Temporal Frequency-Spatial Fusion (TFSF)

• Temporal Frequency-Spatial Fusion (TFSF) is a design paradigm that integrates temporal evolution, frequency content, and spatial structure to enhance the stability and discriminative power of representations.
• It is implemented through methods like STFT, attention mechanisms, and transformer-based temporal modeling in applications such as EEG motor imagery classification, remote sensing, and video fusion.
• Practical insights include the benefits of explicit spectral and spatial attention, hierarchical fusion architectures, and the balance between temporal robustness and spatial detail across domains.

Searching arXiv for papers relevant to Temporal Frequency–Spatial Fusion and closely related formulations. Temporal Frequency–Spatial Fusion (TFSF) denotes a class of fusion strategies that jointly exploit temporal variation, frequency-domain or frequency-like structure, and spatial organization in order to construct a more discriminative or more stable representation than any single domain alone can provide. Across application areas, the term is not universally standardized, but the underlying design pattern recurs: a model first represents temporal evolution, then extracts or reweights frequency content, then incorporates spatial structure, or it performs these operations in a coupled manner within a unified architecture. In electroencephalography (EEG), TFSF is concretely instantiated by the Spatial–Spectral–Temporal Attention Fusion transformer, which uses a time–frequency transform, spectral attention, spatial attention, and transformer-based temporal modeling for cross-subject motor imagery classification (Muna et al., 17 Apr 2025). In remote sensing, the same conceptual objective appears under “spatiotemporal fusion,” where high temporal frequency observations from coarse sensors are combined with high spatial detail from sparse fine-resolution imagery to synthesize dense fine-resolution image series (Albanwan et al., 2021, Cheng et al., 2016, Li et al., 2023). In video fusion, frequency-aware architectures separate low- and high-frequency components and assign them distinct temporal treatments to balance temporal stability and spatial fidelity (Li et al., 2 Apr 2026).

1. Conceptual scope and terminology

TFSF is best understood as an organizing principle rather than a single canonical algorithm. In the EEG setting, the SSTAF model operationalizes this idea under the name “Spatial–Spectral–Temporal Attention Fusion,” using short-time Fourier transform (STFT) to build a time–frequency representation, followed by spectral and spatial attention and a temporal transformer (Muna et al., 17 Apr 2025). In remote sensing, the phrase “Temporal Frequency–Spatial Fusion” does not typically appear verbatim, but the chapter on spatiotemporal fusion explicitly frames the task as combining high spatial resolution and high temporal resolution imagery by exploiting correlations over both space and time, which is functionally the same objective (Albanwan et al., 2021). The STNLFFM paper likewise formulates the problem as blending high spatial resolution but low temporal frequency imagery with low spatial resolution but high temporal frequency imagery to generate a continuous Landsat-like time series (Cheng et al., 2016).

A useful cross-domain interpretation is that TFSF methods operate on a signal defined over space and time, and either decompose or weight the signal along a frequency axis explicitly, as in STFT-based EEG modeling or frequency-aware video fusion, or implicitly, as in temporal filters, temporal graphs, or state-space priors (Muna et al., 17 Apr 2025, Li et al., 2 Apr 2026, Li et al., 2020, Li et al., 2023). This suggests that “frequency” in TFSF can refer either to an explicit spectral representation, such as Fourier coefficients or spectrogram bins, or to model components that behave like temporal or spatial filters without an explicit transform.

The breadth of the concept has produced several near-equivalent formulations in the literature. EEG work uses “spatial–spectral–temporal attention fusion” (Muna et al., 17 Apr 2025); traffic forecasting uses “spatial-temporal fusion” through graph construction without explicit spectral decomposition (Li et al., 2020); remote sensing uses “spatiotemporal fusion” to describe multi-source temporal densification and spatial sharpening (Albanwan et al., 2021, Cheng et al., 2016); and video fusion uses “frequency-aware” low/high-frequency branches combined with temporal perturbation and temporal consistency constraints (Li et al., 2 Apr 2026). The common denominator is joint exploitation of complementary dependencies across temporal evolution, frequency content, and spatial structure.

