# REALPAIR005 — OHRC/TMC-2 local science-pixel registration

**Audit decision: REGISTERED for the tested middle local region, as relative OHRC-to-TMC-2 correspondence on the native 5 m TMC-2 ortho grid.** The decision applies only to OHRC native samples 1,000–11,000 and lines 42,000–58,000 where the camera/DTM projection is valid. **Full-strip registration: ABSTAIN. Absolute ground-coordinate accuracy at 5 m: ABSTAIN.** This is real Chandrayaan-2 science data; no synthetic image is used. A 1.30-pixel residual on the TMC grid is roughly 6.51 m *relative image residual*, not 0.25 m OHRC precision or 6.51 m absolute lunar accuracy.

The original [science-pixel candidate report](SCIENCE_PIXEL_CANDIDATE_REPORT.md) remains an immutable record of earlier failed full-strip transfer and its ABSTAIN decision. This new audit does not turn those failures into passes. The original middle [matcher metrics](science_roi_aligned_middle_test/SIFT/metrics.json) intentionally retain `ABSTAIN_PENDING_INDEPENDENT_GROUND_CONTROL`; this document and the separate [scoped decision](aligned_middle_decision.json) record the later geometry review.

## Source data and reproducibility

| Instrument | Exact product ID | Science SHA-256 |
|---|---|---|
| OHRC | `ch2_ohr_nrp_20241117T2033118394_d_img_d18` | `975a64b823d3129dc64751c0ec82df814479a246a7f0a6a5a578270c04bc1dcd` |
| TMC-2 ortho | `ch2_tmc_ndn_20231101T0125121377_d_oth_d18` | `2648b33e142c47cd08a006702fa1e3360153f87b0458179258349a1b81acd7c5` |
| TMC-2 DTM | `ch2_tmc_ndn_20231101T0125121377_d_dtm_d18` | `01572844b2a7173aa63ffd7a8a851864e888f771773eb9532ec40d2e2287265c` |

The official PDS4 OHRC and TMC ortho labels have SHA-256 `3a47106478ba17c1f5e5f5e782e9c1db1c48eb23769736eb04ae8a2456e6139c` and `84241c0bdb5dc4ab491e923bae4c34265b0ff9571be0fbf76b1985c00a67e3ee`. Their true equal-area polygon intersection is **80.356 km²** ([polygon and CRS](xml_footprint_intersection.json)). This establishes catalogue overlap, not registration. The source manifest is [products.json](../../manifests/products.json); [artifact hashes](aligned_middle_sha256.txt) cover this audit's derived inputs, outputs and figures.

## Findings and fixed test

| Gate | Evidence | Finding |
|---|---|---|
| Real terrain and geometry | OHRC science pixels projected with OHRC SPICE pointing/IK and a TMC DTM aligned by NASA Ames Stereo Pipeline `pc_align` to NASA LOLA 120 m; invalid DTM cells excluded | Local camera/terrain projection valid for 5,444 sampled rays; 72.86% of the output ROI pixels have interpolated source support. |
| Independent navigation check | ASP DEM-to-LOLA alignment plus real TMC ortho against LOLA hillshade; separately, real OHRC against LOLA hillshade under OHRC Sun geometry | TMC DEM correction is supported by ortho; OHRC/LOLA intensity ZNCC rises from 0.110 at unshifted SPICE ground to 0.798 at the *predeclared* image-derived prior rounded to 120 m cells. Both checks support the coarse displacement, not native 5 m absolute accuracy. |
| Frozen correspondence test | New middle region had not supplied science-image pixels to the earlier pointing-prior or matcher tuning; fixed 0.75 ratio, reciprocal match, 500 TMC-pixel candidate window, affine RANSAC ≤4 TMC pixels; spatial-bin held-out points were not used in fitting | SIFT and AKAZE independently pass every fixed local image gate. |
| Spatial and reverse checks | ≥6 training inliers/40%, ≥3 held-out/40%, ≥6 of 18 cells, ≥15% hull, separately fitted reverse-model held-out median ≤3 TMC pixels | Both pass. The reverse model is fitted from training *image* correspondences; it is not an independent ground-control survey. |
| Alternative-location controls | Valid TMC ortho patches 20 km north and 10 km south, selected by pixel validity before matching, same settings | Both SIFT and AKAZE fail fixed support gates at both displaced locations. |
| Accuracy boundary | No surveyed lunar control point or independently validated 5 m OHRC camera/DEM solution | Relative local registration passes; absolute 5 m geolocation and full strip remain ABSTAIN. |

The local [projection provenance](science_roi_aligned_middle/projection.json) records the nominal TMC map ROI `[76795,135470,79715,139760]` m, **584 × 858** 5 m pixels (width × height). The first/final development regions alone supplied the fixed candidate pointing prior of `(-2462.78,-4187.93)` m in the ASP/LOLA-aligned map frame ([prior accounting](pointing_residual_accounting.json), [frozen config](aligned_candidate_config_v1.json)). This is an **image-derived search/projection prior, not a published camera correction**. TMC original image values and OHRC IMG values were unchanged; the local [VRT manifest](aligned_tmc_local/manifest.json) records the navigation-only TMC grid adjustment. DTM heights were adjusted separately by the recorded local vertical plane. The same [TMC/LOLA navigation fit](tmc_navigation_affine.json) is restricted to this region.

