoid coefficients only). It loads Sentinel-1 GRD vv/vh, calibrates both polarisations to sigma0-ellipsoid, and assembles the classic dual-pol SAR false-colour composite used to sanity-check a backscatter product by eye:

Channel	Source	What it emphasises
Red	VH (cross-pol)	Volume scattering -- vegetation, forest canopy
Green	VH / VV (as dB difference: VH(dB) - VV(dB))	Scattering-mechanism contrast -- vegetation vs. bare/smooth surfaces
Blue	VV (co-pol)	Surface / double-bounce scattering -- water (dark), urban structures (bright)

There is no ground truth check here -- this is a plausibility validation (does the scene look like a real Sentinel-1 composite: water dark, vegetation reddish, urban bright), the same spirit as the ADR's own acceptance criterion 1 (S7.8), just visual instead of a single dB number.

Requires: a running local titiler-openeo backend configured against a real STAC catalogue with credentials (see .env.cdse), and the openeo Python client (pip install openeo) -- these notebooks run in their own environment, openeo is not a project dependency.

Import Required Libraries¶

In [1]:

import matplotlib.pyplot as plt
import numpy as np
import rasterio
import openeo

import matplotlib.pyplot as plt import numpy as np import rasterio import openeo

Connect to OpenEO Backend¶

Connect to a local OpenEO backend running on port 8081 and authenticate using OpenID Connect (same pattern as the other notebooks in this folder).

In [2]:

# connection = openeo.connect(
#     url="https://openeo.ds.io"
# ).authenticate_oidc_authorization_code()

connection = openeo.connect(
    url="http://127.0.0.1:8082/"
).authenticate_oidc_authorization_code()

# connection = openeo.connect( # url="https://openeo.ds.io" # ).authenticate_oidc_authorization_code() connection = openeo.connect( url="http://127.0.0.1:8082/" ).authenticate_oidc_authorization_code()

Load Sentinel-1 GRD¶

The AOI below covers Tokyo Bay: open water in the bay, dense urban/industrial structures around it, and hills inland -- enough land-cover variety to exercise all three composite channels in one scene.

Pick an extent that a real pass actually covers. Two traps, both worth knowing:

• temporal_extent must be ISO 8601 -- "2026-07-08T20:42:55Z", with the T. A space instead of the T is rejected by the openEO client.
• Sentinel-1 footprints are slanted parallelograms, so a bbox can intersect an item's axis-aligned bbox (which is what a STAC search matches) while missing the real footprint entirely. When that happens the read returns all no-data, every pixel is masked, and the composite is uniformly black -- which looks like a calibration bug but is not. The mask band is the quickest way to tell the difference: all-zero means "no data reached this grid", not "calibration produced zeros".

A narrow, sub-minute window is a good way to pin exactly one acquisition.

In [3]:

# Tokyo Bay AOI
spatial_extent = {"west": 139.5, "south": 35.2, "east": 140.2, "north": 35.8}
# spatial_extent = {"west": 4.20, "south": 51.85, "east": 4.55, "north": 51.98}
temporal_extent = [
    "2026-07-08T20:42:55Z",
    "2026-07-08T20:43:20Z",
]  # adjust to a real S1 pass over the AOI

s1 = connection.load_collection(
    "sentinel-1-grd",
    spatial_extent=spatial_extent,
    temporal_extent=temporal_extent,
    bands=["vv", "vh"],
)

# Tokyo Bay AOI spatial_extent = {"west": 139.5, "south": 35.2, "east": 140.2, "north": 35.8} # spatial_extent = {"west": 4.20, "south": 51.85, "east": 4.55, "north": 51.98} temporal_extent = [ "2026-07-08T20:42:55Z", "2026-07-08T20:43:20Z", ] # adjust to a real S1 pass over the AOI s1 = connection.load_collection( "sentinel-1-grd", spatial_extent=spatial_extent, temporal_extent=temporal_extent, bands=["vv", "vh"], )

Calibrate to Backscatter¶

sar_backscatter isn't one of the openEO client's built-in convenience methods (it's not a core process), so it's called generically via .process(). coefficient= "sigma0-ellipsoid" and noise_removal=True are within Phase 1's supported set -- gamma0-terrain, the openEO spec's own default, is deliberately rejected by this back-end (see the ADR). mask=True adds a mask band (1.0 valid, 0.0 border/no-data) so the composite below can black out the scene's own padding instead of plotting noise there.

