Spectral-Spatial Unmixing Fusion (SSUF)

• SSUF is a family of techniques that fuses spectral unmixing with spatial and cross-modal data to estimate high-resolution hyperspectral images, endmembers, and abundances.
• It encompasses methods from post-unmixing spatial allocation and LiDAR-guided regularization to dual-stream autoencoders and attention-based fusion, ensuring improved spatial consistency.
• Empirical results show SSUF enhances metrics such as RMSE, SAD, and PSNR, proving effective in edge preservation and robustness even in unregistered imaging scenarios.

Spectral-Spatial Unmixing Fusion (SSUF) denotes a class of hyperspectral fusion and unmixing formulations in which spectral mixture modeling is coupled with spatial, contextual, or cross-modal information to estimate endmembers, abundances, or a high-resolution hyperspectral image. In the supplied literature, SSUF appears both as an explicit module name—introduced as “Spectral-Spatial Unmixing Fusion (SSUF)” for hyperspectral single-image super-resolution (Muhammad et al., 26 Sep 2025)—and as a broader methodological pattern encompassing LiDAR-guided regularization, joint spatial-spectral cofactorization, abundance-subspace PAN-guided fusion, dual-stream spectral-spatial attention, collaborative autoencoder unmixing, and unregistered abundance-domain super-resolution (Uezato et al., 2017, Lagrange et al., 2019, Li et al., 2022, Wang et al., 2024, Qi et al., 2022, Zhang et al., 9 Mar 2026). Taken together, these works indicate that SSUF is best understood as a family of fusion mechanisms rather than a single standardized algorithm.

1. Historical formation and scope

The supplied literature traces SSUF from early unmixing-guided sharpening heuristics to coupled inverse problems, attention-based deep networks, and unregistered fusion. An early precursor used spectral unmixing outputs from hyperspectral imagery together with panchromatic segmentation to infer subpixel material arrangement (Rajabi et al., 2013). A more formal foundation then modeled the latent high-resolution image directly through endmembers and abundances and estimated both jointly from two sensors (Wei et al., 2016). Subsequent work expanded the fusion locus: external LiDAR could guide spatial regularization (Uezato et al., 2017); contextual observations could be factorized as a second modality and coupled to spectral abundances through cofactorization (Lagrange et al., 2019); deep networks could fuse spatial neighborhoods with spectral views or abundance-like latent codes (Qi et al., 2022, Wang et al., 2024, Li et al., 2022); and recent unregistered methods moved fusion into abundance-space translation or blind coupled inverse problems (Zhang et al., 9 Mar 2026, Song et al., 23 Mar 2026).

Formulation family	Representative paper	SSUF role
Unmixing-guided PAN allocation	(Rajabi et al., 2013)	uses unmixing outputs and PAN segmentation to spatially allocate endmembers
Joint fusion and unmixing	(Wei et al., 2016)	estimates fused image, endmembers, and abundances together
External-modality regularization	(Uezato et al., 2017)	injects LiDAR DSM into weighted spatial regularization
Joint spatial-spectral cofactorization	(Lagrange et al., 2019)	treats contextual features as a second observation model
Collaborative spectral-spatial autoencoding	(Qi et al., 2022)	couples spatial and spectral autoencoders through shared decoder and abundance consistency
Abundance-subspace hyperpansharpening	(Li et al., 2022)	injects PAN detail into a projected low-dimensional abundance subspace
Dual-stream or attention-based unmixing fusion	(Wang et al., 2024, Ratnayake et al., 2024)	fuses spatial context, spectral views, or endmember ensembles through attention
Tensor or implicit spectral-spatial fusion	(Peng et al., 2022, Wang et al., 2024)	strong architectural templates, but not direct classical unmixing
Unregistered abundance-domain fusion	(Zhang et al., 9 Mar 2026, Song et al., 23 Mar 2026)	extends abundance-based fusion beyond pixel-aligned settings

A useful way to delimit the field is to distinguish strict SSUF from adjacent spectral-spatial fusion. Some papers in the supplied corpus are described as directly relevant yet not direct spectral unmixing methods, notably U2Net and DTDNML (Peng et al., 2022, Wang et al., 2024). This suggests that SSUF sits inside a broader landscape of spectral-spatial fusion, with explicit unmixing at one end and implicit latent-factor or attention-based fusion at the other.

