SEN4X: Hybrid Super-Resolution for Sentinel-2

• SEN4X is a hybrid super-resolution method that combines multi-image and single-image techniques to enhance Sentinel-2 imagery.
• It fuses eight temporal acquisitions with a learned prior from high-resolution Pléiades Neo data to achieve a 4× upsampling to 2.5 m resolution.
• The method significantly improves urban land-cover segmentation accuracy compared to state-of-the-art approaches by balancing physical fidelity with learned texture details.

Searching arXiv for SEN4X and key related work to ground the article with current citations. SEN4X is a learned super-resolution method for Earth observation that upgrades Sentinel-2 imagery from its native 10 m ground sampling distance in the RGB+NIR bands to an effective 2.5 m ground sampling distance by combining multi-image super-resolution and single-image super-resolution in a single end-to-end architecture. It uses temporal oversampling from repeated Sentinel-2 acquisitions together with a learned prior from high-resolution Pléiades Neo data, and it is evaluated not primarily by image-level similarity metrics but by downstream urban land-cover segmentation in Hanoi, Vietnam. In the reported experiments, SEN4X leads to a significant performance improvement over state-of-the-art super-resolution baselines (Retnanto et al., 30 May 2025).

1. Definition and problem setting

SEN4X is presented as a hybrid super-resolution architecture for Sentinel-2 imagery that combines the advantages of single-image and multi-image techniques. The method operates on the 10 m RGB+NIR Sentinel-2 bands and targets a 4× upsampling in each spatial dimension, yielding output on a 2.5 m grid that is trained to be radiometrically consistent with 2.5 m downsampled Pléiades Neo imagery (Retnanto et al., 30 May 2025).

The central problem addressed by SEN4X is that Sentinel-2, despite being relatively high resolution by traditional Earth observation standards, remains too coarse to resolve many urban features. The reported examples include individual houses and small buildings at the 10–20 m scale, narrow streets and alleyways, hedgerows, narrow rivers, and small sealed surfaces. Pléiades Neo can resolve such objects, but is expensive and not globally nor regularly available. SEN4X is therefore positioned as a method for approaching some of the utility of very-high-resolution imagery while retaining the global coverage and revisit characteristics of Sentinel-2 (Retnanto et al., 30 May 2025).

The method is explicitly framed around two complementary paradigms. Multi-image super-resolution exploits temporal stacks of repeated acquisitions with sub-pixel shifts and independent noise to reconstruct physically supported high-frequency structure. Single-image super-resolution uses a high-capacity learned prior to infer plausible high-resolution textures and edges from a single input image. SEN4X combines these mechanisms so that physically grounded information from temporal oversampling and learned priors from cross-sensor supervision can be used jointly rather than separately (Retnanto et al., 30 May 2025).

A notable feature of the work is its evaluative emphasis. Rather than restricting assessment to peak signal-to-noise ratio, structural similarity, or other image quality metrics, the method is tested on urban land-cover classification. This design reflects the paper’s stated emphasis that super-resolution should be judged by its usefulness for downstream analysis rather than only by visual appearance or pixelwise similarity (Retnanto et al., 30 May 2025).

2. Data, supervision, and preprocessing pipeline

The Sentinel-2 inputs are Level-2A surface reflectance images accessed via SentinelHub. Only the 10 m red, green, blue, and near infrared bands are used; the 20 m and 60 m bands are excluded. The study relies on two satellites, S2A and S2B, in phase-shifted orbits with global 5-day revisit, enabling multiple observations of each location within a 2-year window around the very-high-resolution acquisition date (Retnanto et al., 30 May 2025).

Reflectance values are clipped to the 2nd–98th percentile and normalized to $[0,1]$. The imagery is split into 373 tiles, each 2.5 km × 2.5 km, corresponding to $158 \times 158$ pixels at 10 m. For each tile, 8 Sentinel-2 revisits are selected for multi-image super-resolution using three criteria: temporal proximity to the Pléiades Neo target date, completeness or cloud quality using the Sentinel-2 cloud and scene classification masks, and spectral quality measured by the number of pixels exceeding reflectance 0.8, which often indicates snow or cloud anomalies. This stack of eight revisits constitutes the temporal oversampling used by SEN4X (Retnanto et al., 30 May 2025).

The high-resolution supervision is derived from six Pléiades Neo top-of-atmosphere scenes covering Hanoi, obtained from Airbus OneAtlas. These scenes are divided into the same tiles as Sentinel-2 and normalized to $[0,1]$. A radiometric cross-calibration step is then applied: for each tile, the Pléiades Neo image is histogram-matched to the best Sentinel-2 image according to the same three quality criteria used for revisit selection. After alignment, Pléiades Neo is downsampled to 2.5 m using bilinear resampling to $632 \times 632$ pixels per tile, defining the target resolution for super-resolution (Retnanto et al., 30 May 2025).

