Compressive Spectral Imaging Overview

• Compressive Spectral Imaging is a technique that captures a 3D spatial–spectral datacube from a few multiplexed 2D measurements, bypassing conventional scanning methods.
• It employs a linear measurement model combined with diverse optical architectures—like CASSI and reflective systems—to enable efficient sensing in varied applications.
• Advanced reconstruction methods, ranging from convex optimization with hand-crafted priors to deep unfolding networks, address the severe inverse problem inherent in CSI.

Compressive spectral imaging (CSI) is a sensing and inversion framework in which a three-dimensional spatial–spectral datacube is acquired from one or a few multiplexed two-dimensional measurements rather than by conventional Nyquist sampling or line-by-line spectral scanning. A prototypical implementation is coded aperture snapshot spectral imaging (CASSI), in which a coded aperture modulates the scene, a dispersive element shears the modulated images as a function of wavelength, and a detector integrates the result. Closely related systems replace the prism or grating with diffractive or refractive chromatic optics, use reflective layouts for broadband near-infrared acquisition, combine dual-resolution arms for fusion, or replace array detectors with single-pixel sensing and spectrometric readout (Zhang et al., 2019, Kar et al., 2019, Li et al., 20 Aug 2025, Liu et al., 2015).

1. Forward models and inverse formulation

The common mathematical structure of CSI is a linear measurement model in which the unknown datacube is mapped to compressive measurements by an operator determined by mask modulation, wavelength-dependent shifts or blurs, and sensor integration. Several papers express this in the compact form

$y=\Phi x+n,$

or equivalently

$y=Hf+\omega,$

with $x$ or $f$ the vectorized hyperspectral cube and $\Phi$ or $H$ the calibrated sensing operator (Li et al., 20 Aug 2025, Zhang et al., 2019).

In discrete CASSI, when the coded aperture resolution equals the detector resolution and dispersion shears along the $x$-direction, the detector measurement for shot $k$ is written as

$$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$

where the index $j+l$ accounts for lateral shear. Vectorization yields $y=Hf+\omega,$0, and stacking $y=Hf+\omega,$1 shots gives $y=Hf+\omega,$2. In the same framework, sparse reconstruction is often imposed through a separable basis such as 2D Symmlet-8 wavelets in space and 1D DCT in spectrum, so that $y=Hf+\omega,$3 and $y=Hf+\omega,$4 with $y=Hf+\omega,$5 (Zhang et al., 2019).

Single-pixel and dual-compressed formulations exhibit the same linear structure but with separable sensing. In the dual-domain system using two DMDs and a PMT, the voxel-wise model is

$y=Hf+\omega,$6

and, after stacking measurements,

$y=Hf+\omega,$7

This separability motivates two-stage reconstructions in which spectral recovery is followed by spatial recovery (Liu et al., 2015).

Broadband reflective CASSI in the near infrared keeps the same structure but incorporates the double pass of a prism. The paper formulates

$y=Hf+\omega,$8

followed by

$y=Hf+\omega,$9

and its discrete version $x$0 for each spectral segment (Li et al., 20 Aug 2025).

These models make CSI a severely ill-posed inverse problem. Recovery therefore depends on the conditioning of the sensing matrix and on priors such as sparsity, total variation, low rank, or learned proximal maps. This shared structure also explains why optical design and reconstruction design are tightly coupled across the literature.

2. Optical architectures and sensing strategies

CSI encompasses several optical families that differ mainly in how wavelength diversity is generated and how measurements are collected.

Architecture	Optical mechanism	Representative papers
CASSI	coded aperture, disperser, 2D detector	(Zhang et al., 2019, Wang et al., 2022)
Reflective CASSI	single prism used twice, reflective coded aperture, beam splitter, SWIR FPA	(Li et al., 20 Aug 2025)
Diffractive or chromatic-aberration CSI	coded aperture plus diffractive lens or refractive lens with wavelength-dependent blur	(Kar et al., 2019, Thomas et al., 14 Aug 2025)
Single-pixel CSI	DMD-based spatial or spectral modulation with PMT, spectrometer, or beam-splitting detection	(Liu et al., 2015, Wang et al., 2017, Song et al., 2 Apr 2026)
Dual-resolution fusion systems	low-spatial/high-spectral CASSI arm plus high-spatial/low-spectral arm	(Jacome et al., 2022, Ramirez et al., 2020)
Polarimetric CSI	coded aperture with DAP, QWP, and Wollaston prism for joint spectral and circular-polarization encoding	(Ning et al., 2020)

Within CASSI, mask design itself has become a central research topic. A notable example replaces conventional square binary masks with binary hexagonal coded apertures. Because the detector remains square-sampled, the misalignment between hexagonal mask elements and detector pixels induces an equivalent grayscale modulation through area-weighted overlap. The resulting effective transmission increases the degrees of freedom of the sensing matrix, yields more uniformly distributed nonzeros in $x$1, and improves RIP behavior. The same analysis shows that the optimal hexagonal masks follow a blue noise distribution on the hex lattice and that, in multi-shot operation, masks should be complementary across shots. The paper further derives the ordering

$x$2

for random square, blue-noise square, and blue-noise hexagonal masks, respectively, and reports best reconstruction PSNRs for lateral offset ratio $x$3 at both 50% and 25% transmittance (Zhang et al., 2019).