2. Mathematical structure and fusion loci

A recurrent mathematical core is the linear mixing model. In LiDAR-guided unmixing, each observed pixel satisfies

$$\mathbf{y}_i = \mathbf{E}\mathbf{a}_i + \mathbf{n}_i,$$

with abundance nonnegativity and sum-to-one constraints

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$

(Uezato et al., 2017). Joint estimation then replaces pixelwise fitting by

$$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$

subject to ANC and ASC, where fusion enters through the regularizer. In the LiDAR case, the critical modification is

$$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$

with $w_{ij}$ derived from hyperspectral similarity, principal components, abundances, DSM heights, or their combinations. The main SSUF mechanism is therefore guidance-aware spatial weighting rather than a new mixing law (Uezato et al., 2017).

A second fusion locus is the observation model itself. In matrix cofactorization, the spectral image and contextual features are both factorized: $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$ and the two latent codes are coupled by

$$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$

The full SP2U objective jointly penalizes spectral reconstruction, contextual reconstruction, coupling to shared clusters, and orthogonality-promoting clustering, with simplex constraints on $\mathbf A$, $\mathbf U$, and $\mathbf Z$ (Lagrange et al., 2019). Here SSUF is not a regularizer on abundances; it is a coupled factorization of two observation spaces.

A third locus is abundance-subspace fusion for pansharpening or super-resolution. Pgnet adopts the observation model

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$0

together with

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$1

then performs fusion in a projected low-dimensional abundance subspace rather than directly in spectral space (Li et al., 2022). The paper derives

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$2

and uses PAN Detail Inject Network blocks to inject PAN detail into abundance features, making the abundance representation the primary fusion target.

A fourth locus is feature-level or branch-level fusion. The 2025 SSUF module defines a spectral branch

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$3

and a spectral-spatial branch

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$4

so fusion occurs before residual refinement in a single-image SR network (Muhammad et al., 26 Sep 2025). DSANet uses a full-view spatial network, a multiview spectral network, and a cross-fusion attention network in which

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$5

followed by attention-weighted combination of branch outputs (Wang et al., 2024). These formulations move SSUF toward learned mid-level fusion rather than explicit variational coupling.

A fifth locus is unregistered abundance-domain reconstruction. In abundance fusion learning for unregistered HSI-SR, the network preserves endmembers estimated from the upsampled HSI, predicts enhanced abundances, reconstructs a residual HR-HSI by

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$6

and outputs

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$7

(Zhang et al., 9 Mar 2026). In blind tensor fusion, the latent tensor satisfies

$$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$8

so spectral and spatial degradations are fused through a common HR-HSI tensor while PSF and SRF are jointly estimated (Gao et al., 12 Mar 2026).

3. Principal methodological families

One major SSUF family performs post-unmixing spatial allocation. The 2013 PAN-guided method first estimates endmembers and abundances from the low-resolution HSI, then segments the corresponding PAN patch into as many classes as estimated endmembers, compares segment area proportions with abundance fractions, and assigns endmember spectra to PAN subregions (Rajabi et al., 2013). This is heuristic rather than jointly optimized, but it already embodies the SSUF idea that unmixing recovers composition while the high-resolution guide supplies spatial arrangement.

A second family performs external-data-guided regularization. LiDAR-aided unmixing keeps the spectral model unchanged and injects the DSM into the affinity weights of a weighted anisotropic TV-like regularizer (Uezato et al., 2017). The supplied description explicitly identifies this as an SSUF contribution because smoothing becomes guidance-aware and multimodal. The practical significance is strongest in shadowed regions, where hyperspectral guidance can be corrupted by illumination changes but DSM heights remain reliable.

A third family performs true joint spatial-spectral factorization. SP2U treats contextual features extracted from an image-derived panchromatic representation as explicit observations, factorizes them through a spatial dictionary $$\forall i,\ \forall m,\ a_{mi} \ge 0, \qquad \forall i,\ \sum_{m=1}^{M} a_{mi}=1$$9 and spatial codes $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$0, and couples them to spectral abundances through shared soft clusters (Lagrange et al., 2019). This is structurally different from abundance smoothing: spatial information is not merely a prior but a second matrix factorization problem. A plausible implication is that this family is the most direct precursor of later multimodal latent-space SSUF formulations.

