Regionally-Resolved Spectra Retrieval

• Regionally-resolved spectra retrieval is a method that inverts mixed, indirect measurements to extract localized spectral information using explicit forward models.
• It employs diverse techniques such as direct cube reconstruction, Fourier interpolation, and multidimensional atmospheric modeling to enhance spectral fidelity.
• Applications span solar slit spectroscopy and exoplanet phase curves, providing improved constraints on local thermal structures and compositional properties.

Regionally-resolved spectra retrieval denotes a class of inverse problems in which measurements that are spatially mixed, indirectly encoded, or globally integrated are transformed into spectra associated with localized structures such as spatial pixels, latitude bins, longitude sectors, annular terminator regions, or spectrally defined surface components. In the cited literature, the retrieved object may be a restored spatial-spectral cube, a set of local count spectra, a longitude-resolved spectrum, a latitude-dependent vertical atmospheric profile, or a small number of representative spectra for clustered regions rather than a full map. The unifying idea is that the measured signal is not interpreted as an isolated spectrum per detector sample; instead, spatial mixing, viewing geometry, instrumental encoding, or temporal modulation is modeled explicitly and then inverted or regularized to recover local spectral information (Noort, 2017, Picone et al., 2023, Volpara et al., 2024, Cubillos et al., 2021, Nixon et al., 2022).

1. Conceptual scope and meanings of “regional”

The term is used differently across subfields. In ground-based solar long-slit spectroscopy, it refers to recovering an undegraded spectrum for each spatial location in a scanned field from seeing-mixed slit data, yielding a spatial-spectral cube rather than a degraded slit trace (Noort, 2017). In interferometric hyperspectral imaging, it refers to reconstructing the whole hyperspectral cube jointly and then extracting spectra from selected regions, object interiors, or local neighborhoods; the regional spectrum is therefore a downstream product of cube reconstruction rather than a directly parameterized quantity (Picone et al., 2023).

In hard X-ray solar imaging spectroscopy, “imaging spectroscopy” is defined as “the generation of spatially resolved count spectra and of cubes of count maps at different energies,” so the regional product is a local count spectrum extracted from an energy-indexed image cube reconstructed from visibilities (Volpara et al., 2024). In exoplanet phase-curve work, the regional object may be a longitude-resolved spectrum obtained by inverting phase-dependent spectra into spectra associated with specific longitudes, or a hemisphere-weighted spectrum at selected orbital phases that constrains a shared low-dimensional thermal field (Cubillos et al., 2021, Yang et al., 2024).

A recurring misconception is that regionally-resolved retrieval must imply a full 2D or 3D map with one freely retrieved spectrum per spatial cell. Several of the cited methods do not do that. Some reconstruct a cube first and only afterward permit arbitrary regional extraction (Picone et al., 2023); some retrieve a small number of clustered eigenspectra from eclipse maps (Mansfield et al., 2020); some infer only filling factors of surface components such as quiet Sun and plage from disk-integrated spectra (Pevtsov et al., 2014); and some retrieve time-variable atmospheric parameters whose modulation implies unresolved heterogeneity without yielding a unique geographic map (Wang et al., 22 Feb 2026). This suggests that “regional” in the literature is best understood as a statement about the localization of spectral inference, not about one fixed reconstruction architecture.

2. Forward models and inverse formulations

Most methods are organized around an explicit forward operator that mixes or encodes local spectra before measurement. In restored solar slit spectroscopy, the undegraded solar scene is $s(u,v,\lambda)$, while each degraded slit datum is modeled as a weighted sum of nearby undegraded contributions through a space-variant PSF estimated from strictly synchronized slit-jaw images. Restricting attention to the slit yields

$d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$

and stacking all measurements gives

${\bf J}\cdot{\bf s}={\bf d},$

with normal equations

$${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$

The retrieval capability follows directly from the fact that every degraded slit sample is treated as a known linear combination of unknown localized spectra (Noort, 2017).

In interferometric hyperspectral imaging, the forward operator is spectrally mixing but spatially separable. The unknown cube $X\in\mathbb{R}^{I\times J\times K}$ is mapped to an interferogram-domain tensor $Y\in\mathbb{R}^{I\times J\times L}$ by

$Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$

where $\mathbf{A}\in\mathbb{R}^{L\times K}$ is the transmittance response matrix. For Fabry–Perot interferometers,

$$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$

Here the physics does not mix neighboring spatial pixels in the forward model; spatial coupling is introduced only in the inverse problem through regularization (Picone et al., 2023).

