SatFusion: Unified Satellite Image Fusion

• SatFusion is a comprehensive framework that integrates multi-temporal and multi-source fusion to produce high-resolution satellite imagery while preserving spatial detail and spectral fidelity.
• It employs modular components such as Multi-Temporal Image Fusion and Multi-Source Image Fusion to seamlessly align and inject complementary information despite noise and misregistration.
• Empirical evaluations on datasets like WorldStrat, WV3, and GF2 demonstrate significant improvements in PSNR and ERGAS over traditional MISR and pansharpening methods.

SatFusion is a contemporary term used in multiple, partly overlapping senses across remote sensing and adjacent sensor-fusion literature. In its most specific usage, it denotes a unified framework for enhancing Satellite Internet of Things images by jointly exploiting multi-temporal multispectral observations and a high-resolution panchromatic image through Multi-Temporal Image Fusion, Multi-Source Image Fusion, and Fusion Composition (Tong et al., 9 Oct 2025). In broader usage, the term also describes multi-temporal satellite super-resolution, noise-aware spatio-temporal fusion, pre-fusion standardization across sensors, radar–optical fusion for space surveillance and tracking, and satellite-image-assisted fusion in online HD map construction (Luo et al., 2024, Isono et al., 15 Aug 2025, Goyena et al., 17 Oct 2025, Coelho et al., 2022, Huang et al., 12 Dec 2025).

1. Terminology and scope

Recent arXiv usage shows that “SatFusion” is not restricted to a single algorithmic family. It can name a specific end-to-end image enhancement framework, but it also functions as a general label for fusing complementary information across time, modality, or sensor geometry in satellite-related pipelines.

Usage	Core inputs	Output or objective
SatFusion framework	Multi-temporal LRMS images and one HR PAN	One fused HRMS image
SatDiffMoE as SatFusion	Arbitrary-number sequential LR satellite images	One HR reconstructed image
TSSTF / STIF context	HR-LR reference pair and LR target image	HR target image at target date
PASO radar–optical concept	Co-located radar and optical measurements	Joint estimation and rapid TLE updates
SATMapTR satellite-image fusion	Multi-view RGB cameras and satellite patch	Vectorized online HD map elements

In the strict sense introduced in the 2025 Sat-IoT paper, the input is a set of multi-temporal low-resolution multispectral images $$\{lrms_t\}_{t=1}^T \in \mathbb{R}^{H/\gamma \times W/\gamma \times C_{ms}}$$ and a single high-resolution panchromatic image $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$, with $C_{pan}=1$, and the output is a fused high-resolution multispectral image $SR \in \mathbb{R}^{H \times W \times C_{ms}}$ (Tong et al., 9 Oct 2025). In the broader literature, the same label covers cases where the fused product is not an image in the usual pansharpening sense, but a reconstructed HR scene, a fused state estimate, or a vectorized map representation (Luo et al., 2024, Coelho et al., 2022, Huang et al., 12 Dec 2025).

This suggests that SatFusion is best understood as a family resemblance term: the common thread is the exploitation of complementary information that is unavailable from any single acquisition stream alone.

2. The unified SatFusion framework for Satellite IoT image enhancement

The named SatFusion framework is an end-to-end architecture composed of three modules: Multi-Temporal Image Fusion (MTIF), Multi-Source Image Fusion (MSIF), and Fusion Composition (Tong et al., 9 Oct 2025). Its stated purpose is to transform large-scale multi-temporal and multi-source observations into a single high-resolution multispectral product that preserves both spatial detail from PAN and spectral fidelity from MS, while remaining robust to misregistration and noise.

MTIF first encodes each temporal multispectral input with shared weights, fuses the encoded representations using a pluggable MISR backbone, and decodes the result to PAN resolution by sub-pixel convolution. The paper states that alignment to PAN is achieved implicitly by learning HR spatial features from multi-temporal inputs and decoding them via PixelShuffle to match PAN resolution; no explicit flow-based warping or deformable convolution is introduced (Tong et al., 9 Oct 2025). This design is contrasted with classical pansharpening pipelines that rely on naive pre-interpolation of a single LRMS input.