Before any local image affine was fitted, the frozen candidate geometry already put **544/900 SIFT** and **287/426 AKAZE** candidate pairs within the 4-pixel residual gate ([reproducible pre-fit check](aligned_middle_prefit_support.json)). The independent [OHRC/LOLA screen](ohrc_vs_lola_independent_screen.json) used 480 × 480 *native OHRC science-pixel* bins, direct SPICE rays, aligned terrain and LOLA hillshade, without applying the candidate shift to ground coordinates. Its best 120 m-cell displacement was `(−21,+35)`, exactly the rounded prior; the Sobel structural match peaked at `(−20,+34)`. This checks the kilometre-scale geometry against a separate terrain image, while acknowledging that the terrain projection uses the TMC DTM whose navigation came from the same LOLA reference. The [ASP run audit](asp_lola_alignment_audit.json) reports 67 hillshade matches and 2.40 m post-alignment median nearest-point *DEM shape* error at 120 m; that number is not a ground-control error for this image registration.

| Method | Candidates | Training inliers | Untouched held-out inliers | All accepted controls | Relative RMSE | P90 | Coverage; hull | Held-out reverse median |
|---|---:|---:|---:|---:|---:|---:|---|---:|
| SIFT | 900 | 785/789 = 99.49% | 93/111 = 83.78% | **878** | 1.302 TMC px ≈ **6.51 m** | 1.985 px ≈ 9.93 m | 11/18; 32.71% | 1.263 TMC px |
| AKAZE | 426 | 362/370 = 97.84% | 48/56 = 85.71% | **410** | 1.435 TMC px ≈ **7.18 m** | 2.187 px ≈ 10.93 m | 10/18; 27.93% | 2.201 TMC px |

Accepted controls from the two methods are **not summed**: SIFT and AKAZE can observe the same physical terrain. All residuals above are in **native 5 m TMC-2 ortho pixels** after locally fitting each method's affine transform using training bins only. The threshold was fixed at 4 pixels/20 m. The reverse error uses a separately fitted reverse affine tested at observed held-out points; it is not merely the algebraic inverse of the forward matrix. Full match tables, accepted inliers, reverse matches, transforms, configs and manifests are in [SIFT](science_roi_aligned_middle_test/SIFT) and [AKAZE](science_roi_aligned_middle_test/AKAZE). [Control-point provenance](science_roi_aligned_middle_test/control_point_provenance.json) provides approximate native OHRC sample/line, original native TMC pixel, aligned map coordinate and per-match residual for every accepted correspondence. All 878 SIFT and 410 AKAZE accepted points have a native OHRC lookup. These coordinate records inherit the 5 m absolute geolocation limit.

On **accepted held-out points alone**, SIFT forward RMSE is 1.536 TMC pixels (≈7.68 m) and AKAZE is 1.900 pixels (≈9.50 m). These values are conditional on the fixed 4-pixel inlier gate and therefore exclude rejected held-out candidates; the held-out inlier ratios above must accompany any accuracy statement.

The [held-out correspondence figure](aligned_middle_evidence/sift_heldout_correspondences.png) and [local overlay](aligned_middle_evidence/local_affine_overlay.png) show terrain-scale agreement, with [figure provenance](aligned_middle_evidence/figure_provenance.json). A post-hoc [shadow diagnostic](shadow_diagnostic_aligned_middle.json) found 45.0% of valid OHRC pixels dark under the chosen threshold but only 6.7% of accepted SIFT points dark. The threshold was not used to select matches and does not inflate the fixed held-out test. [Displaced-region controls](negative_controls_aligned_middle/metrics.json) yielded at most 6 SIFT or 4 AKAZE candidates and no accepted held-out support.

## Decision boundary

The audit accepts **relative local correspondence** because real science-image structures, frozen held-out SIFT and AKAZE support, spatial coverage, separately fitted reverse checks, camera/terrain geometry, coarse OHRC/LOLA and TMC/LOLA checks, and rejected alternative-location controls agree. It does **not** establish a lunar absolute control network, full OHRC strip registration, or sub-metre accuracy. The local affine absorbs remaining camera/terrain residuals, so its small image residual cannot itself certify the physical camera parameters. Regions with invalid DTM support or failed transfer retain ABSTAIN. Runtime in matcher-only metrics is below one second per method; full projection/runtime and peak memory were not instrumented, so no end-to-end speed or memory claim is made.

The next scientific extension is an independent 5 m-class surveyed/externally controlled check or a rigorously held-out camera/DEM bundle adjustment, then new untouched regions of the OHRC strip. IIRS must be evaluated as a separate TMC-to-IIRS leg; no three-sensor registration is claimed by this audit.