A fourth family performs collaborative dual-stream unmixing. SSCU-Net explicitly separates a spatial autoencoder and a spectral convolutional autoencoder, shares decoder weights interpreted as endmembers, and aligns abundance estimates through a collaborative loss (Qi et al., 2022). Its central empirical finding is that spatial information is more beneficial for endmember extraction, whereas spectral information is more beneficial for abundance estimation. DSANet pushes the same general logic into attention-based branch fusion: a many-to-one spatial stream estimates the center-pixel abundance from a neighborhood, a multiview spectral stream estimates abundances from low-correlation spectral partitions, and a cross-fusion attention network combines the two (Wang et al., 2024).

A fifth family performs attention-based endmember or representation fusion. FusionNet contextualizes the center pixel using neighboring pixels as keys and values, predicts abundances with self-attention and softmax, and separately fuses multiple endmember sets from different extraction algorithms through endmember-specific attention (Ratnayake et al., 2024). Here SSUF is not only spectral-spatial fusion but also fusion of endmember hypotheses. This broadens the meaning of fusion from “combine spatial and spectral evidence” to “combine spatial evidence, spectral evidence, and multiple candidate material dictionaries.”

A sixth family uses architectural spectral-spatial fusion without strict classical unmixing. U2Net employs a spatial U-Net, a spectral U-Net, and an S2Block that computes spatial self-correlation and spectral self-correlation before multiplicative integration (Peng et al., 2022). DTDNML uses deep Tucker decomposition, shared latent core tensors, spatial-spectral attention, and Laplacian manifold constraints (Wang et al., 2024). The supplied descriptions explicitly state that these are not direct spectral unmixing methods, but they remain relevant because they operationalize modality-specific branches, hierarchical fusion, and latent shared structure.

4. Optimization paradigms and computational organization

Classical model-based SSUF is dominated by alternating optimization with convex subproblems. In FUMI, the fused image is parameterized as $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$1, the MAP objective combines HSI and MSI data-fitting terms under ANC, ASC, and reflectance bounds, and the algorithm alternates between abundance updates and endmember updates (Wei et al., 2016). Each block is solved with ADMM. The abundance update leads to a Sylvester equation and a simplex projection; the endmember update yields a second Sylvester equation and projection onto $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$2. This is one of the clearest examples of joint fusion and unmixing solved purely by constrained variational optimization.

LiDAR-guided regularization also uses ADMM, but on a different splitting. The weighted TV problem is written with auxiliary variables $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$3, positivity and sum-to-one projections, and a soft-thresholding update for the TV term (Uezato et al., 2017). When the weights depend on abundances, an outer reweighted-$$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$4-type loop recomputes the operator $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$5 from updated abundances. The supplied text notes that no formal convergence proof is provided for the reweighted outer loop, although the fixed-$$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$6 subproblems are standard convex proximal updates.

SP2U uses PALM, a block-coordinate proximal gradient method for nonconvex nonsmooth problems (Lagrange et al., 2019). The smooth term contains the spectral fit, contextual fit, coupling penalty, and orthogonality-promoting regularizer; projections are onto the nonnegative orthant or the probability simplex. The paper states that PALM guarantees convergence to a critical point under the required conditions. This optimization style differs from ADMM-based SSUF by working directly with nonconvex coupled factorizations rather than convex subproblems over a fixed operator.

Deep SSUF often replaces explicit constrained solvers with shared-parameter training. SSCU-Net uses alternating collaborative training of two autoencoders coupled through a shared decoder and abundance consistency (Qi et al., 2022). Pgnet is trained end-to-end with Adam for 500 epochs under a combined MSE and SAM loss, while the actual fusion process occurs in a 20-channel abundance subspace and PAN detail is injected through PDIN blocks (Li et al., 2022). Blind tensor fusion returns to optimization-centric design: it formulates blind fusion as a coupled inverse problem and solves it with a partially linearized ADMM algorithm with Moreau envelope smoothing, together with rigorous convergence analysis (Gao et al., 12 Mar 2026). Taken together, these works show a clear split between solver-driven SSUF and architecture-driven SSUF.

