Deep Pushbroom Super-Resolution (DPSR)

• The paper introduces a neural network that aligns with pushbroom physics by processing hyperspectral lines sequentially using a causal memory mechanism.
• It achieves a quality-efficiency tradeoff, delivering competitive reconstruction metrics while using significantly lower memory and computational resources than traditional 2D/3D models.
• The design is optimized for onboard, real-time inference under strict satellite constraints, leveraging causal interpolation and a lightweight state-space memory.

Searching arXiv for the DPSR and related satellite super-resolution papers to ground the article with current records. Deep Pushbroom Super-Resolution (DPSR) is a neural network architecture for single-image hyperspectral super-resolution (HSI-SR) that is explicitly designed for onboard, real-time inference on satellites equipped with pushbroom sensors. Its defining premise is architectural alignment with acquisition physics: instead of processing a full $2$D image or buffered tile, DPSR processes the hyperspectral signal line by line in the along-track direction, using a causal memory mechanism to exploit previously acquired lines while never requiring future lines. In the reported experiments, this design yields a markedly lower memory footprint and lower computational complexity than conventional $2$D/$3$D CNN and transformer-based HSI-SR models, while remaining competitive in reconstruction quality and, for the smaller variant, approaching or meeting the throughput needed to super-resolve a line before the next line is acquired (Piccinini et al., 28 Jul 2025).

1. Problem domain and pushbroom acquisition model

DPSR addresses a central tradeoff in spaceborne hyperspectral imaging: hyperspectral imagers provide fine spectral signatures that are essential for material discrimination, but they typically do so at limited spatial resolution. The reported motivation is practical rather than purely algorithmic. Recent spaceborne hyperspectral missions such as PRISMA, EnMAP, and HySIS are described as operating around $30\ \mathrm{m/pixel}$, so improving spatial resolution is positioned as a preprocessing step for downstream tasks including land cover mapping, vegetation assessment, mineral detection, and change monitoring (Piccinini et al., 28 Jul 2025).

The method is formulated for the pushbroom sensing regime. In that regime, the instrument acquires one line at a time, containing all across-track pixels and all spectral channels for that line, while platform motion provides successive lines in the along-track direction. Accordingly, the incoming low-resolution observation at time $y$ is a line $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$, where $W$ is the across-track width and $C$ is the number of spectral channels. This differs fundamentally from the assumptions made by most image super-resolution pipelines, which presuppose immediate access to a full $2$D image or at least to a spatial tile (Piccinini et al., 28 Jul 2025).

The paper’s operational target is onboard AI under stringent payload constraints. The relevant constraints are limited power, limited memory, and limited compute throughput, which are especially acute for hyperspectral data because of their joint spatial and spectral dimensionality. DPSR is therefore not merely a reconstruction model; it is a streaming inference design whose real-time criterion is explicit: it should “super-resolve a line in the time it takes to acquire the next one.” For the PRISMA VNIR instrument, the line acquisition time is reported as $4.32\ \mathrm{ms}$ in the abstract and approximately $2$0 in the runtime section (Piccinini et al., 28 Jul 2025).

A common misconception is to read DPSR as a conventional hyperspectral SR network that happens to be evaluated on embedded hardware. That characterization is incomplete. The defining claim is that the model’s causal structure, memory layout, and output schedule are chosen to match the pushbroom data stream itself, rather than to retrofit a full-image SR backbone to a constrained deployment setting.

2. Causal interpolation formulation

DPSR treats pushbroom HSI-SR as a causal interpolation problem. For a super-resolution factor $2$1, the model does not attempt to extrapolate beyond the latest observation. Instead, once line $2$2 has been acquired, it estimates the $2$3 high-resolution lines located between low-resolution lines $2$4 and $2$5. The paper emphasizes that this interpolation framing is easier and more accurate than extrapolation (Piccinini et al., 28 Jul 2025).

This design determines both the information set available at inference time and the shape of the output. When processing line $2$6, DPSR has access to the current line $2$7, the previous line $2$8 for bilinear interpolation, and a compact memory state summarizing previously processed lines. It does not have access to any future lines. A low-resolution input line of size $2$9 yields an output of size $3$0, corresponding to $3$1 super-resolved lines and an across-track upscaling by $3$2 (Piccinini et al., 28 Jul 2025).

The final prediction is residual with respect to bilinear interpolation: $3$3 The learned component therefore corrects a lightweight interpolation baseline rather than replacing it entirely (Piccinini et al., 28 Jul 2025).