2. Canonical architecture in EEG: SSTAF as a concrete TFSF realization

The clearest direct realization of TFSF in the provided literature is the SSTAF transformer for EEG-based motor imagery classification (Muna et al., 17 Apr 2025). It addresses cross-subject upper-limb motor imagery classification on EEGMMIDB and BCI Competition IV-2a, where EEG non-stationarity and inter-subject variability make robust subject-independent modeling difficult. The architecture comprises a spectral transformer and a spatial transformer, followed by a transformer block and a classifier network, with attention mechanisms that attend to spectral frequencies, spatial electrode locations, and temporal dynamics (Muna et al., 17 Apr 2025).

On EEGMMIDB, the paper uses 103 subjects, 64 channels, and 160 Hz sampling, retaining only motor imagery runs and extracting 13,831 epochs of length 4 s, yielding 640 time points by 64 channels after preprocessing (Muna et al., 17 Apr 2025). Preprocessing includes an 8–30 Hz bandpass filter with a 5th-order Butterworth filter,

$X^{bf}_{c,t} = \mathcal{B}(\varphi(c,t)),$

a 50 Hz notch filter,

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$

and common average reference,

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$

followed by per-channel standardization with $\epsilon = 10^{-8}$ (Muna et al., 17 Apr 2025). The paper explicitly states that this preprocessing ensures the data fed to the spectral and spatial modules are already frequency-constrained and amplitude-normalized, which is a crucial step for stable TFSF-type fusion (Muna et al., 17 Apr 2025).

The temporal–frequency stage is built with STFT. For each channel and epoch,

$$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$

with Hann window, $n_{fft}=128$, hop size $R = H_l = 64$, and sampling frequency $f_s = 160$ Hz (Muna et al., 17 Apr 2025). The resulting tensor per epoch has shape $(C, f_{bin}, t_f)$ with $f_{bin}=65$, and over batches,

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$0

This representation already couples temporal and spectral information per channel before any attention mechanism is applied (Muna et al., 17 Apr 2025).

The frequency-side of TFSF is realized by spectral attention. SSTAF computes a global frequency descriptor by averaging over channels and time,

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$1

producing a tensor of shape $X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$2, and maps it through a single-head MLP with Softmax:

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$3

The resulting per-frequency weights are broadcast and applied as

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$4

thereby reweighting each frequency bin across all channels and time frames (Muna et al., 17 Apr 2025).

The spatial-side of TFSF follows by averaging over frequency and time to obtain a channel descriptor and applying a structurally similar MLP:

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$5

with output applied as

$X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$6

Spectral attention is applied first and spatial attention second, yielding what the paper characterizes as a spatio-spectral representation (Muna et al., 17 Apr 2025). Because both attention weights are computed after collapsing over time, the frequency–spatial weighting is conditioned on temporal context.

Explicit temporal fusion is then performed by a transformer encoder that consumes a sequence of shape $X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$7, where the paper describes $X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$8, 2 encoder layers, 8 heads, and feed-forward dimension $X^{nf}_{c,t} = \mathcal{N}(X^{filtered}_{c,t}),$9, with GELU, dropout, and LayerNorm (Muna et al., 17 Apr 2025). Multi-head temporal self-attention follows the standard form

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$0

with head aggregation

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$1

Time-average pooling and a two-layer classifier complete the pipeline (Muna et al., 17 Apr 2025).

The paper explicitly interprets this as a three-stage TFSF pipeline: temporal–frequency representation via STFT, frequency–spatial fusion via dual attention, and temporal fusion via transformer self-attention (Muna et al., 17 Apr 2025). It reports subject-independent accuracy of 76.83% and F1 73.52% on EEGMMIDB, and 68.30% accuracy and 70.63% F1 on BCI IV-2a (Muna et al., 17 Apr 2025). The ablation study on EEGMMIDB further shows 76.83 Acc and 0.744 F1 for full SSTAF, 70.21 and 0.688 without spectral attention, 72.96 and 0.713 without spatial attention, and 63.47 and 0.608 without transformer (Muna et al., 17 Apr 2025). The largest drop occurs when the temporal transformer is removed, while spectral and spatial attention also contribute materially, supporting the claim that all three axes of TFSF are functionally significant in this setting (Muna et al., 17 Apr 2025).