In STIX hard X-ray imaging spectroscopy, the measurement model at one energy channel is the sparse Fourier relation

${\bf V}={\bf A}{\bf f},$

with $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$0 the measured visibilities and $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$1 the count image. RIS adds an interpolation stage,

$d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$2

where the basis functions depend on a scale function derived from the previously reconstructed adjacent-energy image, thereby transferring morphology across contiguous energy bins in count space (Volpara et al., 2024).

Exoplanet multidimensional transmission retrievals generalize the forward model from a single column to a line-of-sight integral through a spatially varying terminator. In AURA-3D, the total transit depth is

$d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$3

with $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$4 determined by slant optical depths $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$5 computed along rays through a 3D atmosphere $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$6 (Nixon et al., 2022). The common structural point across all of these examples is that regional spectral inference emerges only after the spatial or spectral mixing operator has been made explicit.

3. Direct reconstruction of spatial-spectral cubes

The clearest literal realization of regionally-resolved spectra retrieval is direct cube reconstruction. In solar long-slit restoration, the missing spatial dimension is supplied by synchronized slit-jaw imaging, PSF estimation with MOMFBD, and slit scanning across the solar surface. Because the PSF influence radius is only a couple of arcseconds under moderate to good seeing, the normal-equation matrix is fairly local and “sufficiently diagonal,” enabling a damped iterative solve with a multigrid-like strategy. The implementation is explicitly unregularized; iteration is stopped empirically when the reconstructed power spectrum begins to rise at the highest spatial frequencies, signaling deconvolution-driven noise amplification (Noort, 2017).

That method also makes clear why long-slit spectroscopy is a special inverse problem. A spectrograph sacrifices one spatial dimension at the slit in exchange for spectral resolution, so atmospheric turbulence does not merely blur the image: it mixes spectra from neighboring solar locations before dispersion. The paper therefore reconstructs monochromatic images independently under a wavelength-decoupling approximation and stacks them into a restored spectral cube. In active scanning mode at the Swedish 1-m Solar Telescope, scans covered roughly $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$7, with vertically binned pixels of about $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$8, and 40 frames summed over 400 ms corresponded to a scan displacement of $d_{i,y,x_{\rm slit}}={\bf a}_i^T\cdot{\bf s},$9, close to critical sampling relative to the ${\bf J}\cdot{\bf s}={\bf d},$0 diffraction limit at 6300 Å. A “sit-and-stare” example with slit width ${\bf J}\cdot{\bf s}={\bf d},$1 recovered spectra over a ${\bf J}\cdot{\bf s}={\bf d},$2-wide strip in 23 s, albeit with the highest S/N concentrated in the central four pixel rows (Noort, 2017).

Interferometric hyperspectral reconstruction tackles a different encoding but reaches the same cube-first endpoint. The reconstruction objective is a penalized least-squares problem with a spectro-spatial operator

${\bf J}\cdot{\bf s}={\bf d},$3

so that spectral sparsity in DCT coordinates is combined with spatial total variation on the coefficient maps. The resulting collaborative norm ${\bf J}\cdot{\bf s}={\bf d},$4 defines a convex collaborative total variation prior, optimized by an over-relaxed Loris–Verhoeven primal-dual splitting scheme (Picone et al., 2023).

The practical consequence is that regional spectra are more stable because the cube is reconstructed jointly rather than pixel by pixel. In simulated Fabry–Perot acquisitions based on a real Specim IQ scene, with ${\bf J}\cdot{\bf s}={\bf d},$5 spatial sampling, ${\bf J}\cdot{\bf s}={\bf d},$6 spectral channels, ${\bf J}\cdot{\bf s}={\bf d},$7 interferogram samples, and ${\bf J}\cdot{\bf s}={\bf d},$8, the joint LV-CTV method achieved RMSE ${\bf J}\cdot{\bf s}={\bf d},$9 and SSIM $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$0, outperforming ADMM-TV, ADMM-BM3D, LV-DCT, and LV-TV. The paper’s interpretation is that combining spectral and spatial regularization improves both local spectral signatures and their spatial arrangement, which is precisely the condition needed for reliable region-level spectral extraction after reconstruction (Picone et al., 2023).