This formulation has a direct consequence for training and evaluation at sequence boundaries. Because the model emits lines between $3$4 and $3$5, the first $3$6 lines of the network output and the last $3$7 lines of the ground truth are discarded. The paper states that these discarded regions correspond, respectively, to positions before the start of the image and positions that would require a line after the end of the image. It also states that in a real streaming scenario, no line would need to be discarded because the input stream would be continuous (Piccinini et al., 28 Jul 2025).

The causal interpolation perspective is central to understanding both DPSR’s efficiency and its limits. It avoids the latency and memory cost of buffering future context, but it also relinquishes access to potentially informative future lines that non-causal $3$8D models can exploit.

3. Network architecture and state-space memory

The architecture comprises four main components: a Shallow Feature Extraction (SFE) block, two Cross-Line Feature Fusion (CLFF) blocks, an upsampler, and the residual bilinear interpolation path. Spatially, the network processes one line at a time; temporally, it links lines through a recurrent state (Piccinini et al., 28 Jul 2025).

The SFE block uses $3$9D operations over the across-track direction, treating spectral bands as channels. It contains a $30\ \mathrm{m/pixel}$0D convolution with kernel size $30\ \mathrm{m/pixel}$1, layer normalization, SiLU activation, and a Channel Attention Block. The input line is transformed into $30\ \mathrm{m/pixel}$2, where $30\ \mathrm{m/pixel}$3 is the internal feature dimension. In the paper’s interpretation, this stage removes noise and extracts shallow features that jointly represent local across-track structure and spectral correlations (Piccinini et al., 28 Jul 2025).

The backbone then applies two CLFF blocks: $30\ \mathrm{m/pixel}$4 Each CLFF block contains a NAFBlock and a Mamba block. The NAFBlock is described as inspired by NAFNet and uses separable convolutions, specifically a size-$30\ \mathrm{m/pixel}$5 $30\ \mathrm{m/pixel}$6D convolution followed by a depthwise $30\ \mathrm{m/pixel}$7D convolution, together with SimpleGate and Simplified Channel Attention. Its role is local feature processing within the current line, enlarging the receptive field in the across-track direction while deepening spectral-spatial features efficiently (Piccinini et al., 28 Jul 2025).

The Mamba block is the core causal memory mechanism. After NAFBlock processing, the features of a line are treated as $30\ \mathrm{m/pixel}$8 tokens evolving over the sequence of lines. The block first expands the feature dimension from $30\ \mathrm{m/pixel}$9 to $y$0, applies a causal convolution of kernel size $y$1 along the along-track direction, and then applies SiLU. The resulting features $y$2 drive a selective state-space model (SSM) update: $y$3

$y$4

Here $y$5 is the latent state at line $y$6, $y$7 is the processed current-line feature, and $y$8 is the SSM output. The latent state has shape

$y$9

with $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$0 the SSM state size (Piccinini et al., 28 Jul 2025).

The memory cost is therefore concentrated in two structures: the causal convolution buffer of size $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$1 and the SSM latent state of size $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$2. A concrete example is provided: with $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$3, $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$4, $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$5, and $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$6, the total memory to store these tensors in single precision is about $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$7 (Piccinini et al., 28 Jul 2025).

The upsampler is deliberately lightweight. It consists of a $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$8D convolution expanding channels to $$\mathbf{x}^{\mathrm{LR}}_y \in \mathbb{R}^{W \times C}$$9, Pixel Shuffle, and a final $W$0D convolution restoring the channel count to $W$1. The parameter $W$2 denotes the number of features remaining after Pixel Shuffle, and the paper relates its choice to the tradeoff between low FLOPs and preserving enough spectral information (Piccinini et al., 28 Jul 2025).

A defining architectural feature is the asymmetric treatment of dimensions. The across-track spatial dimension is handled explicitly with $W$3D convolutions, the spectral dimension is treated as channels and fused jointly with spatial structure, and the along-track spatial dimension is modeled as a causal sequence using the Mamba-based SSM recurrence. This decoupling is the core of the “pushbroom” design principle.

4. Training protocol, datasets, and implementation regime

The reported base model uses internal feature dimension $W$4, Mamba expansion factor $W$5, SSM state dimension $W$6, upsampler reduced-feature factor $W$7, and causal convolution kernel size $W$8. A smaller variant, DPSR-S, reduces the internal feature dimension to $W$9 (Piccinini et al., 28 Jul 2025).