3. Explicit versus implicit frequency modeling

One major axis of variation among TFSF-like methods is whether the frequency dimension is explicit. SSTAF uses an explicit time–frequency representation through STFT and then manipulates a literal spectral axis with learned attention (Muna et al., 17 Apr 2025). FTPFusion also uses explicit frequency decomposition, but in a different sense: it decomposes feature maps into low-frequency and high-frequency spatial components with a learned weighted multi-scale smoothing operator and residual high-pass branch (Li et al., 2 Apr 2026). For a feature map $$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$2, the low-frequency component is constructed from average-pooling kernels of sizes $$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$3,

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$4

and

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$5

This is an explicit frequency-aware split, but it is spatial-frequency decomposition rather than temporal-frequency decomposition (Li et al., 2 Apr 2026).

By contrast, STFGNN contains no explicit frequency-domain transform. It constructs a spatial graph, a temporal graph based on Dynamic Time Warping similarity, and a temporal connectivity graph, then fuses them structurally into a block adjacency and applies graph propagation with gated linear units (Li et al., 2020). The paper is explicit that STFGNN does not perform Fourier transforms, wavelets, or spectral graph convolution using Laplacian eigenvectors; its dilated convolutions provide implicit multi-scale temporal filtering rather than explicit frequency decomposition (Li et al., 2020). A plausible implication is that TFSF can be realized without an explicit spectral basis when temporal and spatial fusion mechanisms have sufficiently rich filtering behavior.

The remote-sensing chapter presents a similar case. It emphasizes temporal weighting, temporal windows, and temporal similarity in multi-date imagery, but not explicit Fourier-domain temporal analysis (Albanwan et al., 2021). Temporal frequency is instead realized by sampling density and the choice of temporal weighting functions, such as

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$6

within 3D spatiotemporal bilateral filtering (Albanwan et al., 2021). This suggests that in remote sensing, “temporal frequency” is often operationalized as dense revisit rate rather than frequency-domain coefficients.

The state-space fusion method for multiresolution multispectral imagery makes this implicitness even clearer. It uses a random walk

$$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$7

with $$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$8 in experiments and time-varying process covariance $$X^{CAR}_{c,t} = X^{notched}_{c,t} - \frac{1}{N}\sum_{c=1}^{N} X^{notched}_{c,t},$$9 calibrated from historical data (Li et al., 2023). There is no spectral transform in time, but the process covariance regulates how much rapid temporal variation each state component can express. This suggests an interpretation of TFSF in which frequency is represented by temporal smoothness priors or temporal variance structure rather than by explicit spectral coordinates.

The literature therefore supports two technically distinct interpretations of TFSF. One is explicit, where a model contains a frequency axis or frequency bands, as in STFT-based EEG fusion and low/high-frequency video branches (Muna et al., 17 Apr 2025, Li et al., 2 Apr 2026). The other is implicit, where temporal and spatial operators have frequency-selective behavior without explicit Fourier coordinates, as in graph fusion, remote-sensing weighting schemes, and Bayesian state-space filtering (Li et al., 2020, Albanwan et al., 2021, Li et al., 2023).

4. Domain-specific realizations beyond EEG

In remote sensing, TFSF corresponds to the synthesis of outputs that simultaneously exhibit high spatial detail and high temporal completeness. The chapter on spatiotemporal fusion describes the canonical problem as combining high-spatial/low-temporal sensors such as Landsat or Sentinel-2 with low-spatial/high-temporal sensors such as MODIS or VIIRS, using redundant information in space and time to infer unobserved fine-resolution states (Albanwan et al., 2021). A generic weighting form is

$\epsilon = 10^{-8}$0

with weights factorized into spatial, spectral, and temporal terms (Albanwan et al., 2021). In the 3D spatiotemporal bilateral filter, the fused value is

$\epsilon = 10^{-8}$1

showing an explicitly multiplicative three-way fusion of spatial, spectral, and temporal similarity (Albanwan et al., 2021).