4. Retrieval from sparse Fourier data, eclipse maps, and phase curves

A second family of methods recovers regional spectra from sparse measurements in transform space rather than from directly imaged scenes. RIS for STIX is the most explicit count-space example. It first reconstructs a trigger image at one count-energy channel, uses its Fourier transform as the scale function in Variably Scaled Kernel interpolation of visibilities at the adjacent channel, and then reconstructs the next count map from the interpolated visibility surface. Repeating this sequentially over contiguous channels imposes a smoothness constraint across energy without deconvolving the bremsstrahlung cross-section and without constructing electron-flux maps. Applied to the November 11, 2022 flare in the 01:30:00–01:32:00 UT window, using nine 2-keV channels from 4–6 keV to 20–22 keV and one 3-keV channel at 22–25 keV, the method produced maps whose morphology “smoothly evolves from one energy channel to the contiguous one,” and the local count spectra extracted at selected points were numerically more stable with systematically smaller error bars than spectra from independently reconstructed MEM_GE maps (Volpara et al., 2024). The method retrieves count spectra, not electron spectra; that distinction is central.

Secondary-eclipse mapping uses a different route. “Eigenspectra” first infers monochromatic dayside maps from spectroscopic eclipse light curves using eigencurves, then clusters map pixels with similar spectra using K-means, and finally averages spectra within each cluster to obtain a small number of regional eigenspectra. The framework is explicitly a preprocessing and dimensionality-reduction step rather than a full atmospheric retrieval. In simulated JWST/NIRCam-like data, it can isolate broad hot versus cool or compositionally distinct regions, but it struggles to identify sharp edges and can generate spurious structure in asymmetric cases because only a few low-order eigencurves are supported (Mansfield et al., 2020).

Thermal phase-curve inversion addresses a related but more restricted geometry. ReSpect expands the observed phase curve at each wavelength as a Fourier series and converts the phase-dependent spectra into longitudinally resolved spectra that can then be retrieved with standard 1D tools. The paper emphasizes that a disk-integrated phase spectrum is a hemispherically integrated mixture of many longitudes, so a 1D retrieval on that spectrum is formally inconsistent when strong longitudinal heterogeneity is present. On simulated JWST observations of a longitudinally varying WASP-43b atmosphere, the average absolute Z-scores for retrieved $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$1, CO, $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$2, and $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$3 improved from $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$4, $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$5, $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$6, and $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$7 for disk-integrated retrievals to $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$8, $${\bf A}\cdot{\bf s}\equiv{\bf J}^T{\bf J}\cdot{\bf s}={\bf J}^T{\bf d}\equiv{\bf b}.$$9, $X\in\mathbb{R}^{I\times J\times K}$0, and $X\in\mathbb{R}^{I\times J\times K}$1 for longitudinally resolved retrievals. For extant Hubble and Spitzer observations, by contrast, the difference between the two approaches was negligible; the method’s practical relevance therefore depends strongly on data quality (Cubillos et al., 2021).

5. Multidimensional atmospheric retrievals

In exoplanet and brown-dwarf atmospheres, regional spectral retrieval is often implemented by embedding spatial structure directly in the atmospheric model. AURA-3D treats transmission spectra as integrals through a 3D atmosphere parameterized in spherical coordinates $X\in\mathbb{R}^{I\times J\times K}$2, with three anchor pressure–temperature profiles on the dayside edge of the terminator, the terminator center, and the nightside edge. Its main retrieval demonstrations are effectively 2D in $X\in\mathbb{R}^{I\times J\times K}$3–$X\in\mathbb{R}^{I\times J\times K}$4, with optional $X\in\mathbb{R}^{I\times J\times K}$5-dependent structure for clouds or future extensions. On a synthetic JWST spectrum generated from a 3D GCM, a traditional 1D retrieval returned abundance estimates that were all inconsistent with the true values at $X\in\mathbb{R}^{I\times J\times K}$6 and, except for HCN, more than $X\in\mathbb{R}^{I\times J\times K}$7 away, whereas the retrieval including a day–night temperature gradient recovered all major abundances within $X\in\mathbb{R}^{I\times J\times K}$8. The paper further reports that mean 1D–3D spectral differences exceed 50 ppm for many hot Jupiters with photospheric terminator temperatures $X\in\mathbb{R}^{I\times J\times K}$9 K and day–night contrasts $Y\in\mathbb{R}^{I\times J\times L}$0 K, indicating that JWST-quality transmission spectra can require explicit regional thermal structure (Nixon et al., 2022).

POSEIDON generalizes the same idea within an open-source Bayesian retrieval framework for 1D, 2D, and 3D transmission spectra. The public release officially supports modelling and retrieval of exoplanet transmission spectra in 1D, 2D, and 3D and is motivated by inhomogeneous temperatures, compositions, and cloud properties such as day–night structure, terminator asymmetries, and patchy clouds. The JOSS paper is explicit, however, that multidimensional retrieval support in version 1.0 is for transmission spectra only; thermal emission is a beta capability limited to cloud-free, 1D atmospheres with no scattering (MacDonald, 2024). This is a useful corrective to the assumption that all multidimensional retrieval frameworks are already general across observing modes.

