04 Merge PS and DS, refine with n2ft

Pipeline: examples/04_refine.toml. Tutorial: nbs/Tutorials/CLI/CampiFlegrei/04_refine.ipynb. Needs 02_ps and 03_ds.

moraine run examples/04_refine.toml --workdir WORK

Steps

1. pc-union of the DS (their phase history) and the PS candidates (their rslc), with e/n; a point in both sets keeps the DS phase history (the first input of pc-union wins); ras2pc adds lon/lat.
2. image-pairs (sequential) and n2ft: Noise2Fringe Transformer filtering of the point cloud interferograms (chunks of 20 000 points with k nearest neighbour halos).
3. temp-coh between the filtered and the raw phase; points with temporal coherence > 0.8 are kept (pc-logic-pc, pc-select-data).

Outputs (WORK/pc/): merged candidates pc_can_*; refined points pc_hix.zarr, pc_ph.zarr (n_points, nimages), their coordinates pc_e.zarr, pc_n.zarr, pc_lon.zarr, pc_lat.zarr, their temporal coherence pc_temp_coh.zarr, and the geometry the deformation products need: height pc_hgt.zarr, look vector pc_theta.zarr, pc_phi.zarr (GAMMA elevation and orientation angles, radians) and slant range pc_range.zarr (m); pyramids pc_can_temp_coh_pyramid, pc_ph_pyramid.

Parameters

• Temporal coherence threshold: 0.7-0.85; lower keeps more points with more noise.
• temp_coh runs on the CPU (cuda = false in the step): it reads the data once and reduces them, starting a GPU worker costs more than the GPU saves.
• n2ft chunks: points per chunk (20 000); smaller uses less GPU memory.
• n2ft compile: the model is compiled with torch.compile in every worker when points times image pairs is at least 1e8 (15-40 s once per worker, less when torch has cached it; the model then runs about 3 times faster on a GPU); compile = true or false forces it. Not worth it for the sample data.

Expected results (sample data, Campi Flegrei, 981 x 4160 x 92)

result	sample value	sane range
merged candidates	675 687	about PS + DS minus overlap
pc_can_temp_coh_pyramid	p50 0.79, p01 0.21, p99 0.995	0..1
refined points (pc_hix)	330 754 (49 % of the candidates; 99.9 % of the DS, 32 % of the PS candidates that are not DS)

Run time: about 40 s with an A100 (n2ft 28 s for 91 image pairs).

Figures

Quicklooks of the sample data run of 2026-10-10 (the numbers above are from the same run):

Temporal coherence of the merged candidates after the Noise2Fringe Transformer (pc/pc_can_temp_coh_pyramid).

Sequential interferogram 45 (2020-07-07 / 2020-07-19) of the refined points (pc/pc_ph_pyramid --show intf_seq --index 45): continuous phase, no noise.

Interferogram 2019-01-02 / 2021-12-29 of the refined points (--show intf_all --index 0 91): the concentric fringes of the caldera uplift centred on Pozzuoli, about ten fringes over three years.

Checks

• moraine quicklook pc/pc_ph_pyramid --show intf_seq --index I -o pc_intf.png: interferograms of the refined points should show smooth, spatially continuous phase (fringes), not noise.
• Compare with the raw interferogram: moraine quicklook raw/rslc_pyramid --show intf_seq --index I.

Problems

• n2ft is slow on CPU; use a GPU.
• Too few points: lower the temporal coherence threshold or revisit PS / DS thresholds.