Spatial-Spectral Chromatic Coding

• Spatial-spectral chromatic coding is a family of techniques that integrates spectral information with spatial structure to enable advanced imaging and data analysis.
• These methods employ physical encoding schemes, computational inversion, and deep learning to achieve precise hyperspectral reconstruction and robust visual diagnostics.
• The approach finds applications in hyperspectral imaging, SAR visualization, and computer vision, offering improved resolution, reduced ill-posedness, and efficient data interpretation.

Spatial-spectral chromatic coding denotes a family of representations in which chromatic or spectral information is explicitly coupled to spatial structure. In optical and computational imaging, the coupling is usually physical: wavelength is transformed into spatially varying speckle patterns, diffraction patterns, coded-aperture modulations, or defocus signatures that can later be inverted into a multispectral or hyperspectral estimate (French et al., 2017, Wang et al., 2017, Salazar et al., 2018, Aydin et al., 26 Mar 2025). In learning-based hyperspectral reconstruction, the coupling is feature-theoretic: RGB channels are independently embedded, cross-channel spatio-spectral dependencies are combinatorially queried, or illumination is separated from chromaticity before reconstruction (Yang et al., 2023, Wang et al., 20 Sep 2025). In other literatures, the term extends to formal spatial coding schemes whose particles receive unique chromatic codes, to chromatic visualization of SAR interference, and to joint spatial-chromatic descriptors for image analysis (Zhu, 2016, Yang et al., 10 Sep 2025, Bandara et al., 2018, Shoeiby et al., 2019). Across these meanings, the unifying idea is that spatial arrangement and spectral content are not treated as separable afterthoughts, but as mutually constraining signals.

1. Conceptual scope and recurring design pattern

A recurring pattern in spatial-spectral chromatic coding is a two-stage pipeline. First, a system constructs a code in which spectral variation changes a spatial response, or spatial context changes a chromatic representation. Second, an inversion, decoding, or interpretation stage uses that coupling for reconstruction, recognition, topology, or visual diagnosis.

Domain	Coding object	Representative mechanism
Hyperspectral imaging	Wavelength or spectrum	Speckle patterns, diffractive PSFs, coded apertures, chromatic blur
Learning-based reconstruction	Cross-channel cues	Independent channel modeling, attention-based fusion, chromaticity-intensity decomposition
Interferometry and SAR	Interference signatures	Color erasure detection, spectral subband decomposition with RGB/HSV coding
Spatial data models and vision descriptors	Spatial particles or local patterns	Chromatic codes, spatial-chromatic histograms, learned chromatic mapping

A common misconception is to equate chromatic coding with simple false-color display. In several systems the color or spectral code is part of the forward sensing model itself, as in $y = T x$, $\mathbf{I} = \mathbf{A}\mathbf{S}$, $y = C H x$, or the SSCSI sensing equation $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$ (French et al., 2017, Wang et al., 2017, Aydin et al., 26 Mar 2025, Salazar et al., 2018). In other systems the code is explicitly designed for interpretation rather than inversion, as in SAR chromatic visualization or full-OACD spatial tessellations (Yang et al., 10 Sep 2025, Zhu, 2016).

2. Physical encoding in optical and compressive spectral imaging

In speckle-based hyperspectral imaging with disordered nanowire mats, spectral information is encoded into wavelength-dependent spatial speckle fingerprints. Strong, diffuse light scattering in a semiconductor nanowire mat transforms different wavelengths into distinct speckle patterns with nanometer sensitivity, and the reconstruction is posed through a spectral intensity transmission matrix $T$ satisfying $y = T x$ (French et al., 2017). Tikhonov regularization minimizes $\| T x - y \|^2 + \lambda \| x \|^2$, while compressive sensing uses $l_1$-minimization, $\min_x \|x\|_1$ subject to $T x = y$ or $\mathbf{I} = \mathbf{A}\mathbf{S}$0. The reported spectral correlation width is $\mathbf{I} = \mathbf{A}\mathbf{S}$1 at $\mathbf{I} = \mathbf{A}\mathbf{S}$2, the nanowire mat is $\mathbf{I} = \mathbf{A}\mathbf{S}$3 thick with relatively high transmission ($\mathbf{I} = \mathbf{A}\mathbf{S}$4), and the method is best suited for scenarios with few spectral lines over a wide range (French et al., 2017). This establishes a canonical form of spatial-spectral chromatic coding: wavelength is not dispersed onto separate detector columns, but mapped into a high-dimensional spatial interference signature.