A more constrained but observationally powerful example is the simultaneous retrieval of four phase-resolved JWST/MIRI emission spectra of WASP-43b. The spectra at orbital phases $Y\in\mathbb{R}^{I\times J\times L}$1, $Y\in\mathbb{R}^{I\times J\times L}$2, $Y\in\mathbb{R}^{I\times J\times L}$3, and $Y\in\mathbb{R}^{I\times J\times L}$4 are retrieved together with a longitude-dependent 2D thermal structure and a chemically homogeneous atmosphere over the observed pressure region. The regional information is therefore thermal, not chemical. This shared-chemistry, shared-geometry approach yielded statistically significant evidence of $Y\in\mathbb{R}^{I\times J\times L}$5 at $Y\in\mathbb{R}^{I\times J\times L}$6, $Y\in\mathbb{R}^{I\times J\times L}$7 at $Y\in\mathbb{R}^{I\times J\times L}$8, CO at $Y\in\mathbb{R}^{I\times J\times L}$9, weak evidence of $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$0 at $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$1, and a non-detection of $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$2, while also retrieving a moderate eastward hotspot offset of $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$3 degrees and a day–night temperature contrast of at least 700 K across the pressures shown (Yang et al., 2024). The method does not retrieve separate dayside, nightside, or terminator compositions, so its regionality is deliberately partial.

Time-domain heterogeneity retrieval makes the same distinction even more explicit. Tempawral does not recover a geographic map; instead it retrieves a baseline atmosphere from the time-averaged spectrum and then infers sinusoidal temporal variations in atmospheric parameters whose modulation implies unresolved regions with different properties. PCA is performed on the spectral time series, and the retrieval is carried out in eigen-spectrum space. Applied to JWST/NIRISS-SOSS observations of the variable T2.5 brown dwarf SIMP 0136, the preferred explanation involved a thermal perturbation of about 300 K near 1 bar, variations in the abundances of $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$4, CO, and FeH, and changes in the thickness of the iron cloud deck. The paper states directly that the output is regionally informative but not geographically unique (Wang et al., 22 Feb 2026).

6. Latitudinal strips, center-to-limb atlases, and component mixtures

Not all regional retrievals are map-like. In Uranus spectroscopy, the region is a latitude bin along a slit. Near-infrared H and K band spectra obtained with the Hale Telescope and the PHARO adaptive-optics camera were aligned roughly along the central meridian, binned to an effective resolution of about $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$5, and then inverted latitude by latitude with a custom non-linear constrained retrieval. Each latitude yields a 1D vertical aerosol profile, and in some cases a simplified methane parameter, but there is no joint 2D inversion across latitude. The retrieval consistently required two aerosol layers: a thin upper haze near the 100-mb tropopause and a lower cloud around $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$6 bar. The preferred solutions supported deep methane depletion toward high latitudes and a drop by more than a factor of 3 in the integrated scattering optical depth of the deeper layer at high southern latitudes from 2001 to 2007 (Roman et al., 2017). Here, “regionally resolved” means a meridional cross-section built from independent local column inversions.

Solar and stellar work shows a different variant: component-fraction inference. SOLIS reconciles disk-resolved and disk-integrated Ca II 854.21 nm spectra of the Sun by modeling the Sun-as-a-star line profile as an integral over local feature-class spectra with position-dependent Doppler shifts and limb darkening. The retrieval demonstration is intentionally modest: two classes only, quiet Sun and plage, with the disk-integrated line-core intensity modeled as

$Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$7

where $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$8 is the fractional area of the disk covered by class $Y=\mathbb{A}(X)+E,\qquad \mathbb{A}(X)=\mathbf{A}\,_3 X,$9. What is retrieved is the plage filling factor, not a latitude–longitude map or separate regional spectra from unresolved data alone (Pevtsov et al., 2014). This establishes that component-level regional inference can be spectrally meaningful even when exact spatial reconstruction is impossible.

High-resolution reference atlases serve as empirical priors for such work. The IAG center-to-limb solar atlas provides 165 quiet-Sun spectra from 4200 Å to 8000 Å at $\mathbf{A}\in\mathbb{R}^{L\times K}$0, corresponding to $\mathbf{A}\in\mathbb{R}^{L\times K}$1 at $\mathbf{A}\in\mathbb{R}^{L\times K}$2 Å, sampled at 14 heliocentric positions from $\mathbf{A}\in\mathbb{R}^{L\times K}$3 to $\mathbf{A}\in\mathbb{R}^{L\times K}$4. Each spectrum averages a $\mathbf{A}\in\mathbb{R}^{L\times K}$5 patch and multiple azimuths to suppress local-flow biases. The atlas shows explicitly that line depth decreases, line width increases, convective blueshift weakens, and line asymmetries change toward the limb, so a single disk-center template is inadequate for regionally aware spectral synthesis (Ellwarth et al., 2023). Such atlases do not themselves solve a retrieval problem, but they provide the empirical basis functions needed when regional spectra are assigned to annuli or occulted patches in Sun-as-a-star or transit calculations.