STNLFFM offers a more algorithmic remote-sensing realization. It blends coarse-resolution but high temporal frequency imagery with fine-resolution but low temporal frequency imagery by estimating a local linear transformation between fine-resolution images at different dates, constrained by coarse-resolution observations (Cheng et al., 2016). Its core prediction takes the form

$\epsilon = 10^{-8}$2

where weights combine non-local patch similarity and a whole-date weight derived from temporal reliability (Cheng et al., 2016). Although no explicit frequency axis is present, the problem setting is fundamentally one of temporal frequency–spatial fusion: daily MODIS-like dynamics are mapped to Landsat-like spatial detail (Cheng et al., 2016).

The online Bayesian fusion model advances the same objective in state-space form. The latent high-resolution sequence $\epsilon = 10^{-8}$3 is observed through sensor-specific degradation operators,

$\epsilon = 10^{-8}$4

and propagated in time by a random walk with calibrated process noise (Li et al., 2023). The paper explicitly targets the generation of high-resolution image sequences at higher revisiting rates and evaluates the method on water mapping using Landsat and MODIS (Li et al., 2023). A plausible implication is that TFSF in remote sensing can be formalized equally well through deterministic weighted fusion models or probabilistic dynamical systems.

In traffic forecasting, STFGNN constructs a block adjacency $\epsilon = 10^{-8}$5 that embeds spatial graph structure, temporal similarity, and temporal connectivity across a local window of $\epsilon = 10^{-8}$6 consecutive time steps (Li et al., 2020). Graph propagation is then performed with a GLU:

$\epsilon = 10^{-8}$7

The method does not perform explicit frequency-domain operations, but its dilated temporal convolutions are described as implicitly multi-scale and therefore frequency-like (Li et al., 2020). The paper itself proposes extending the model toward explicit TFSF by adding Fourier transforms, frequency graphs, or learnable frequency filters (Li et al., 2020). This is notable because it frames TFSF not as a fixed architecture but as a generalization path from spatial–temporal fusion to explicit frequency-aware modeling.

In infrared-visible video fusion, FTPFusion separates low-frequency and high-frequency branches and treats them differently over time (Li et al., 2 Apr 2026). The high-frequency branch uses sparse cross-modal spatio-temporal interaction, block selection with $\epsilon = 10^{-8}$8, and top-$\epsilon = 10^{-8}$9 block retention based on a sparse ratio $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$0 to focus computation on motion-related and detail-rich regions (Li et al., 2 Apr 2026). The low-frequency branch applies temporal perturbation,

$$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$1

local 3D enhancement,

$$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$2

and temporal mean preservation (Li et al., 2 Apr 2026). It further imposes an offset-aware temporal consistency loss

$$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$3

defined on low-frequency video components (Li et al., 2 Apr 2026). This architecture makes explicit the idea that different frequency bands should receive different temporal treatments, which is one of the clearest modern formulations of TFSF outside EEG.

5. Fusion mechanisms and recurring design patterns

Across these domains, several fusion mechanisms recur. The first is multiplicative or attention-based reweighting. In SSTAF, spectral and spatial attention are both Softmax-normalized MLP outputs that reweight the original tensor along frequency and channel axes (Muna et al., 17 Apr 2025). In remote sensing, spatiotemporal bilateral filters likewise multiply spatial, spectral, and temporal weights before normalization (Albanwan et al., 2021). These are mathematically distinct implementations but share the same structural principle: domain-specific saliency is computed from a global or local descriptor and then broadcast back onto the original representation.

