[mgovorcin]

Change Summary

baseline.compute got the perpendicular baseline from the two ranges by the law of cosines, then sin_theta = np.sqrt(1 - cos_theta**2). That fails in two ways:

1. NaN where B⊥ crosses zero. With the baseline nearly parallel to the look, rounding pushes |cos_theta| just above 1 and the square root returns NaN. In DISP-S1 products this shows up as a NaN band in corrections/perpendicular_baseline along the B⊥ = 0 line. In one set of 376 archived products, 15 have it, 20k–2.7M pixels each.
2. Bias away from zero. compute_baselines (and disp-s1's _baselines.compute_baselines) removes the along-track offset from the secondary position, but passes the secondary range measured before that shift. On S1A–S1B pairs this is a 0.1–0.5 mm range mismatch, which the law of cosines turns into centimetres of B⊥ error near zero: up to 19 cm at a 32 m along-track shift.

The fix computes B⊥ directly as the baseline component perpendicular to the unit look vector, |B × l̂|. It doesn't use the ranges, so neither problem can occur. The sign rule is unchanged. ref_range / sec_range stay in the signature, now unused, because disp-s1 calls dolphin.baseline.compute with them positionally.

Away from zero nothing changes: on tests/data/baseline_cslcs (B⊥ ≈ 135–142 m) the new values are within 0.04 mm of the old ones.

Verification on archived DISP-S1 products

For each of the 15 affected products, B⊥ was recomputed through disp-s1's own code (disp_s1._baselines.compute_baselines on the product grid every 50th pixel, then _interpolate_data, as in product.py). The inputs were the first reference and secondary CSLC by name from ASF, run once with this branch and once with main:

• With main the recomputation reproduces every archived layer: the same NaN pixels and values within 0.03 mm. So the NaNs come from this function and not from anything downstream.
• With this branch all 15 have 0 NaN.
• Given ranges consistent with the shifted secondary position, the old formula equals the new one to within 1e-5 m (usually 1e-7 m) at 400 points per product. So the remaining differences are the old formula's bias, not a change in definition.

Top: an S1A–S1A pair, where the NaN band disappears. Bottom: an S1A–S1B pair with a 21–32 m along-track shift, where the band disappears and the bias away from the zero line is removed too. In the figure titles, "where |B⊥| > 0.5 m" is measured on the fixed layer; in the table below it's measured on the archived layer.

Frame	Reference → secondary	NaN px, archived	NaN px, old code rerun	NaN px, fix	Along-track shift removed (max)	Fix − archived where |B⊥| > 0.5 m
F03848	20170827 → 20171026	1,026,502	1,026,502	0	14.0 m	0.75 cm
F09154	20160705 → 20160729	2,336,590	2,336,590	0	18.6 m	0.72 cm
F09154	20160705 → 20170101	176,955	176,955	0	13.4 m	1.33 cm
F09154	20160705 → 20170119	2,653,491	2,653,491	0	32.1 m	7.51 cm
F16943	20170906 → 20180715	523,506	523,506	0	1.5 m	0.01 cm
F23443	20170516 → 20170925	20,351	20,351	0	1.1 m	0.00 cm
F24196	20160720 → 20161129	817,400	817,400	0	5.6 m	0.28 cm
F34482	20160711 → 20160828	312,504	312,504	0	4.3 m	0.02 cm
F34482	20171010 → 20171115	1,224,014	1,224,014	0	14.4 m	0.22 cm
F34482	20171010 → 20180102	461,156	461,156	0	21.0 m	0.46 cm
F34732	20170314 → 20171121	140,400	140,400	0	1.4 m	0.01 cm
F44329	20160714 → 20161223	55,450	55,450	0	5.7 m	0.11 cm
F44329	20160714 → 20170116	221,000	221,000	0	2.9 m	0.01 cm
F44329	20160714 → 20170609	2,196,815	2,196,815	0	19.8 m	0.62 cm
F44329	20170609 → 20171112	2,287,969	2,287,969	0	24.2 m	3.65 cm

Same comparison on one burst (t035_073233_iw3, 20160705 → 20160729), on the CSLC lon/lat grid:

How to verify

pytest tests/test_baseline.py

The new tests put the secondary 100 m or 1000 m along the look and 10 m to 0 m across it, and check that |B⊥| is recovered to 1e-6 m and is finite. On main they fail at 0.01 m and 0 m. A second test checks that the sign flips with the side.

Pre-flight: pre-commit run --all-files clean. pytest: 781 tests run (tests/test_io_paths.py skipped, moto not installed): 757 passed, 21 skipped, 24 failed. The same 24 also fail on main in that environment (phass not installed, gdal_merge.py not on PATH), so none come from this change.

Not in this PR

• Products already in the archive keep their NaN bands and bias until they're reprocessed with a release that includes this.
• test_baseline_compute has two bare pytest.approx(...) lines that never assert, and its expected minimum (134.98) doesn't match what the function returns on main (134.51). That needs a separate fix.

Related issue number

None.

Checklist

• The pull request title is a good summary of the changes - it will be used in the changelog
• Unit tests for the changes exist
• Tests pass on CI
• Documentation reflects the changes where applicable
• My PR is ready to review

🤖 Generated with Claude Code