Optical compactness has been pursued by replacing conventional disperser assemblies with chromatic imaging elements. Diffractive-lens CSI uses a coded aperture followed by a photon sieve and measures at a few planes with a monochrome detector. Its wavelength-dependent point spread functions define a shift-invariant forward model

$x$4

and simulations show PSNR values from $x$5 dB to $x$6 dB as the number of measurements increases from $x$7 to $x$8 under input SNRs from $x$9 dB to $f$0 dB (Kar et al., 2019). A related Earth-observation study replaces the diffractive lens with a classical refractive lens whose chromatic aberration produces wavelength-dependent defocus. There the compressive ratio is $f$1, and with $f$2 bands and $f$3 measurements the reported ratio is about $f$4, with RGB PSNR $f$5 dB in noiseless simulation (Thomas et al., 14 Aug 2025).

Broadband operation has also driven architecture changes. A reflective near-infrared system extends R-CASSI to $f$6–$f$7 nm by segmenting the spectrum into $f$8–$f$9 nm and $\Phi$0–$\Phi$1 nm, swapping prism and beam-splitter components optimized for each segment, and concatenating the reconstructed cubes after spatial registration. The system reports $\Phi$2 spectral channels, average channel spacing $\Phi$3 nm, and calibration error below the nominal channel spacing (Li et al., 20 Aug 2025).

3. Reconstruction algorithms and prior models

Early CSI reconstruction emphasized convex optimization with hand-crafted priors. In CASSI, basis-pursuit denoising with sparsifying transforms is commonly written as

$\Phi$4

or, in penalized form,

$\Phi$5

The hexagonal blue-noise study uses GPSR with spatial Symmlet-8 wavelets and spectral DCT, while broadband reflective CASSI employs TwIST with TV regularization and reports that TwIST provided superior reconstructions to GAP-TV on the experimental data (Zhang et al., 2019, Li et al., 20 Aug 2025).

Total-variation-driven and operator-splitting methods remain important in architectures with blur-based sensing. The refractive-lens Earth-observation formulation reconstructs in a separable 3D basis by basis pursuit solved with Douglas–Rachford splitting, while the progressive pushbroom architecture senses spectral rows and reconstructs them iteratively with TV on the $\Phi$6–$\Phi$7 plane, exploiting along-track correlation through predictors from neighboring rows (Thomas et al., 14 Aug 2025, Kuiteing et al., 2014). Convolutional sparse coding has also been adapted to CSI by representing high-frequency structure as the convolution sum of filters and coefficient maps, constraining the coefficients with the $\Phi$8 norm to exploit spectral correlation, and adding a global TV term for low-frequency estimation; this paper reports improvements of up to $\Phi$9 dB in PSNR and $H$0 in SSIM over mainstream optimization methods (Wang et al., 2022).

A second line of work keeps the physical forward model but replaces explicit regularizers with learned or training-free priors. A training-free deep prior framework embeds a low-rank Tucker tensor in the first layer and fits both the generator weights and Tucker factors directly to the measurements. On simulated CASSI with $H$1 bands and SNR $H$2 dB, the reported variants achieve around $H$3 dB PSNR and SSIM around $H$4–$H$5, while real-data tests report spectral angle mapper values of $H$6 and $H$7 in binary and colored coded-aperture settings, respectively (Bacca et al., 2021). A related MAP framework learns a Gaussian scale mixture prior through a DCNN that predicts both local means and scale weights, and reports average PSNR $H$8 dB and SSIM $H$9 on simulated CASSI, surpassing TSA-Net and DNU on the same setup (Huang et al., 2021).