The second recurring mechanism is hierarchical fusion. SSTAF first couples time and frequency through STFT, then fuses frequency and space with dual attention, and only afterwards performs temporal modeling through a transformer (Muna et al., 17 Apr 2025). FTPFusion likewise first decomposes features into low and high frequency components, then applies branch-specific temporal and cross-modal modeling, and finally recombines them (Li et al., 2 Apr 2026). STFGNN similarly constructs a fused graph before applying graph propagation and a parallel gated convolution module (Li et al., 2020). This suggests that TFSF architectures often do not fuse all dimensions simultaneously at the first stage; instead they build increasingly abstract fused representations.

A third pattern is the use of global descriptors to derive axis-specific weights. SSTAF averages over channels and time to derive frequency weights, and over frequency and time to derive channel weights (Muna et al., 17 Apr 2025). FTPFusion pools low-frequency features to obtain temporal activity descriptors and shared temporal context (Li et al., 2 Apr 2026). In remote sensing, temporal weights may depend on image-to-image discrepancy or class-specific temporal variability (Albanwan et al., 2021). This suggests that TFSF often couples one domain to another through a summary statistic rather than direct dense pairwise interaction.

A fourth pattern is branch specialization. In FTPFusion, the low-frequency branch is responsible for temporal stability, flicker resistance, and robustness to misalignment, while the high-frequency branch concentrates on edges, textures, and motion-related detail (Li et al., 2 Apr 2026). In TSFmicro for dynamic micro-expression recognition, temporal and spatial features are extracted in parallel and fused late:

$$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$4

with the authors describing the resulting representation as complementary “where-how” semantics (Liu et al., 22 May 2025). Although TSFmicro does not explicitly use frequency-domain features, its late-fusion design shows that preserving modality-specific structure until a high-level fusion stage can be advantageous when branches encode different aspects of a phenomenon (Liu et al., 22 May 2025). A plausible implication is that explicit TFSF models may benefit from keeping temporal-frequency and spatial branches separate until semantically rich intermediate features have formed.

A final pattern is that temporal consistency is often enforced separately from spatial-detail preservation. FTPFusion’s offset-aware temporal consistency loss changes temporal metrics while leaving spatial metrics largely intact in ablations (Li et al., 2 Apr 2026). SSTAF’s transformer removal produces the largest performance drop even though spectral and spatial attention remain, indicating that temporal modeling cannot be reduced to static pooling over time (Muna et al., 17 Apr 2025). These examples indicate that in TFSF systems, temporal fusion is not merely an optional refinement layered on top of spatial-frequency modeling; it is often a coequal component that must be explicitly optimized.

6. Evaluation, interpretability, and limitations

Because TFSF spans disparate application domains, evaluation criteria vary widely. In EEG motor imagery classification, SSTAF is evaluated with subject-independent accuracy and F1 under $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$5-fold cross-validation and Leave-One-Subject-Out protocols, achieving 76.83% accuracy on EEGMMIDB and 68.30% on BCI IV-2a (Muna et al., 17 Apr 2025). Its ablations provide direct evidence that temporal, spectral, and spatial modules each contribute to performance (Muna et al., 17 Apr 2025). The paper also includes interpretability analyses showing that after training, spectral attention emphasizes $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$6 and $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$7 rhythms and spatial attention emphasizes sensorimotor electrodes such as C3, C4, and Cz, aligning with known motor-imagery neurophysiology (Muna et al., 17 Apr 2025).

In remote sensing, performance is assessed with reconstruction metrics such as SAM, RMSE, and $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$8, as well as downstream task metrics such as water mapping accuracy. The online fusion model reports superior performance to ESTARFM and PSRFM on Oroville and Elephant Butte, with smoothing variants outperforming filtering variants when future information is available (Li et al., 2023). STNLFFM shows lower RMSE and higher $$X[m, k] = \sum_{n=0}^{L-1} x[n + m R] w[n] e^{-j \frac{2\pi}{L} k n},$$9 than STARFM and ESTARFM across heterogeneous and temporally dynamic landscapes, especially in edge regions and flooded areas (Cheng et al., 2016). These results suggest that TFSF benefits are especially pronounced when either spatial heterogeneity or temporal dynamics are strong.