MSIF then performs multi-source fusion by injecting fine-grained PAN texture into the high-resolution multispectral features produced by MTIF. It is implemented by plugging in a pansharpening module such as PNN, PanNet, INNformer, or Pan-Mamba, so the framework is orchestration-oriented rather than tied to one fixed feature injector (Tong et al., 9 Oct 2025).

Fusion Composition adaptively merges the MTIF and MSIF outputs and performs spectral adjustment through $1 \times 1$ convolutions. The paper gives the fusion-composition equation as

$$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$

The first residual merge aggregates complementary information, while the final $1 \times 1$ convolution emphasizes spectral consistency and channel mixing (Tong et al., 9 Oct 2025).

The training objective is a weighted composite loss

$$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$

with $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$ and default weights $\lambda_1 = 0.3$, $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$0, $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$1, $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$2 (Tong et al., 9 Oct 2025). Before loss evaluation, the framework applies channel-wise brightness compensation,

$Pan \in \mathbb{R}^{H \times W \times C_{pan}}$3

followed by

$Pan \in \mathbb{R}^{H \times W \times C_{pan}}$4

and uses Lanczos-based cropping and shift compensation to select the minimum loss across shifted variants (Tong et al., 9 Oct 2025).

A notable architectural property is modularity. MTIF and MSIF are explicitly described as pluggable from prior MISR and pansharpening work, including CNN, residual, transformer-based, invertible-network, and Mamba-based backbones. SatFusion itself contributes the shared encoders, PixelShuffle decoder, and lightweight Fusion Composition that bridge the temporal and source dimensions within one trainable pipeline (Tong et al., 9 Oct 2025).

3. Datasets, metrics, and empirical behavior of the SatFusion framework

The framework is evaluated on WorldStrat as a real dataset and on WV3, QB, and GF2 as synthetic datasets (Tong et al., 9 Oct 2025). WorldStrat provides multiple multi-temporal LRMS images, one HRPAN, and one HRMS per location. WV3, QB, and GF2 are used with Wald-protocol-derived LRMS/PAN pairs, but the paper modifies this protocol by introducing pixel shifts and noise to create synthetic multi-temporal LRMS sets more consistent with Sat-IoT conditions (Tong et al., 9 Oct 2025).

Evaluation uses PSNR, SSIM, SAM, and ERGAS. The paper reports that, on real WorldStrat data, MISR baselines achieve PSNR approximately $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$5–$Pan \in \mathbb{R}^{H \times W \times C_{pan}}$6, pansharpening baselines achieve PSNR approximately $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$7–$Pan \in \mathbb{R}^{H \times W \times C_{pan}}$8, and SatFusion combinations consistently outperform both categories (Tong et al., 9 Oct 2025). Representative WorldStrat results include SRCNN+PNN with PSNR $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$9, SSIM $C_{pan}=1$0, ERGAS $C_{pan}=1$1; HighRes-Net+PNN with PSNR $C_{pan}=1$2, SSIM $C_{pan}=1$3, ERGAS $C_{pan}=1$4; and RAMS+INNformer with PSNR $C_{pan}=1$5 and ERGAS approximately $C_{pan}=1$6 (Tong et al., 9 Oct 2025).

The paper summarizes average gains on WorldStrat as follows: compared to MISR, SatFusion improves PSNR by $C_{pan}=1$7 and ERGAS by $C_{pan}=1$8; compared to pansharpening, it boosts PSNR by $C_{pan}=1$9 and ERGAS by $SR \in \mathbb{R}^{H \times W \times C_{ms}}$0 (Tong et al., 9 Oct 2025). On synthetic WorldStrat with perturbations, SatFusion combinations reach approximately PSNR $SR \in \mathbb{R}^{H \times W \times C_{ms}}$1, with RAMS+INNformer reporting PSNR $SR \in \mathbb{R}^{H \times W \times C_{ms}}$2 and ERGAS approximately $SR \in \mathbb{R}^{H \times W \times C_{ms}}$3, while RAMS+PanNet reports SAM $SR \in \mathbb{R}^{H \times W \times C_{ms}}$4 (Tong et al., 9 Oct 2025).