5. Empirical behavior and application regimes

The empirical literature shows that SSUF is not tied to a single task. It appears in hyperspectral unmixing, hyperspectral–multispectral fusion, panchromatic-guided hyperpansharpening, single-image super-resolution, and unregistered reference-guided HSI super-resolution. Across these settings, the practical motivation remains stable: recover spectral fidelity without sacrificing spatial detail.

For edge preservation and robustness to illumination, the LiDAR-guided study is especially explicit. On SIM1, the unweighted method gives $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$7 and $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$8, whereas the DSM-guided method reaches $$\min_{\mathbf{A}} \frac{1}{2}\|\mathbf{Y}-\mathbf{E}\mathbf{A}\|_F^2 + \lambda \phi(\mathbf{A})$$9, and combined methods such as w-HI-DSM and w-A-DSM reach the same whole-image RMSE with nearly identical edge RMSE (Uezato et al., 2017). On SIM2, where the DSM is real and imperfectly correlated with all abundance edges, gains are smaller but systematic, and the real-data analysis emphasizes that DSM-guided methods are less sensitive to shadows.

For two-stream abundance estimation, DSANet reports the best average SAD and RMSE on both Urban and Jasper Ridge among the listed baselines. On Urban, DSANet reaches average SAD $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$0 and average RMSE $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$1; on Jasper Ridge, it reaches average SAD $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$2 and average RMSE $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$3 (Wang et al., 2024). The supplied text also notes that the paper lacks a dedicated branch ablation, so the value of each stream is inferred indirectly from comparison with other methods rather than isolated experimentally.

FusionNet provides complementary evidence that attention-based contextualization and endmember fusion can improve both abundance and signature quality. On Samson, FusionNet achieves average RMSE $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$4 and average SAD $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$5, outperforming the listed transformer, CNN, and factorization baselines; on Jasper-Ridge it achieves the best average abundance RMSE $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$6 and the second-best average SAD $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$7 (Ratnayake et al., 2024). The EEA-combination ablation further shows that fusing all candidate endmember sets improves over any single initialization.

For hyperpansharpening, Pgnet emphasizes extremely large fusion ratio $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$8. On JiaXing at ratio 16, it reports PSNR $$\phi(\mathbf{A}) = \sum_{i=1}^{N}\sum_{j\in\mathcal{N}(i)} w_{ij}\|\mathbf{a}_i-\mathbf{a}_j\|_p^p,$$9, SSIM $w_{ij}$0, SAM $w_{ij}$1, ERGAS $w_{ij}$2, and SCC $w_{ij}$3; on Chikusei, PSNR $w_{ij}$4, SSIM $w_{ij}$5, SAM $w_{ij}$6, ERGAS $w_{ij}$7, and SCC $w_{ij}$8; on XiongAn, PSNR $w_{ij}$9, SSIM $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$0, SAM $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$1, ERGAS $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$2, and SCC $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$3 (Li et al., 2022). The supplied description attributes these gains to fusion inside a low-dimensional abundance subspace rather than direct spectral-space detail injection.

For collaborative autoencoder unmixing, SSCU-Net consistently improves both endmember SAD and abundance RMSE over EndNet, TANet, CNNAEU, and SSAE. On Urban, it reports average endmember SAD $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$4 and average abundance RMSE $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$5; on Jasper Ridge, $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$6 and $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$7; on Samson, $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$8 and $$\mathbf Y \approx \mathbf M \mathbf A,\qquad \mathbf S \approx \mathbf D \mathbf U,$$9, all in the units reported by the paper’s tables (Qi et al., 2022). The ablations are especially significant because they support a specific SSUF claim: spatial collaboration improves endmember extraction, while spectral collaboration improves abundance estimation.