In video fusion, FTPFusion is evaluated with both spatial and temporal metrics. The paper reports that it achieves top or second-best results on $n_{fft}=128$0, $n_{fft}=128$1, EN, SCD, BiSwE, MS2R, MMCI, and TCPE across M3SVD, HDO, and VTMOT (Li et al., 2 Apr 2026). Its ablation study shows that removing the dual-branch frequency-aware module causes the largest drop, while removing the temporal consistency loss mainly degrades temporal metrics (Li et al., 2 Apr 2026). This provides evidence that explicit frequency-aware branching and explicit temporal regularization contribute different but complementary effects.

Interpretability in TFSF systems is often strongest where the frequency or spatial weighting is explicit. SSTAF attention maps can be inspected directly (Muna et al., 17 Apr 2025). In remote sensing, weighting terms and state covariances have physical interpretation as similarity, spatial proximity, or temporal change uncertainty (Albanwan et al., 2021, Li et al., 2023). By contrast, graph-based or deep video models with implicit frequency selectivity are more difficult to interpret mechanistically, even if their operators can be described as frequency-like (Li et al., 2020, Li et al., 2 Apr 2026).

Several misconceptions follow from the heterogeneous terminology. One is that TFSF necessarily requires explicit Fourier analysis. The literature does not support this: STFGNN, STNLFFM, and the state-space Bayesian model all instantiate the temporal–spatial part of the paradigm without explicit frequency transforms (Li et al., 2020, Cheng et al., 2016, Li et al., 2023). Another is that “frequency” always refers to temporal frequency. In SSTAF, the spectral axis is the EEG frequency axis after STFT (Muna et al., 17 Apr 2025); in FTPFusion, the low/high-frequency decomposition is spatial frequency decomposition applied framewise (Li et al., 2 Apr 2026). A plausible implication is that the exact meaning of the frequency dimension in TFSF is application-dependent and should always be defined operationally.

The principal limitations also recur across domains. Remote-sensing methods note difficulties with abrupt land-cover change, mixed pixels, radiometric inconsistency, and cloud contamination (Albanwan et al., 2021). STNLFFM highlights the limitations of linear temporal change assumptions and non-local search cost (Cheng et al., 2016). The Bayesian online fusion model assumes linear-Gaussian dynamics and relies on historical data to calibrate $n_{fft}=128$2 (Li et al., 2023). FTPFusion remains local in time and uses small translation windows for offset modeling (Li et al., 2 Apr 2026). SSTAF does not detail positional encoding explicitly and is evaluated on moderate-sized EEG datasets, which constrains architectural scale (Muna et al., 17 Apr 2025). These limitations suggest that while TFSF is a broadly useful design paradigm, its success depends heavily on how each domain instantiates temporal priors, frequency structure, and spatial correspondence.

Taken together, the literature supports a technically precise view of Temporal Frequency–Spatial Fusion as a family of architectures and models that combine temporal information, frequency content or frequency-selective processing, and spatial structure through explicit transforms, learned weighting, graph construction, or dynamical inference. SSTAF provides a direct end-to-end template in which STFT, spectral attention, spatial attention, and temporal self-attention form a unified pipeline (Muna et al., 17 Apr 2025). Remote-sensing methods show that the same idea can be realized as weighted spatiotemporal filtering or state-space estimation to synthesize dense fine-resolution series from heterogeneous sensors (Albanwan et al., 2021, Cheng et al., 2016, Li et al., 2023). Frequency-aware video fusion demonstrates a modern branch-specialized version in which low- and high-frequency components are assigned different temporal treatments (Li et al., 2 Apr 2026). The diversity of these implementations indicates that TFSF is best regarded as a general fusion doctrine: the joint modeling of temporal evolution, frequency-selective structure, and spatial organization in a representation that is more stable, more informative, or more discriminative than any of its constituent views alone.