The dataset is split into 70% training, 20% validation, and 10% test with geographical stratification. Two contiguous regions in the north and east are reserved entirely for testing to reduce spatial leakage. The area of interest is Hanoi, Vietnam, chosen specifically because dense urban morphology, narrow streets, water bodies, vegetation, and agricultural fields make 10 m sampling particularly inadequate for key objects (Retnanto et al., 30 May 2025).

For training, each Sentinel-2 tile is divided into $64 \times 64$ patches with stride 48, corresponding to 25% overlap. The matching high-resolution targets are $256 \times 256$ patches at 2.5 m. Masked pixels caused by clouds or defects are imputed by averaging valid reflectance values at the same location across the temporal stack. This preprocessing defines the paired learning problem as direct regression from 8 low-resolution 4-channel inputs to one high-resolution 4-channel target (Retnanto et al., 30 May 2025).

3. Hybrid MISR–SISR architecture

The SEN4X architecture is organized around a default early-fusion design in which multi-image fusion is performed at low resolution before a high-capacity single-image super-resolution backbone performs high-level reconstruction and upsampling. The low-resolution inputs are denoted $I_{LR}^{(t)} \in \mathbb{R}^{4 \times H \times W}$ for $t = 1,\dots,T$ with $T=8$, and the target is $I_{HR} \in \mathbb{R}^{4 \times sH \times sW}$ with upsampling factor $158 \times 158$0. The overall mapping is written as

$158 \times 158$1

This formalization makes explicit that SEN4X is a multi-view regression model rather than a single-image model with auxiliary metadata (Retnanto et al., 30 May 2025).

Each low-resolution input first passes through a shared shallow feature extractor consisting of a single $158 \times 158$2 convolution. This produces low-level feature maps that encode edges and color gradients. The channel dimension is 258, matching the embedding dimension used by the Swin2SR-derived backbone (Retnanto et al., 30 May 2025).

The eight shallow feature tensors are then recursively fused using a HighResNet-style multi-image super-resolution module. Fusion proceeds pairwise. Each pair is first processed by a two-layer convolutional residual block applied separately to both feature maps, and then merged by a residual convolution operating on their concatenation. The same fusion block is shared across all recursive merges. This parameter sharing makes the design relatively efficient and order-invariant in spirit, although the implementation uses a fixed pairing order. The result is a single fused low-resolution feature map that integrates the temporal oversampling signal from all eight acquisitions (Retnanto et al., 30 May 2025).

The fused representation is then processed by a modified Swin2SR-style single-image backbone. This backbone contains 6 Residual Swin Transformer Blocks with shifted-window self-attention, window size 8, 6 attention heads per layer, and embedding dimension 258. The Swin-based design is used to capture non-local spatial context and complex textures that are less accessible to purely convolutional models. After the transformer stack, a final $158 \times 158$3 convolution expands the channels to $158 \times 158$4, and a pixel shuffle layer rearranges them into a $158 \times 158$5 output (Retnanto et al., 30 May 2025).

No explicit geometric warping or optical flow is used. Alignment is handled implicitly by the convolutional fusion layers, which must learn to reconcile misaligned content during training. The method is therefore distinct from approaches that rely on explicit registration pipelines or analytically specified sub-pixel motion models (Retnanto et al., 30 May 2025).

A late-fusion variant is also evaluated. In that alternative, each of the 8 inputs is passed independently through shallow feature extraction and the full Swin2SR-style backbone, and only the resulting deep features are recursively fused. The paper reports that this variant is less effective and substantially more expensive, indicating that low-level temporal fusion is the preferable architecture in the reported setting (Retnanto et al., 30 May 2025).

4. Training objective and methodological choices

SEN4X is trained directly on real cross-sensor pairs rather than on synthetically degraded high-resolution images. The paper contrasts this with the common observation model

$158 \times 158$6

where $158 \times 158$7 denotes blur, $158 \times 158$8 downsampling, and $158 \times 158$9 sensor noise. Instead of specifying a parametric degradation operator $[0,1]$0, the method uses actual Sentinel-2 observations and downsampled Pléiades Neo targets, allowing the network to learn the effective cross-sensor mapping directly (Retnanto et al., 30 May 2025).

The training objective is a pure $[0,1]$1 reconstruction loss: $[0,1]$2 No perceptual loss, adversarial loss, or explicit sensor-consistency loss is used in SEN4X. This differentiates it from ESRGAN-style baselines that employ GAN objectives to emphasize perceptual realism (Retnanto et al., 30 May 2025).