Optimization uses Adam with a fixed learning rate of $C$0 and no weight decay. Training runs for a variable number of epochs until convergence or overfitting. The paper states that DPSR uses the same loss as CST, with $C$1 and $C$2, but it does not reproduce the explicit loss formula in the provided text (Piccinini et al., 28 Jul 2025). This omission matters for reproducibility. A plausible implication is that architectural analysis is substantially better specified than the exact optimization objective.

Training is supervised and relies on synthetic low-resolution data generated from high-resolution hyperspectral imagery through bicubic downsampling. For HySpecNet-11k, the inputs are $C$3 for $C$4 and $C$5 for $C$6. The dataset contains $C$7 non-overlapping image patches of size $C$8, with $C$9 usable bands, ground sampling distance $2$0, and a $2$1 train/validation/test split. The experiments use the hard split rather than the easy patchwise split. The paper notes that approximately $2$2 of images contain zeroed-out channels; these were kept during training, predicted at test time, and excluded from metric computation for those bands (Piccinini et al., 28 Jul 2025).

Additional experiments are conducted on Houston, Pavia, and Chikusei. For HySpecNet-11k, “classic augmentations” increase the training sample count by a factor of $2$3, although the specific augmentations are not enumerated in the provided text. Training and testing are performed on one Nvidia TITAN RTX GPU (Piccinini et al., 28 Jul 2025).

The dependence on synthetic bicubic degradation is one of the clearest limitations of the reported setup. Unlike real sensor-paired low-resolution/high-resolution benchmarks, the training protocol assumes a simplified degradation model. This suggests that deployment performance could depend on the gap between bicubic simulation and actual sensor degradations.

5. Quantitative performance, efficiency, and onboard viability

DPSR’s reported contribution is not maximal reconstruction quality in isolation, but a quality-efficiency tradeoff that remains competitive against much more expensive methods. On HySpecNet-11k at $2$4, DPSR achieves MPSNR $2$5, MSSIM $2$6, SAM $2$7, and RMSE $2$8. These values exceed those of GDRRN, SSPSR, SNLSR, and EUNet, while trailing MSDformer at MPSNR $2$9 and CST at $4.32\ \mathrm{ms}$0 (Piccinini et al., 28 Jul 2025). At $4.32\ \mathrm{ms}$1, DPSR reaches MPSNR $4.32\ \mathrm{ms}$2, MSSIM $4.32\ \mathrm{ms}$3, SAM $4.32\ \mathrm{ms}$4, and RMSE $4.32\ \mathrm{ms}$5, again outperforming several CNN and unfolding baselines while remaining below the strongest transformer-based models (Piccinini et al., 28 Jul 2025).

The Houston results are especially close to the best baseline. At $4.32\ \mathrm{ms}$6, DPSR attains MPSNR $4.32\ \mathrm{ms}$7, MSSIM $4.32\ \mathrm{ms}$8, SAM $4.32\ \mathrm{ms}$9, and RMSE $2$00, versus CST at $2$01, $2$02, $2$03, and $2$04. At $2$05, DPSR records $2$06, $2$07, $2$08, and $2$09, again close to CST’s $2$10, $2$11, $2$12, and $2$13 (Piccinini et al., 28 Jul 2025).

On Pavia and Chikusei, the pattern is similar but somewhat more mixed. At Pavia $2$14, DPSR’s MPSNR $2$15, MSSIM $2$16, and RMSE $2$17 are close to CST’s $2$18, $2$19, and $2$20, but its SAM $2$21 is worse than CST’s $2$22. On Chikusei $2$23, DPSR reaches MPSNR $2$24, MSSIM $2$25, SAM $2$26, and RMSE $2$27, essentially matching MSDformer and remaining within $2$28 of CST (Piccinini et al., 28 Jul 2025). The paper’s qualitative example on HySpecNet-11k further states that DPSR does not introduce visible artifacts attributable to line-by-line processing (Piccinini et al., 28 Jul 2025).

The efficiency results are more distinctive. For HySpecNet-11k at $2$29, DPSR uses $2$30 parameters and $2$31 FLOPs/pixel, compared with SSPSR at $2$32 and $2$33, MSDformer at $2$34 and $2$35, CST at $2$36 and $2$37, EUNet at $2$38 and $2$39, SNLSR at $2$40 and $2$41, and GDRRN at $2$42 and $2$43 (Piccinini et al., 28 Jul 2025). At $2$44, DPSR uses $2$45 parameters and $2$46 FLOPs/pixel (Piccinini et al., 28 Jul 2025).