Deep unfolding has become especially prominent. GAP-CCoT embeds a hybrid convolution-and-contextual-transformer denoiser inside generalized alternating projection and reports average PSNR $x$0 dB and SSIM $x$1 on simulated benchmarks (Wang et al., 2022). PGDUDST inserts a Dense-spatial Spectral-attention Transformer into a proximal-gradient unfolding framework and reports average PSNR $x$2 dB and SSIM $x$3, while requiring only $x$4 of the training time of RDLUF-MixS$x$5-9stg to achieve comparable results (Chen et al., 2023). CIDNet introduces chromaticity–intensity decomposition in a dual-camera CASSI system and reports average PSNR $x$6 dB and SSIM $x$7 on KAIST simulation, together with chromaticity PSNR $x$8 dB and SSIM $x$9 (Wang et al., 20 Sep 2025). Phy-CoSF extends unfolding to continuous spectral reconstruction and spectral super-resolution; on unseen wavelengths it reports SAM $k$0, PSNR $k$1 dB, and SSIM $k$2, while also achieving PSNR $k$3 dB and SSIM $k$4 on discrete reconstruction (Chen et al., 13 May 2026).

Deployment constraints have motivated algorithmic compression as well. BiSRNet binarizes a spectral-redistribution network for snapshot compressive imaging and reports average PSNR/SSIM $k$5 dB / $k$6 across KAIST scenes, with $k$7 K parameters versus $k$8 K for $k$9-Net and $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$0 G operations versus $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$1 G for $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$2-Net (Cai et al., 2023). This suggests that CSI reconstruction is no longer defined only by accuracy; memory footprint, operator count, and hardware compatibility have become part of the reconstruction problem itself.

4. Compressed-domain analysis and task-oriented CSI

CSI is not restricted to full-cube reconstruction. Several studies treat the compressive measurements, or deliberately reduced spectral features, as sufficient representations for downstream tasks.

One direction performs inference directly in the compressive domain. Spatially regularized sparse subspace clustering on CASSI measurements assumes that compressed signatures lie in a union of low-dimensional subspaces and augments SSC with a $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$3 spatial regularizer on the coefficient cube. On Indian Pines, the compressed-domain method with optimized codes reports OA $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$4 versus $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$5 for random-coded compressed measurements, while reducing runtime to $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$6 s compared with $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$7 s for full-data spatial SSC and $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$8 s for full-data SSC (Zhu et al., 2019).

A related line fuses features directly from dual-resolution compressive measurements instead of first reconstructing an HSI cube. In a dual-arm 3D-CASSI formulation, the fused high-spatial-resolution low-dimensional feature bands are estimated by solving an inverse problem with sparsity and TV regularization. On Pavia University, the reported classification pipeline with MLPNN achieves OA $$Y^k_{i,j}=\sum_{l=1}^{L} F_{i,j+l,l}T^k_{i,j+l}+\omega^k_{i,j},$$9, AA $j+l$0, and $j+l$1; on Indian Pines it reports OA $j+l$2 (Ramirez et al., 2020). End-to-end sensing-and-reconstruction co-design has been pushed further in D$j+l$3UF, which jointly optimizes the CASSI coded aperture, the MCFA colored coded aperture, and an ADMM-inspired unrolling network, reporting PSNR $j+l$4 dB, SSIM $j+l$5, and SAM $j+l$6 in its best ICVL configuration (Jacome et al., 2022).

Task-oriented sensing can also eliminate the explicit datacube. HyPIS maps each pixel’s spectrum to a two-dimensional phasor through sine and cosine optical encoders and reconstructs only the phasor images with single-pixel detection. The reported system reduces data volume by up to two orders of magnitude, reduces stored data to $j+l$7 of raw hyperspectral data on CAVE simulations, and demonstrates about $j+l$8 fps at $j+l$9 resolution over $y=Hf+\omega,$00 frames, while remaining robust under low light and uneven illumination (Song et al., 2 Apr 2026). This suggests a different endpoint for CSI: not spectral reconstruction per se, but hardware-level generation of features sufficient for classification or recognition.

Single-pixel dual compressed sensing reveals a different lesson. In the PMT-based two-DMD system, increasing spectral modulations improves spectral reconstruction and suppresses ghost images, whereas increasing spatial modulations primarily reduces image noise. The reported experiment uses $y=Hf+\omega,$01 spatial modulations and $y=Hf+\omega,$02 spectral modulations on a $y=Hf+\omega,$03 cube, producing $y=Hf+\omega,$04 scalar measurements, and explicitly concludes that sufficient spectral modulation is critical to avoid ghost images (Liu et al., 2015).

5. Applications, operating regimes, and domain-specific extensions

A major application driver is broadband near-infrared imaging. The reflective NIR system covering $y=Hf+\omega,$05–$y=Hf+\omega,$06 nm demonstrates spectral accuracy on letter targets illuminated at $y=Hf+\omega,$07 nm, $y=Hf+\omega,$08 nm, $y=Hf+\omega,$09 nm, and $y=Hf+\omega,$10 nm, and reports correlations $y=Hf+\omega,$11 and $y=Hf+\omega,$12 between reconstructed and spectrometer ground-truth spectra for a real and a fake apple, respectively. It also notes a decline for the real apple between $y=Hf+\omega,$13–$y=Hf+\omega,$14 nm and lower intensity above $y=Hf+\omega,$15 nm, enabling authenticity discrimination (Li et al., 20 Aug 2025).