7. Assumptions, failure modes, and calibration prerequisites

Across domains, the main limitations arise from conditioning, prior structure, and calibration. Solar slit restoration assumes that the slit-jaw and spectral cameras share the same atmospheric realization and optical degradation, that exposures are short enough for a per-frame PSF to remain meaningful, and that scan motion and alignment can be reconstructed accurately. The implementation is sensitive to PSF errors, especially in the wings; interpolation of PSFs patch by patch can blur the effective kernel; incorrect slit-width modeling causes vertical stripes; and without the flatfield-like intensity renormalization of both $\mathbf{A}\in\mathbb{R}^{L\times K}$6 and the operator, strong striping appears in the reconstructions (Noort, 2017).

Joint hyperspectral reconstruction depends just as strongly on modeling assumptions. The interferometric HSI method assumes an accurate transmittance matrix $\mathbf{A}\in\mathbb{R}^{L\times K}$7, additive white Gaussian noise, and a scene structure well matched by DCT sparsity plus total variation. Its experiments are simulated, the spatial prior is especially favorable for patch-based scenes, hyperparameter tuning was manual, and the method does not explicitly enforce region segmentation beyond local TV regularity (Picone et al., 2023). RIS inherits the limitations of count-space imaging spectroscopy: it regularizes only count maps across energy, not electron-flux maps, and its stability depends on contiguous energy bins, sparse Fourier coverage, and SNR that becomes poor at high energies (Volpara et al., 2024).

Multidimensional atmospheric retrieval adds a separate layer of identifiability issues. AURA-3D’s demonstrated retrievals are mainly $\mathbf{A}\in\mathbb{R}^{L\times K}$8–$\mathbf{A}\in\mathbb{R}^{L\times K}$9 rather than fully arbitrary 3D inversions, and inhomogeneous chemistry was forward-modeled but not retrieved because of likely degeneracies (Nixon et al., 2022). POSEIDON’s public multidimensional support is mature for transmission but not yet general across emission or phase-curve use cases (MacDonald, 2024). The simultaneous WASP-43b emission retrieval assumes chemically homogeneous gas abundances over the observed atmosphere and neglects clouds, so its regional interpretation is intentionally limited to thermal structure (Yang et al., 2024). Tempawral further assumes one sinusoid per variable parameter, depends on the baseline mean-spectrum retrieval, and does not yield a unique geography; the authors note that phase shifts are hard to recover and that residuals in the $$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$0–$$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$1 water band may reflect multiple asynchronous water-bearing features beyond the model’s single-mode parameterization (Wang et al., 22 Feb 2026).

Calibration can itself determine whether regional products are trustworthy. Relative SRF retrieval for hyperspectral atmospheric spectrometers models one instrument as $$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$2 and estimates a robust relative SRF matrix

$$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$3

explicitly incorporating collocation mismatch and radiometric noise. The paper’s central claim is that even when the exact $$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$4 is not uniquely recovered, the harmonized spectra $$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$5 reproduce the covariance structure of $$a_{lk}=\frac{(1-\mathcal{R})^2}{1+\mathcal{R}^2-2\mathcal{R}\cos(2\pi\sigma_k\delta_l)}.$$6 to second order in the mismatch under favorable conditions, and a practical residual criterion is that the normalized difference should generally remain below about 10% (Dussarrat et al., 28 Apr 2025). A plausible implication is that regionally resolved geophysical retrievals are only as reliable as the cross-instrument spectral harmonization that precedes them.

Taken together, these works show that regionally-resolved spectra retrieval is not one method but a family of inverse strategies. Some reconstruct local spectra by explicitly undoing spatial mixing; some regularize a full cube and then permit post hoc regional extraction; some infer low-dimensional regional atmospheres from viewing geometry; some recover only component fractions or empirical basis spectra. The shared technical lesson is that localized spectral interpretation becomes feasible when the mixing operator—optical, geometric, temporal, or instrumental—is modeled well enough, and when the available diversity in scans, phases, wavelengths, or collocations is sufficient to constrain the inverse problem.