The explicit SSUF module for single-image super-resolution reports that HDL with SSUF achieves strong MPSNR and SAM while using only $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$0M parameters. On Chikusei at $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$1, the paper reports $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$2 MPSNR, $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$3 CC, $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$4 RMSE, and $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$5 SAM; on PaviaU at $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$6, $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$7 MPSNR and $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$8 SAM (Muhammad et al., 26 Sep 2025). This is significant because it shows the term SSUF being used not only for cross-sensor fusion, but also for single-image hyperspectral SR through early spectral-spatial fusion inside a 2D CNN.

Recent unregistered work extends empirical SSUF behavior beyond co-registered settings. The abundance fusion learning framework reports state-of-the-art super-resolution performance, with ICVL $$\begin{pmatrix} \mathbf A\ \mathbf U \end{pmatrix} \approx \mathbf B \mathbf Z.$$9 results of PSNR $\mathbf A$0, SSIM $\mathbf A$1, and SAM $\mathbf A$2, and REAL $\mathbf A$3 results of $\mathbf A$4, $\mathbf A$5, and $\mathbf A$6, respectively (Zhang et al., 9 Mar 2026). The important empirical implication is that abundance-domain fusion remains viable even when the reference image is not spatially aligned with the hyperspectral observation.

6. Conceptual boundaries, limitations, and open directions

A persistent misconception is that SSUF always means explicit endmember-abundance unmixing with hard physical constraints. The supplied literature shows otherwise. Some methods enforce ANC and ASC explicitly in convex or nonconvex optimization (Uezato et al., 2017, Lagrange et al., 2019, Wei et al., 2016). Others use abundance-like latent variables without hard nonnegativity or sum-to-one mechanisms being specified, such as DSANet and Pgnet (Wang et al., 2024, Li et al., 2022). The 2025 SSUF module goes further: its “spectral unmixing” branch is implemented as stacked $\mathbf A$7 convolutions and explicitly “does not rely on a fixed number of endmembers” (Muhammad et al., 26 Sep 2025). This suggests that SSUF has both strict and loose usages.

Another boundary concerns what should count as SSUF at all. The supplied corpus explicitly states that DSCN is “not a true spectral-spatial unmixing fusion method” because it is spectral-only (Ozkan et al., 2018). U2Net is described as not an explicit spectral unmixing paper, and DTDNML as not a direct SSUF method in the classical endmember-abundance sense (Peng et al., 2022, Wang et al., 2024). These cases matter because strong spectral-spatial fusion performance does not automatically imply explicit unmixing semantics.

The literature also identifies clear failure modes. LiDAR guidance can become detrimental when abundance discontinuities do not correspond to height discontinuities, especially for sufficiently large $\mathbf A$8 on SIM2 (Uezato et al., 2017). PAN-guided subpixel allocation is limited by segmentation quality, ambiguity when abundance fractions are close, and the fact that PAN has no spectral specificity (Rajabi et al., 2013). Branch-attention methods can be difficult to interpret when notation or implementation details are incomplete, as noted for DSANet’s CFAN and missing hard abundance constraints (Wang et al., 2024). These are not merely implementation issues; they mark structural trade-offs between physical interpretability and architectural flexibility.

Open directions are also explicit in the supplied works. SP2U leaves more advanced spatial features, including convolutional representations, for future work (Lagrange et al., 2019). Pgnet notes that the generality of the already trained model on different hyperspectral datasets with different spectral bands still needs further research (Li et al., 2022). SSCU-Net states that its SAE and SCAE are simple networks and that more advanced reconstruction losses could further improve the framework (Qi et al., 2022). The 2026 unregistered HMF paper adds a different direction by supplying recoverability results for MSI-side and HSI-side SRIs in unregistered settings, which indicates that registration-free SSUF is becoming theoretically analyzable rather than purely heuristic (Song et al., 23 Mar 2026).

Taken together, these limitations and trajectories indicate that SSUF is evolving along three axes. First, the fusion variable is moving from explicit abundance maps toward more general latent representations. Second, the coupling mechanism is moving from regularization and joint likelihoods toward attention, shared decoders, and distribution matching. Third, the sensing scenario is moving from co-registered fusion toward blind and unregistered settings. A plausible implication is that future SSUF research will continue to hybridize physically constrained unmixing with multiscale deep fusion, blind degradation estimation, and misregistration-robust abundance modeling.