0026 EMI is regularized by default; DS are selected by temporal coherence¶
Status¶
Accepted
Date¶
2026-10-05
Context¶
Decision 0023 added the adaptive regularization of EMI with regularize defaulting to False, so that the
verified examples did not change. Without it EMI silently returns a phase history that follows the
estimation noise wherever the coherence magnitude matrix is not positive definite, and that is the common
case for real stacks: 97 % of the DS candidates of Campi Flegrei (92 images) and 98 % of Xinpu (60 images)
with 11 x 11 windows, because 121 SHPs are only about 45 independent looks (decision 0025). With the
regularization the EMI quality of a point depends on its β and cannot be compared between points, so the
example's selection by EMI quality (1.0 to 1.05) and temporal coherence no longer fits.
Decision¶
regularizedefaults to True inemiandemperical_co_emi_temp_coh_pc(API and commands). Points with a positive definite, well conditioned coherence magnitude matrix are unchanged.examples/03_ds.tomlselects DS by temporal coherence alone (0.8 to 1.0); the EMI quality range and the intersection are removed. The EMI quality pyramid stays as a diagnostic (removed with the EMI quality output by decision 0027).
Consequences¶
- The sample data (17 images) change little: see
docs/workflows/03_ds.mdto05_unwrap.mdfor the numbers of the rerun examples. - Code that used negative EMI qualities to find points with a matrix that is not positive definite needs
regularize=False. - Decision 0023 stays accepted; its default is changed by this record.
Do not¶
- Do not select regularized DS by a fixed range of the EMI quality.
- Do not switch the regularization off for stacks of many images without checking the fraction of negative EMI qualities.