The memory results reinforce the same point. The paper states that DPSR uses $2$47 to $2$48 less memory than other methods, and that for a $2$49 image representative of a PRISMA VNIR frame, DPSR can process the input with less than $2$50 of memory, whereas competing state-of-the-art models exceeded the $2$51 VRAM of the GPU used for those tests (Piccinini et al., 28 Jul 2025). Because processing is line-based, memory usage is said to be constant with the number of lines by design and to grow only weakly with the number of columns $2$52.

The onboard experiments are conducted on an Nvidia Jetson Orin Nano in $2$53 mode, in half precision and without further optimizations. For DPSR-S, the line processing time is $2$54 at $2$55 and $2$56 at $2$57. For the larger DPSR, it is $2$58 at $2$59 and $2$60 at $2$61 (Piccinini et al., 28 Jul 2025). Since the PRISMA VNIR line acquisition time is quoted as approximately $2$62 in this section, DPSR-S at $2$63 meets the real-time criterion, DPSR-S at $2$64 is slightly above the strict threshold, and the larger DPSR remains faster than existing baselines but does not meet PRISMA-rate real time in the reported unoptimized implementation (Piccinini et al., 28 Jul 2025).

The ablation studies support the Mamba-based memory choice. Replacing the Mamba block with simple causal convolution yields MPSNR $2$65, $2$66 parameters, and $2$67 FLOPs/pixel; with LSTM, the corresponding values are $2$68, $2$69, and $2$70; with Mamba-2, $2$71, $2$72, and $2$73. The reported Mamba configuration achieves $2$74, $2$75, and $2$76 (Piccinini et al., 28 Jul 2025). This suggests that, within the tested design space, selective SSM memory is the most favorable balance of quality and cost.

6. Relation to satellite super-resolution literature, scope, and limitations

DPSR belongs to the broader family of remote-sensing super-resolution methods, but its problem formulation differs sharply from multi-frame satellite SR. A useful contrast is “HighRes-net: Recursive Fusion for Multi-Frame Super-Resolution of Satellite Imagery,” which reconstructs a single high-resolution image from a set of low-resolution views of the same site acquired over time, using implicit co-registration, recursive fusion over an arbitrary number of views, and a registered loss implemented through ShiftNet and differentiable Lanczos shifting (Deudon et al., 2020). That framework is not pushbroom-specific; it assumes an unordered set of temporally separated observations, not a causal line stream. Its design addresses multi-view aliasing, latent alignment, and fusion-before-upsampling. DPSR, by contrast, is a single-image hyperspectral SR method whose central innovation is matching pushbroom acquisition with a line-by-line causal state (Piccinini et al., 28 Jul 2025).

This distinction matters because “deep pushbroom super-resolution” can be misunderstood as a generic label for any satellite SR model applied to line-scanning data. In the narrower technical sense established by the paper, DPSR is not a multi-frame fusion method, does not rely on explicit registration among multiple revisits, and does not assume future context. Its setting is closer to streaming single-image SR under hardware constraints than to revisit-based Earth observation reconstruction (Piccinini et al., 28 Jul 2025).

The paper also makes several limitations explicit or implicit. DPSR gives up access to future lines, which is the main tradeoff behind its causal efficiency and may limit quality relative to non-causal full-image methods. Its design is specifically tailored to pushbroom line-scanning sensors and is therefore less natural for frame-based imagers. The interpolation formulation requires special boundary handling in finite test images. The training protocol uses bicubic downsampling rather than real low-resolution/high-resolution sensor pairs. Some implementation details, including the full loss formula, exact epoch counts, and batch size, are omitted from the provided text. Real-time $2$77 SR is not fully achieved in the reported implementation, and spectral-angle performance can lag on some datasets, as seen on Pavia $2$78 (Piccinini et al., 28 Jul 2025).

Within those bounds, DPSR reframes HSI-SR as a sensor-matched streaming inference problem. Its significance lies less in establishing a universally superior reconstruction model than in demonstrating that causal line-wise super-resolution, with a compact state-space memory and lightweight upsampling head, can make onboard hyperspectral SR plausible under low-power satellite constraints. The future direction stated in the paper is to integrate super-resolution directly with detection or segmentation, yielding an all-in-one super-resolved inference model that also follows pushbroom acquisition in real time (Piccinini et al., 28 Jul 2025).