A related but distinct implementation uses a micro-structured diffractive filter mounted $\mathbf{I} = \mathbf{A}\mathbf{S}$5 mm above a monochrome CMOS image sensor. There, spectral information is converted to a spatial code because the PSF is both spatially and spectrally variant, and the calibrated forward model is $\mathbf{I} = \mathbf{A}\mathbf{S}$6, or in vectorized form $\mathbf{I} = \mathbf{A}\mathbf{S}$7 (Wang et al., 2017). Reconstruction uses Tikhonov regularization and an SVD solution with filter factor $\mathbf{I} = \mathbf{A}\mathbf{S}$8. The system experimentally demonstrated spectral resolution of $\mathbf{I} = \mathbf{A}\mathbf{S}$9 within the visible band ($y = C H x$0 to $y = C H x$1), spatial resolution enhanced by over $y = C H x$2 compared to the case without the diffractive filter, and software-controlled trade-offs between spectral resolution and field of view without hardware change as long as sufficient sensor pixels are utilized for information encoding (Wang et al., 2017). Because no absorptive color filters are utilized, sensitivity is preserved as well.

Spatial Spectral Compressive Spectral Imagers formalize chromatic coding through a binary coded aperture placed between disperser and detector. Their key parameter is the normalized mask position $y = C H x$3, with $y = C H x$4 at the detector and $y = C H x$5 at the spectral plane (Salazar et al., 2018). The attainable spectral resolution is

$y = C H x$6

and the number of resolvable bands is

$y = C H x$7

The paper terms the increase of spectral resolution by moving the mask toward the spectral plane “spectral zooming” (Salazar et al., 2018). This makes the spatial-spectral trade-off explicit: increasing $y = C H x$8 yields more unique chromatic modulation but can degrade spatial resolution.

“Spectrum from Defocus” exploits chromatic aberration rather than engineered dispersers or coded apertures. Two lenses and a grayscale sensor capture a chromatic focal stack in which each lens position $y = C H x$9 brings a different wavelength $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$0 into focus and leaves other wavelengths blurred (Aydin et al., 26 Mar 2025). The forward model is $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$1, where each block satisfies $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$2, and reconstruction solves

$$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$3

with Plug-and-Play ADMM. Reported performance includes PSNR $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$4, SSIM $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$5, SAM $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$6, compute time $$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$7s, and four optical components (“2 lenses + actuator + sensor”) (Aydin et al., 26 Mar 2025). The paper explicitly frames the method as passive spectral encoding via chromatic blur, emphasizing photon efficiency, optical simplicity, and physical modeling.

3. Learned spatio-spectral feature coding for hyperspectral reconstruction

In RGB-to-HSI reconstruction, spatial-spectral chromatic coding appears as a feature design principle rather than a hardware principle. CESST argues that existing methods often fuse RGB first and project “brutally” into hyperspectral space, losing channel-unique features and subtle cross-channel correlations, especially in marginalized spectral bands (Yang et al., 2023). Its alternative is explicit combinatorial modeling. Each RGB channel is treated as a separate “view” and passed through its own feature extraction block in a three-scale encoder-decoder architecture; only afterward are pairwise and triple-wise interactions queried in a combinatorial manner. The Spatio-Spectral Attention Block runs in parallel with Spatial-MSA and Spectral-MSA, where the spatial branch introduces a spatial shuffle mechanism and a depthwise convolution bridge, and the spectral branch attends over channels. The Spectrum-Fusion Attention Module then computes six pairwise branches such as

$$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$8

followed by

$$g(x, y) = \int_{\Lambda} T(x(1-s) + s\alpha\lambda, y)\, f(x, y, \lambda)\, d\lambda$$9

The loss combines $T$0, SSIM, and mean relative absolute error (MRAE), training uses Adam optimizer, $T$1 crops, data augmentation, and 300 epochs, and the paper reports that CESST outperforms all SOTA baselines on NTIRE2022 and ICVL in MRAE, RMSE, PSNR, and SAM while using fewer parameters (Yang et al., 2023). The stated qualitative effect is particularly improved long-wavelength fidelity.

