Abundance Super-Resolution Neural Network

• Abundance Super-Resolution Neural Network is a neural architecture that enhances hyperspectral abundance maps by incorporating domain-specific constraints like nonnegativity and sum-to-one.
• The network employs a scale-recurrent design with multi-residual dense blocks and sub-pixel convolution, ensuring efficient feature propagation and accurate high-resolution estimation.
• Empirical evaluations on urban hyperspectral datasets demonstrate that the method outperforms traditional techniques, highlighting its potential in unsupervised super-resolution tasks.

An Abundance Super-Resolution Neural Network (AbSRNet) is a neural architecture designed to address the problem of spatial resolution enhancement for abundance maps in hyperspectral imaging and, more generally, for multichannel images where the channel structure encodes material fractions, quantitative decompositions, or analogous semantic attributes. Unlike standard single-image super-resolution networks that focus on general image textures, AbSRNets leverage domain priors such as nonnegativity and sum-to-one constraints, structurally tailored for the problem of hyperspectral unmixing.

1. Mathematical Formulation and Problem Context

In hyperspectral remote sensing, each low-resolution (LR) pixel can be modeled as a convex mixture of $N$ endmembers (pure spectral signatures), given as

$X = EA + N$

where $X \in \mathbb{R}^{L \times (h \cdot w)}$ is the stacked hyperspectral image, $E \in \mathbb{R}^{L \times N}$ is the endmember matrix, $A \in \mathbb{R}^{N \times (h \cdot w)}$ is the abundance matrix, and $N$ is additive noise. The abundance vectors $a_{:,i}$ for each pixel satisfy

$a_{:,i} \geq 0,\quad \sum_{n=1}^N a_{n,i} = 1$

imposing the Abundance Nonnegativity Constraint (ANC) and the Abundance Sum-to-one Constraint (ASC) (Xu et al., 23 Jan 2026).

The goal is to estimate a high-resolution (HR) abundance map $\hat{A}_{HR} \in \mathbb{R}^{N \times H \times W}$ (with $H>h$, $X = EA + N$0), and then reconstruct the HR hyperspectral image as

$X = EA + N$1

even in the absence of any real high-resolution ground-truth data.

2. Network Architecture and Scale-Recurrent Design

The AbSRNet typically leverages a scale-recurrent architecture, in which a single set of network weights is reused at successively higher scales. This approach, rooted in work on scale-recurrent dense networks, enables multi-factor super-resolution while maintaining parameter efficiency (Purohit et al., 2022).

Key structural modules include:

• Shallow feature extractor: Two $X = EA + N$2 convolutional layers with ReLU activations.
• Residual Dense Blocks (RDBs) / Multi-Residual Dense Blocks (MRDBs): Each RDB comprises $X = EA + N$3 convolutional layers (growth rate $X = EA + N$4), where each layer receives the concatenation of all previous feature outputs. MRDBs introduce additional 1$X = EA + N$51 shortcut connections from the block input $X = EA + N$6 to each intermediate layer, enhancing gradient propagation.
• Global Feature Fusion (GFF): Aggregates outputs from multiple RDBs.
• Upsampling module: Utilizes sub-pixel convolution (PixelShuffle) to increase spatial resolution by factor 2 (or higher in repeated recurrences).
• Final reconstruction and constraint enforcement: A $X = EA + N$7 convolution outputs the HR abundance estimate, followed by a pixel-wise softmax ensuring the ASC:

$X = EA + N$8

This ensures $X = EA + N$9, $X \in \mathbb{R}^{L \times (h \cdot w)}$0 and nonnegativity (Xu et al., 23 Jan 2026).

3. Synthetic Training Data via the Dead-Leaves Model

A defining feature of state-of-the-art AbSRNet pipelines for hyperspectral super-resolution is the unsupervised training paradigm that leverages fully synthetic data. This uses the dead-leaves model, which generates HR abundance fields by stochastically dropping random rectangles ("leaves")—each parameterized by size, orientation, and abundance value—on a canvas, followed by normalization to enforce the ANC and ASC (Xu et al., 23 Jan 2026).