All super-resolution models in the comparison, including SEN4X and the baselines, are trained under the same optimization setting. The implementation uses PyTorch 2.4, the Adam optimizer, an initial learning rate of $[0,1]$3, linear warm-up followed by cosine annealing, 100 training epochs, and a batch size of 4 patches per iteration. Training is reported on a single NVIDIA L4 GPU. No staged pretraining is used; the MISR and SISR components are optimized end-to-end from the start (Retnanto et al., 30 May 2025).

The methodological design of SEN4X reflects a specific compromise between physical fidelity and learned prior strength. Pure MISR can exploit genuine sub-pixel offsets across repeated acquisitions, but is susceptible to blur under imperfect alignment or averaging. Pure SISR can generate sharp details, but these may be hallucinated rather than physically supported. SEN4X seeks to use temporal information where it exists and learned priors where temporal evidence is weak. This suggests a model class in which multi-image and single-image super-resolution should not be treated as mutually exclusive but as complementary sources of information (Retnanto et al., 30 May 2025).

5. Evaluation framework and empirical results

The core evaluation task is land-cover classification at 2.5 m. A segmentation model is trained on true Pléiades Neo images and ground-truth labels only, and is then applied at test time to different image sources: true Pléiades Neo, bicubic-upsampled Sentinel-2, and the outputs of several learned super-resolution models. This protocol isolates the effect of the super-resolution method on a practical downstream task because the segmentation network remains fixed while the input imagery changes (Retnanto et al., 30 May 2025).

The segmentation model uses the SATLAS foundation model as a Swin-based encoder, modified from 3 to 4 input channels by initializing the NIR channel weights as the average of the RGB weights. The architecture includes Swin v2 feature maps at $[0,1]$4, $[0,1]$5, $[0,1]$6, and $[0,1]$7 of the $[0,1]$8 input, a Feature Pyramid Network neck, and a U-Net-style decoder. It is trained with masked cross-entropy on valid labeled pixels only, using Adam with initial learning rate $[0,1]$9, cosine annealing to $632 \times 632$0, batch size 16, up to 1000 epochs, and early stopping after 25 epochs without improvement (Retnanto et al., 30 May 2025).

The downstream labels comprise 7 classes: buildings, sealed surfaces, water bodies, forest, grassland, cropland, and bare soil. Labels are manually annotated in QGIS on native-resolution Pléiades Neo imagery, with Google Open Buildings used to guide building delineation. To avoid mixed pixels and label ambiguity, only pixels whose neighborhood is fully covered by a single class are retained after rasterization and downsampling to 2.5 m (Retnanto et al., 30 May 2025).

The paper reports overall accuracy, mean IoU, and micro IoU, with each super-resolution model trained five times using different random seeds and results summarized as mean ± standard deviation. The upper bound using true Pléiades Neo is Accuracy 0.856 and mIoU 0.663. Bicubic Sentinel-2 yields Accuracy 0.440 and mIoU 0.278. Among learned methods, SEN4X early fusion achieves Accuracy $632 \times 632$1, mIoU $632 \times 632$2, and micro IoU $632 \times 632$3, outperforming Swin2SR, HighResNet, ESRGAN, and the SEN4X late-fusion variant (Retnanto et al., 30 May 2025).

The reported performance table is as follows:

Method	Accuracy	mIoU
PNEO	0.856	0.663
Bicubic Sentinel-2	0.440	0.278
Swin2SR	$632 \times 632$4	$632 \times 632$5
HighResNet	$632 \times 632$6	$632 \times 632$7
ESRGAN	$632 \times 632$8	$632 \times 632$9
SEN4X early fusion	$64 \times 64$0	$64 \times 64$1
SEN4X late fusion	$64 \times 64$2	$64 \times 64$3

These results show that the early-fusion SEN4X variant improves mIoU by 2.7 points over Swin2SR, 12.9 points over HighResNet, and 2.3 points over ESRGAN. The paper further notes that SISR-only Swin2SR outperforms MISR-only HighResNet, indicating that simple MISR can be too blurry for segmentation despite higher PSNR (Retnanto et al., 30 May 2025).

Per-class accuracies for SEN4X early fusion are reported as follows: buildings $64 \times 64$4, sealed surfaces $64 \times 64$5, water $64 \times 64$6, forest $64 \times 64$7, grassland $64 \times 64$8, cropland $64 \times 64$9, and bare soil $256 \times 256$0. The paper highlights that improvements are especially notable for buildings and sealed surfaces, which are critical urban classes (Retnanto et al., 30 May 2025).

6. Image-quality metrics, ablations, and limitations

In addition to downstream segmentation, SEN4X is evaluated with PSNR, SSIM, LPIPS, and OpenSR-Test metrics including hallucination, omission, and improvement. The image metrics do not rank methods in the same order as the downstream task. HighResNet achieves the highest PSNR at 16.968, while SEN4X early fusion yields 16.676. SSIM differences are small across learned methods, clustering around approximately 0.415–0.419. LPIPS is lowest for SEN4X early fusion at 0.444, followed by ESRGAN at 0.459 and Swin2SR at 0.470, while HighResNet is worst at 0.490 (Retnanto et al., 30 May 2025).