Cross-sensor experiments on WV3, GF2, and QB show the same pattern. On WV3, FusionNet attains PSNR $SR \in \mathbb{R}^{H \times W \times C_{ms}}$5, whereas SatFusion combinations reach up to PSNR $SR \in \mathbb{R}^{H \times W \times C_{ms}}$6 with RAMS+FusionNet, SSIM up to $SR \in \mathbb{R}^{H \times W \times C_{ms}}$7, and ERGAS down to $SR \in \mathbb{R}^{H \times W \times C_{ms}}$8 (Tong et al., 9 Oct 2025). On GF2, FusionNet reports PSNR $SR \in \mathbb{R}^{H \times W \times C_{ms}}$9, while TR-MISR+FusionNet reaches PSNR $1 \times 1$0, SSIM $1 \times 1$1, and ERGAS $1 \times 1$2 (Tong et al., 9 Oct 2025). On QB, the strongest SatFusion performance is near PSNR $1 \times 1$3, SSIM approximately $1 \times 1$4, and the lowest SAM is $1 \times 1$5 for HighRes-Net+FusionNet (Tong et al., 9 Oct 2025).

Robustness analysis uses a perturbation parameter $1 \times 1$6 controlling joint disturbances: pixel displacement up to $1 \times 1$7 pixels, noise intensity $1 \times 1$8, and brightness shift in $1 \times 1$9. The paper states that, as $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$0 increases, SatFusion degrades more slowly than pansharpening baselines, which is presented as evidence of improved robustness under blur, misalignment, and cross-modal discrepancies (Tong et al., 9 Oct 2025).

Ablation results indicate that increasing the number of input frames $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$1 yields significant gains initially and then saturates; that removing any one loss component degrades at least one metric; and that removing the final $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$2 convolution generally reduces PSNR and SSIM while introducing color distortions (Tong et al., 9 Oct 2025). Parameter counts are given for individual backbones, such as SRCNN at $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$3M, HighRes-Net at $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$4M, RAMS at $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$5M, and TR-MISR at $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$6M, with SatFusion combinations described as roughly the sum of the selected MISR and pansharpening modules. FLOPs, runtime, and memory footprint are not reported (Tong et al., 9 Oct 2025).

4. SatFusion as multi-temporal and spatio-temporal image fusion

A broader SatFusion interpretation encompasses methods that do not fuse PAN imagery but still combine observations across time to reconstruct a high-resolution product. SatDiffMoE is a prominent example: it reconstructs a single HR image $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$7 from an arbitrary number of LR satellite observations $$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$8 acquired at different times, under the contextual observation model

$$SR = \text{Conv}\big( \text{ReLU}(\text{Conv}( SR_{\text{MTIF}} + SR_{\text{MSIF}} )) \big).$$9

The method does not explicitly invert this physical model; instead, it learns an end-to-end latent diffusion prior conditioned on LR observations and relative time differences, then fuses the multi-temporal information during inference (Luo et al., 2024).

SatDiffMoE operates in latent space with a Stable Diffusion 1.2 latent U-Net, a VAE-style encoder–decoder, and a time-aware conditioning mechanism in which the relative time difference $1 \times 1$0 is embedded by a cloned time-embedding network and added to the denoising U-Net’s time embedding (Luo et al., 2024). Its distinguishing feature is Mixture-of-Estimation rather than Mixture-of-Experts: for each input frame, the model computes a clean latent estimate via Tweedie’s formula, then solves for a robust latent-space center

$1 \times 1$1

interpolates each per-frame clean component toward that center with weight $1 \times 1$2, and preserves the DDIM noise component during reverse updates (Luo et al., 2024). The method is explicitly presented as a SatFusion approach because it allows arbitrary $1 \times 1$3, uses permutation-invariant latent fusion, and leverages a strong generative prior to sample plausible HR reconstructions conditioned on LR inputs and $1 \times 1$4 (Luo et al., 2024).

Quantitatively, SatDiffMoE reports on WorldStrat overall LPIPS $1 \times 1$5 and FID $1 \times 1$6, both best among the listed baselines, together with PSNR $1 \times 1$7 and SSIM $1 \times 1$8 (Luo et al., 2024). On fMoW overall it reports LPIPS $1 \times 1$9, best in that metric, and FID $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$0, while diffusion baselines obtain lower FID there (Luo et al., 2024). The ablation “no dt, no fusion” versus “with dt only” versus “with dt + fusion” shows PSNR $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$1, SSIM $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$2, LPIPS $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$3, and FID $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$4, indicating that both time conditioning and latent fusion materially contribute (Luo et al., 2024).