Corresponding LR samples are obtained by Gaussian blurring (PSF, $X \in \mathbb{R}^{L \times (h \cdot w)}$1) and bicubic downsampling by a factor of 4. A local-variation layer is further added to introduce sub-pixel disuniformities. The resulting synthetic datasets faithfully reproduce the spatial statistics of real abundance maps in urban scenes, providing an effective training set in the absence of HR ground-truth.

4. Training Paradigms and Loss Functions

For training on synthetic pairs $X \in \mathbb{R}^{L \times (h \cdot w)}$2, AbSRNets employ an $X \in \mathbb{R}^{L \times (h \cdot w)}$3-reconstruction objective:

$X \in \mathbb{R}^{L \times (h \cdot w)}$4

where $X \in \mathbb{R}^{L \times (h \cdot w)}$5 is the neural mapping. The Adam optimizer is employed, and no explicit regularization is used beyond the ASC-enforcing softmax. The GAN-based extensions described for general image domains—combining pixel-wise, deep-feature (VGG), and adversarial losses—can be applied when perceptual realism is prioritized, with respective weightings $X \in \mathbb{R}^{L \times (h \cdot w)}$6 (Purohit et al., 2022).

Performance can thus be flexibly traded along the perception–distortion curve by adjusting these loss weights. A plausible implication is that similar multi-loss strategies could enhance fidelity in scenarios beyond classical abundance SR.

5. Quantitative Performance and Benchmarking

Empirical evaluation on the Urban hyperspectral dataset (307$X \in \mathbb{R}^{L \times (h \cdot w)}$7307 pixels, 162 spectral bands, 6 materials) demonstrates that AbSRNets—specifically RDN-DL variants trained exclusively on synthetic dead-leaves maps (6$X \in \mathbb{R}^{L \times (h \cdot w)}$8500$X \in \mathbb{R}^{L \times (h \cdot w)}$9500)—outperform several supervised single-image SR baselines (Xu et al., 23 Jan 2026).

Method	Mean PSNR (dB)	Mean SAM (°)	Mean ERGAS (%)
Bicubic	26.49	13.88	7.33
MCNet	27.48	12.45	6.58
SSPSR	26.38	13.67	7.26
HSISR	27.55	12.33	6.51
RDN-DL	27.78	12.14	6.37

RDN-DL achieves the best or competitive performance across all metrics, despite never accessing real HR data during training. This suggests that the combination of synthetic priors and adapted dense network architectures captures sufficient spatial statistics for effective super-resolution.

6. Algorithmic Structure and Implementation Recipes

The AbSRNet forward pass for multi-scale super-resolution is controlled by an outer scale-recurrent loop. Pseudocode for the full architecture illustrates its recursive execution:

$E \in \mathbb{R}^{L \times N}$0 For each MRDB: $E \in \mathbb{R}^{L \times N}$1 This structure realizes the dense, multi-residual connections crucial for propagating abundant local features and stable gradient flow (Purohit et al., 2022).

7. Connections, Extensions, and Future Implications

The AbSRNet design encapsulates a general blueprint for resolution enhancement tasks where structural or physical priors (positivity, simplex constraints, spatial statistics) govern channel relationships. The integration of realistic synthetic data generation (e.g., dead-leaves), constraint-enforcing output layers, and scale-recurrent, dense network blocks enables robust unsupervised performance—even outperforming supervised counterparts given appropriate priors.

A plausible implication is the adaptability of these methods for other semantic super-resolution settings (e.g., semantic segmentation, multi-source fusion) where domain-appropriate constraints and priors can be synthetically instantiated, combined with scale-recurrent dense architectures for high-fidelity mapping. Extensions to full GAN-based perceptual objectives further expand the application spectrum, permitting controlled trade-offs between distortion metrics and perceptual realism (Purohit et al., 2022).

• "Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images Using Fully Synthetic Training" (Xu et al., 23 Jan 2026)
• "Image Superresolution using Scale-Recurrent Dense Network" (Purohit et al., 2022)