In [4]:

backscatter = s1.process(
    "sar_backscatter",
    data=s1,
    coefficient="sigma0-ellipsoid",
    noise_removal=True,
    mask=True,
)

# `firstpixel` collapses the temporal dimension by taking the first valid pixel per
# location, which also stitches adjacent slices of the same pass (each slice is its
# own timestamp, so each is its own temporal slice) into one gap-free image.
backscatter = backscatter.reduce_dimension(dimension="t", reducer="firstpixel")
backscatter = backscatter.resample_spatial(projection="EPSG:3857")

backscatter.download("sar_backscatter_vv_vh_mask.tif", format="GTiff")

backscatter = s1.process( "sar_backscatter", data=s1, coefficient="sigma0-ellipsoid", noise_removal=True, mask=True, ) # `firstpixel` collapses the temporal dimension by taking the first valid pixel per # location, which also stitches adjacent slices of the same pass (each slice is its # own timestamp, so each is its own temporal slice) into one gap-free image. backscatter = backscatter.reduce_dimension(dimension="t", reducer="firstpixel") backscatter = backscatter.resample_spatial(projection="EPSG:3857") backscatter.download("sar_backscatter_vv_vh_mask.tif", format="GTiff")

Build the VH / VV / VH:VV Composite¶

Bands come out in the order sar_backscatter documents: the requested polarisations first (vv, vh, matching bands= above), then mask (ellipsoid_incidence_angle wasn't requested, so it's absent). The calibrated values are still linear power -- SAR composites are conventionally built in decibels, which turns the VH/VV ratio into a difference, dB(VH/VV) = dB(VH) - dB(VV), and conveniently sidesteps dividing by a near-zero VV.

In [5]:

def to_db(power, floor=1e-6):
    """Linear power -> dB, flooring at `floor` so log10(0) doesn't warn/-inf."""
    return 10 * np.log10(np.maximum(power, floor))


def stretch(band, valid, p_low=5, p_high=95):
    """Percentile-stretch `band` to 0-1 using only its valid pixels."""
    lo, hi = np.percentile(band[valid], [p_low, p_high])
    return np.clip((band - lo) / (hi - lo), 0, 1)


with rasterio.open("sar_backscatter_vv_vh_mask.tif") as src:
    vv, vh, mask_band = src.read()

valid = mask_band > 0.5

vv_db = to_db(vv)
vh_db = to_db(vh)
ratio_db = vh_db - vv_db  # dB(VH / VV)

red = stretch(vh_db, valid)
green = stretch(vv_db, valid)
blue = stretch(ratio_db, valid)

rgb = np.dstack([red, green, blue])
rgb[~valid] = 0  # black out border/no-data rather than plotting noise

def to_db(power, floor=1e-6): """Linear power -> dB, flooring at `floor` so log10(0) doesn't warn/-inf.""" return 10 * np.log10(np.maximum(power, floor)) def stretch(band, valid, p_low=5, p_high=95): """Percentile-stretch `band` to 0-1 using only its valid pixels.""" lo, hi = np.percentile(band[valid], [p_low, p_high]) return np.clip((band - lo) / (hi - lo), 0, 1) with rasterio.open("sar_backscatter_vv_vh_mask.tif") as src: vv, vh, mask_band = src.read() valid = mask_band > 0.5 vv_db = to_db(vv) vh_db = to_db(vh) ratio_db = vh_db - vv_db # dB(VH / VV) red = stretch(vh_db, valid) green = stretch(vv_db, valid) blue = stretch(ratio_db, valid) rgb = np.dstack([red, green, blue]) rgb[~valid] = 0 # black out border/no-data rather than plotting noise

Visualize the Composite¶

In [6]:

plt.figure(figsize=(10, 10))
plt.imshow(rgb)
plt.title(
    "Sentinel-1 backscatter sigma0 composite: R=VH, G=VV (dB), B=VH/VV (Tokyo Bay)"
)
plt.axis("off")
plt.show()

plt.figure(figsize=(10, 10)) plt.imshow(rgb) plt.title( "Sentinel-1 backscatter sigma0 composite: R=VH, G=VV (dB), B=VH/VV (Tokyo Bay)" ) plt.axis("off") plt.show()

Interpreting the Composite¶

• Water (port basins, rivers): dark in both polarisations -- a smooth surface scatters the radar pulse away from the sensor rather than back to it -- reads as near-black.
• Urban / industrial (port structures, buildings, containers): bright in VV from double-bounce reflection off vertical structures -- reads bright blue/white.
• Vegetation / farmland: relatively brighter in VH, from volume scattering inside the canopy -- reads reddish/pink, the classic "vegetation red" of this composite.

If the scene instead looks uniformly noisy, inverted (water bright, urban dark), or has a visible seam or gradient across the image, that is a real signal something in the calibration path is wrong -- worth cross-checking against the reference numbers in scripts/sar_backscatter_prototype.py (ADR S7.3) before trusting the pipeline further.

Scope reminder: this is Phase 1 -- ellipsoid geometry only, no terrain correction. Expect visible brightness gradients with local incidence angle over hilly terrain; that is expected, not a bug (see the ADR's scientific accuracy statement, S5).