Noise-aware spatio-temporal fusion represents another branch of the broader SatFusion landscape. TSSTF predicts an HR multispectral image at a target date $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$5 from a noisy HR-LR reference pair at $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$6 and an LR target image, with additive Gaussian noise and sparse outlier terms on HR and LR observations (Isono et al., 15 Aug 2025). Its two core mechanisms are Temporally-Guided Total Variation,

$$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$7

and Temporally-Guided Edge Constraint,

$$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$8

with the paper recommending $$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\mathrm{MAE}}(SR, GT) + \lambda_2 \mathcal{L}_{\mathrm{MSE}}(SR, GT) + \lambda_3 \mathcal{L}_{\mathrm{SSIM}}(SR, GT) + \lambda_4 \mathcal{L}_{\mathrm{SAM}}(SR, GT),$$9, $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$0, $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$1, $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$2, and $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$3 (Isono et al., 15 Aug 2025). On simulated noisy cases, TSSTF is reported as consistently achieving the highest PSNR across all sites; for example, Site1 Case4 gives PSNR $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$4 and MSSIM $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$5, compared with ROSTF at PSNR $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$6 and MSSIM $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$7 (Isono et al., 15 Aug 2025).

A related but distinct line of work argues that standardization should precede spatio-temporal fusion. The standardization paper compares optimized upscaling of fine-resolution images with ABSIS, an anomaly-based sharpening method that blends the “overall features” of the fine-resolution time series with the “distinctive attributes” of a specific coarse-resolution image (Goyena et al., 17 Oct 2025). Injected into USTFIP, both approaches improve fusion accuracy, and ABSIS yields the largest reported gains: in New Cairo, RMSE decreases from $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$8 under baseline coarse harmonization to $\lambda_1 + \lambda_2 + \lambda_3 + \lambda_4 = 1$9, a $\lambda_1 = 0.3$0 reduction, while the spatial Edge metric improves from $\lambda_1 = 0.3$1 to $\lambda_1 = 0.3$2, corresponding to $\lambda_1 = 0.3$3 improvement (Goyena et al., 17 Oct 2025).

5. Other uses of SatFusion beyond image reconstruction

The term also appears in fusion problems whose outputs are not super-resolved satellite images. In ground-based space surveillance and tracking at the Pampilhosa da Serra Space Observatory, SatFusion describes a radar–optical concept built around a co-located LEO tracking radar and double wide-field optical telescope system, together with an existing deployable optical sensor for MEO and GEO surveillance (Coelho et al., 2022). The radar is monostatic at $\lambda_1 = 0.3$4 GHz, has beamwidth approximately $\lambda_1 = 0.3$5, and tracks LEO up to approximately $\lambda_1 = 0.3$6 km for objects with RCS $\lambda_1 = 0.3$7 cm$\lambda_1 = 0.3$8 at $\lambda_1 = 0.3$9 km. The optical system consists of two $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$00 cm telescopes with maximum FoV approximately $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$01, mount slewing up to $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$02/s, and a site at altitude approximately $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$03 m with dark-sky conditions and more than $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$04 clear nights per year. The sensors are separated by approximately $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$05 m and share a common time-stamping system (Coelho et al., 2022).

In that SST setting, SatFusion refers to real-time correlation of radar observables $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$06 and optical observables $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$07 through a joint nonlinear estimator such as an EKF, UKF, or batch WLS. The conceptual motivation is to extend observation arcs within a single pass, reduce latency in TLE generation, and improve initial orbit determination and reentry monitoring by combining precise radial information from radar with high-quality sky angles from optics (Coelho et al., 2022).