The OpenSR-Test metrics further illustrate this mismatch. HighResNet has the lowest hallucination score, 0.270, indicating a more conservative reconstruction with fewer added details, whereas Swin2SR has the highest hallucination score at 0.299. SEN4X early fusion lies between these extremes at 0.286, but achieves the best omission score, 0.331, and the highest improvement score, 0.383. The reported interpretation is that PSNR and hallucination metrics can favor smoothness and penalize incorrect high-frequency content while failing to capture the semantic harm caused by blur. LPIPS correlates better with land-cover performance, though still imperfectly (Retnanto et al., 30 May 2025).

Qualitative comparisons reported in the paper are consistent with the quantitative findings. Bicubic upsampling is described as heavily blurred, with washed-out edges and invisible small buildings. Swin2SR appears sharp but sometimes oversharpened, with artifacts or hallucinations. HighResNet appears smooth and blurred, with fewer artifacts but lower detail. ESRGAN appears sharp and vivid but can produce inconsistent patterns. SEN4X early fusion is described as sharp with more consistent textures and better continuity of roads, roofs, and water edges, and its resulting land-cover maps better recover boundaries for buildings, sealed surfaces, and roads, particularly in dense urban cores (Retnanto et al., 30 May 2025).

The architectural ablation between early and late fusion is especially important. Early fusion outperforms late fusion by approximately 1.6 mIoU points. It is also far cheaper computationally. Inference time per $256 \times 256$1 patch on a T4 GPU is reported as 133.6 ms for Swin2SR, 65.1 ms for HighResNet, 31.9 ms for ESRGAN, 189.6 ms for SEN4X early fusion, and 1471.5 ms for SEN4X late fusion. Parameter counts are 24.5M for Swin2SR, 13.0M for HighResNet, 16.7M for ESRGAN, and 30.5M for both SEN4X variants, with about 24M of SEN4X’s parameters located in the SISR backbone (Retnanto et al., 30 May 2025).

The paper explicitly identifies several limitations. All data come from Hanoi, and no cross-region generalization study is reported. Only four Sentinel-2 bands are used, excluding the 20 m and 60 m bands because no co-registered very-high-resolution references exist for those wavelengths. Only one downstream task, land-cover segmentation, is evaluated. The method depends on expensive high-quality Pléiades Neo supervision. Sensitivity to cloud contamination and seasonality is not analyzed in detail. The model is also heavier than Swin2SR or ESRGAN, with non-trivial inference cost for large-scale deployment (Retnanto et al., 30 May 2025).

7. Position within Sentinel-2 super-resolution research

Within the broader super-resolution literature for remote sensing, SEN4X is characterized by three methodological choices. First, it uses real cross-sensor supervision rather than synthetic low-resolution generation. Second, it combines a dedicated MISR module with a high-capacity SISR backbone rather than committing to a purely temporal or purely single-image design. Third, it evaluates methods by downstream urban land-cover performance rather than relying principally on image similarity scores (Retnanto et al., 30 May 2025).

The paper compares SEN4X with several classes of baseline. Bicubic upsampling provides a learning-free reference. Swin2SR represents a strong SISR transformer baseline. HighResNet represents recursive-fusion MISR. ESRGAN, in the SATLAS-style variant used here, stacks the eight Sentinel-2 revisits across channels and applies a GAN-based architecture as a hybrid alternative. SEN4X differs from that stack-and-SISR strategy by preserving a dedicated temporal fusion mechanism before high-capacity single-image reconstruction (Retnanto et al., 30 May 2025).

The empirical findings lead to several technical implications. MISR alone does not guarantee better downstream performance than SISR, even when it yields better PSNR and lower hallucination. A strong learned prior can be more useful than conservative temporal averaging when the downstream task depends on crisp boundaries and small-object recognition. At the same time, SISR-only approaches can introduce artifacts that reduce physical realism. SEN4X’s hybrid formulation is reported to achieve the best balance between these tendencies by combining physically grounded details with a strong reconstruction prior (Retnanto et al., 30 May 2025).

The future directions proposed in the paper are geographically broader evaluation, inclusion of additional Sentinel-2 bands such as red-edge and SWIR, extension to downstream tasks beyond land-cover segmentation, adaptation to other sensors such as Landsat and PlanetScope, and improved temporal modeling through explicit alignment, deformable convolutions, or spatiotemporal transformers. A plausible implication is that SEN4X is best understood not as a terminal architecture but as a design template for hybrid super-resolution in Earth observation, particularly where repeated acquisitions and sparse high-resolution supervision are both available (Retnanto et al., 30 May 2025).