A later formulation reframes spatial-spectral chromatic coding through chromaticity-intensity decomposition in dual-camera CASSI. The hyperspectral cube is separated into a spatially smooth intensity map and a spectrally variant chromaticity cube via

$T$2

with

$T$3

The measurement model is $T$4, and with known intensity simplifies to $T$5 (Wang et al., 20 Sep 2025). CIDNet uses a hybrid spatial-spectral Transformer with Swin Transformer blocks in the spatial branch and Top-K Spectral Attention in the spectral branch,

$T$6

inside an HQS-style unfolding architecture with a degradation-aware estimator and analytical data-fidelity step. The paper reports average HSI PSNR $T$7, SSIM $T$8, chromaticity PSNR $T$9, chromaticity SSIM $y = T x$0, and an ablation showing that using intensity in both model-based and learning-based methods brings a jump of $y = T x$1dB in reconstruction accuracy (Wang et al., 20 Sep 2025). This suggests that, in compressive systems, coding the illumination-dependent and illumination-invariant components separately can reduce ill-posedness before any deep prior is applied.

4. Interference, phase, and visual interpretation

Chromatic intensity interferometry introduces a different use of spatial-spectral coding: spatial separation is inferred from wavelength diversity after color erasure detection. The experimental system used two $y = T x$2 mm-aperture telescopes and a $y = T x$3 m baseline to resolve the distance between a $y = T x$4 nm source and a $y = T x$5 nm source separated by $y = T x$6 mm at a distance of $y = T x$7 km, surpassing the diffraction limit of a single telescope by about $y = T x$8 times (Liu et al., 2021). The phase is

$y = T x$9

and the measured second-order correlation is

$\| T x - y \|^2 + \lambda \| x \|^2$0

with $\| T x - y \|^2 + \lambda \| x \|^2$1 (Liu et al., 2021). Unlike conventional intensity interferometry, the chromatic formulation allows recovery of the phase of the Fourier transform of the imaged objects. Here the chromatic code is temporal and interferometric rather than image-plane colorimetric.

In SAR imagery, spatial-spectral chromatic coding is explicitly a visual-analytic tool. The method partitions the range frequency domain into $\| T x - y \|^2 + \lambda \| x \|^2$2 equal-width subbands with $\| T x - y \|^2 + \lambda \| x \|^2$3, reconstructs spatial-spectral images by inverse FFT, optionally multilooks them, then assigns each subband a basis color selected to span the hue circle uniformly in HSV space while satisfying a “sum-white constraint” (Yang et al., 10 Sep 2025). Per-pixel normalized amplitudes

$\| T x - y \|^2 + \lambda \| x \|^2$4

are multiplied by reference colors and averaged into a composite RGB image. Clean regions with flat spectra are rendered in achromatic grayscale; interference with non-uniform spectral energy appears as vivid color patches, smooth color gradients, or azimuthal color gradients depending on whether the signal is CW-NBI, pulsed LFM interference, or an unfocused echo (Yang et al., 10 Sep 2025). The method is therefore not a spectral reconstruction algorithm, but a mechanism for rapid visual interpretation, quality assessment, and data diagnosis.