SATMapTR extends the term into autonomous-driving map construction, although the paper explicitly notes that it does not define a separate module named “SatFusion.” Instead, satellite-image fusion is realized through a Gated Feature Refinement module and a Geometry-Aware Fusion module (Huang et al., 12 Dec 2025). SATMapTR takes multi-view RGB cameras, produces BEV features via Lift-Splat-Shoot, extracts and refines satellite features with ResNet18 and hierarchical gated CNN blocks, then fuses $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$08 and $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$09 by strict grid-to-grid addition followed by an MLP:

$Pan \in \mathbb{R}^{H \times W \times C_{pan}}$10

The fused representation is decoded by MapTRv2 into vectorized lane dividers, pedestrian crossings, and road boundaries (Huang et al., 12 Dec 2025).

On nuScenes at the default $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$11 perception range, SATMapTR reports mAP $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$12 with per-class APs $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$13, compared with MapTRv2(C) at $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$14 in the reproduced baseline and MapTRv2(C+L) at $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$15 (Huang et al., 12 Dec 2025). At $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$16 m, it reports mAP $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$17, versus $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$18 for MapTRv2 and $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$19 for SatforHDMap; under fog, snow, FrameLost, CameraCrash, and low-light, the degradation is also smaller than the camera-only baseline (Huang et al., 12 Dec 2025). This usage broadens SatFusion from remote-sensing image enhancement toward multi-view geometric perception aided by overhead imagery.

6. Limitations, misconceptions, and open technical directions

A common misconception is that SatFusion always implies a single architecture or a single sensing geometry. The literature does not support that view. The 2025 SatFusion framework is one specific instantiation for multi-temporal LRMS plus PAN enhancement, whereas SatDiffMoE, TSSTF, ABSIS-assisted STIF, PASO radar–optical fusion, and SATMapTR each operationalize fusion under materially different assumptions, observation models, and outputs (Tong et al., 9 Oct 2025, Luo et al., 2024, Isono et al., 15 Aug 2025, Coelho et al., 2022, Huang et al., 12 Dec 2025).

Another misconception is that fusion methods in this area are uniformly physics-based. Several of the most recent approaches explicitly relax or omit explicit forward modeling. The SatFusion framework is presented as a general learnable fusion pipeline rather than a physically parameterized inverse model, and it does not specify physical image formation models such as $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$20 or $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$21 (Tong et al., 9 Oct 2025). SatDiffMoE likewise states that it does not explicitly invert $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$22, $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$23, or $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$24, and the authors identify the absence of physical measurement constraints as a limitation and future-work direction (Luo et al., 2024). By contrast, TSSTF makes the blur/downsampling operators $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$25 and $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$26 explicit and formulates the problem as a convex constrained optimization program, but its performance depends on accurate co-registration and an adequate sensor degradation model (Isono et al., 15 Aug 2025).

The principal limitations are therefore method-specific. The unified SatFusion framework does not yet leverage multi-temporal PAN, does not include explicit physical sensor modeling or deformable alignment, and does not report FLOPs or runtime (Tong et al., 9 Oct 2025). SatDiffMoE mitigates registration sensitivity through LPIPS-based latent fusion and relative time conditioning, but large parallax, major scene changes, and severe occlusion in most frames remain difficult (Luo et al., 2024). TSSTF can be challenged by genuine structural changes between dates because TGEC constrains edge locations, even though adaptive $Pan \in \mathbb{R}^{H \times W \times C_{pan}}$27 alleviates this to some extent (Isono et al., 15 Aug 2025). The standardization study notes that ABSIS is sensitive to aliasing in aggregated fine images, particularly for circular agricultural patterns, even though it still improves downstream USTFIP accuracy (Goyena et al., 17 Oct 2025). SATMapTR remains vulnerable to clouds, vegetation and building occlusions, outdated satellite imagery, and very large localization errors, although GFR and geometry-aware per-grid fusion mitigate modest misalignment (Huang et al., 12 Dec 2025). The PASO concept improves local fusion and latency but does not replace the need for a wider SST network, and PASO-specific radar and optical accuracy values remain to be established during commissioning and operations (Coelho et al., 2022).

Taken together, these limitations indicate that SatFusion is not a settled design space but an active research area spanning generative priors, convex inverse problems, sensor standardization, and cross-domain fusion architectures. A plausible implication is that future SatFusion systems will increasingly combine explicit degradation models, learned priors, and calibration-aware multi-modal conditioning rather than relying exclusively on any one of those paradigms.