Earth observation motivates compact, low-mass systems and lower downlink burden. The refractive-lens chromatic-aberration architecture is explicitly framed as an alternative to diffractive-lens CSI for spaceborne instruments. Its simulations on a 29-band CAVE scene with three coded measurements report RGB PSNR $y=Hf+\omega,$16 dB at a compressive ratio of about $y=Hf+\omega,$17, illustrating the feasibility of compact chromatic-aberration-based CSI for EO payloads (Thomas et al., 14 Aug 2025). The older progressive TV architecture is likewise motivated by remote acquisition and pushbroom onboard sensors, emphasizing that sensors progressively acquire spectral rows rather than spectral channels (Kuiteing et al., 2014).

Biomedical, fluorescence, and low-light applications motivate different optical trade-offs. The compressive fluorescence spectral imaging system replaces mechanical scanning with DMD-based spatial multiplexing and a fiber spectrometer, reporting spectral resolution $y=Hf+\omega,$18 nm and approximately $y=Hf+\omega,$19 light energy collection from the object per measurement. It also observes that image quality improves markedly above about $y=Hf+\omega,$20 sampling rate (Wang et al., 2017). The circular-polarization snapshot spectral imager extends CSI to a four-dimensional datacube $y=Hf+\omega,$21, where $y=Hf+\omega,$22, and reports PSNR $y=Hf+\omega,$23 dB for all reconstructions, SSIM $y=Hf+\omega,$24 when input SNR $y=Hf+\omega,$25 dB, and improved delineation in DoCP and AoCP images for low-light or scattering scenarios such as defogging and underwater imaging (Ning et al., 2020).

These examples show that CSI is best understood not as a single camera architecture but as a family of coded, multiplexed measurement systems adapted to different spectral ranges, detector constraints, and task definitions.

6. Limitations, controversies, and current directions

Several recurring trade-offs structure the field. First, optical simplicity often shifts burden to calibration and inversion. Many models assume linear dispersion, integer-pixel shear, shift-invariant PSFs, or exact overlap geometry; the hexagonal-mask analysis explicitly notes that linear dispersion and integer-pixel shear are assumed in the discrete model, and the refractive-lens EO system notes that space-invariant bandwise PSFs approximate field-dependent effects (Zhang et al., 2019, Thomas et al., 14 Aug 2025). Broadband reflective CSI likewise emphasizes the need for wavelength-to-pixel shift calibration and careful alignment, with explicit note of calibration complexity and sub-band switching (Li et al., 20 Aug 2025).

Second, “snapshot” is not universal across CSI. CASSI and several of its variants are single-exposure systems, but dual-DMD single-pixel schemes, fluorescence systems with repeated random patterns, and progressive pushbroom architectures remain sequential. This suggests that CSI should be defined by compressive measurement design rather than by snapshot operation alone (Liu et al., 2015, Wang et al., 2017, Kuiteing et al., 2014).

Third, task-oriented compression changes the meaning of spectral fidelity. HyPIS shows that many classification tasks can bypass full 3D recovery, but it also states that quantitative spectral analysis requiring fine spectral lines or recovery of absolute spectra often still needs full hyperspectral reconstruction. Similarly, CIDNet requires a dual-camera setup and accurate registration to exploit chromaticity–intensity decomposition (Song et al., 2 Apr 2026, Wang et al., 20 Sep 2025).

Current directions therefore combine optical co-design, stronger priors, and broader spectral parameterizations. Joint optimization of sensing architectures and reconstruction networks appears in compressive spectral image fusion (Jacome et al., 2022). Continuous spectral fields extend reconstruction from discrete bands to arbitrary target wavelengths (Chen et al., 13 May 2026). Hexagonal blue-noise apertures suggest that binary fabrication can still induce effective grayscale modulation through geometry rather than through true grayscale masks (Zhang et al., 2019). Broadband NIR work explicitly identifies streamlined sub-band switching, automated alignment, flatter spectral efficiency, and data-driven reconstruction as future directions, particularly for compact or UAV-scale platforms (Li et al., 20 Aug 2025).

Across these developments, the central problem remains stable: designing measurement operators whose optical multiplexing preserves enough spatial–spectral structure that inversion, inference, or both remain reliable under extreme dimensional compression.