5. Formal chromatic coding of space and topology

A non-optical but conceptually important branch of the literature defines chromatic coding as a formal code assigned to spatial particles. In the spatial chromatic model, full-coded ordinary arranged chromatic diagrams partition space into cells, edges, and vertices generated by all perpendicular bisectors $\| T x - y \|^2 + \lambda \| x \|^2$5 between $\| T x - y \|^2 + \lambda \| x \|^2$6 generator points (Zhu, 2016). Each particle receives a unique chromatic code

$\| T x - y \|^2 + \lambda \| x \|^2$7

where the entries are accumulated from half-plane and bisector assignments over all generator pairs. Cells have all integer and distinct $\| T x - y \|^2 + \lambda \| x \|^2$8, edges have two equal half-integers, 2-I vertices have four equal half-integers, and 3-I vertices have three identical integers (Zhu, 2016).

The framework also defines a chromatic base, chromatic distance $\| T x - y \|^2 + \lambda \| x \|^2$9 as Manhattan distance between two codes, and code distance $l_1$0 as Hamming distance. Topology is then recoverable directly in code space: two cells are adjacent iff their codes differ by exactly $l_1$1 at two entries, equivalently $l_1$2, and the chromatic code of an edge is the average of the codes of its two incident cells (Zhu, 2016). In this literature, “chromatic” does not refer to optical wavelength, but to a mathematically structured labeling of spatial relations. A plausible implication is that the broader term “spatial-spectral chromatic coding” spans both physical color encodings and abstract code systems in which structured tuples play the role of chromatic signatures.

6. Spatial-chromatic descriptors and chromatic mapping in computer vision

In image analysis, spatial-spectral chromatic coding often appears as a descriptor that jointly quantizes appearance and geometry. The locally salient dither pattern feature with a 2D spatial-chromatic histogram first extracts salient $l_1$3 dither patterns from block-averaged images, computes color differences to eight neighbors through

$l_1$4

keeps locally maximal patterns, and bins each retained pattern by its centroid distance and quantized color (Bandara et al., 2018). The centroid is

$l_1$5

and radial distances $l_1$6 determine the spatial bins. The paper reports that the optimal configuration uses $l_1$7 spatial bins $l_1$8 color levels, and that $l_1$9 of rotated/scaled images were classified to their correct class with mean SVM confidence $\min_x \|x\|_1$0 for both rotation and scaling (Bandara et al., 2018). The descriptor encodes how colors are distributed at different spatial zones relative to the object’s center.

A sensor-oriented variant appears in snapshot mosaic chromatic mapping. Instead of interpolating a sparse multispectral mosaic and then applying a color matching function, TSRCAN forms a zero-padded, uninterpolated $\min_x \|x\|_1$1-channel input in which each channel corresponds to one wavelength band of a $\min_x \|x\|_1$2 mosaic, preserving the on-chip spatial-wavelength arrangement (Shoeiby et al., 2019). The network computes

$\min_x \|x\|_1$3

then fuses them as

$\min_x \|x\|_1$4

The reported results show Bicubic+CMF at PSNR $\min_x \|x\|_1$5, SSIM $\min_x \|x\|_1$6; RCAN baseline at PSNR $\min_x \|x\|_1$7, SSIM $\min_x \|x\|_1$8; and TSRCAN at PSNR $\min_x \|x\|_1$9, SSIM $T x = y$0, with a dataset of $T x = y$1 registered stereo multi-spectral/RGB image pairs (Shoeiby et al., 2019). This is a learned chromatic code in which spatial undersampling and narrow, uneven spectral sampling are not corrected separately; they are jointly mapped into RGB by a residual channel attention network equipped with a texture sensitive block.

Spatial-spectral chromatic coding is therefore not a single algorithmic doctrine but a cross-domain principle. It includes physically invertible encodings of wavelength into spatial signatures, learned decompositions and cross-channel attentional embeddings for hyperspectral reconstruction, chromatic readouts of interference and artifacts, formal topological codes for spatial particles, and compact descriptors that bind local color content to spatial configuration. The principal differences across the literature concern what is being coded, whether the code is invertible or interpretive, and whether chromaticity denotes optical spectrum, learned feature dependence, or abstract symbolic structure. The principal commonality is that chromatic information acquires meaning only through its spatial organization, and spatial organization becomes more discriminative when represented through chromatic or spectral structure.