learnable spectral preprocessing

This study addresses the issue of erroneous invariance attribution in preprocessing for spectral foundation models by proposing a decoupled evaluation paradigm that distinguishes the contributions of normalization from representation learning. Through Raman spectroscopy analysis, controlled experiments, and numerical validation, we demonstrate that the invariance observed in multimodal models primarily stems from parameter-free preprocessing normalization rather than learned representations. These findings reveal that current models fail to surpass simple normalization baselines, effectively correcting prevailing cognitive biases regarding spectral representation learning capabilities within the field. Consequently, this work establishes a new benchmark for evaluation methodologies, emphasizing the necessity of rigorously isolating preprocessing effects when assessing the true efficacy of spectral foundation models.

Hyperspectral unmixing is severely hindered by scale-induced spectral variability—arising from terrain topography, illumination conditions, and shadow effects—leading to degraded estimation accuracy and poor convergence. This work presents the first systematic mathematical formulation of such variability and proposes a novel pre-processing framework grounded in geometric modeling and radiative transfer analysis. By formulating and solving an optimization problem, the method isolates and compensates for large-scale multiplicative distortions, thereby achieving spectral response scale correction. The approach is generic and seamlessly integrates into diverse unmixing pipelines. Extensive experiments on two synthetic and two real hyperspectral datasets demonstrate that the proposed pre-processing reduces abundance estimation errors of mainstream unmixing algorithms by approximately 50% on average, significantly enhancing unmixing accuracy, robustness, and convergence stability.

This study addresses the instability, poor auditability, and sensitivity to small sample sizes inherent in traditional near-infrared spectroscopic modeling, which relies on external preprocessing searches. To overcome these limitations, the authors propose an Operator-Adaptive Calibration framework (AOM) that embeds common preprocessing techniques—such as Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and Asymmetric Least Squares (ASLS)—as learnable linear operators integrated within local ensemble branches, thereby preventing information leakage. This approach unifies preprocessing and model training while preserving the interpretability of wavelength coefficients. Built upon PLS and Ridge regression, AOM leverages efficient algorithms and dual kernel formulations to enable rapid training and exact coefficient recovery. Evaluated across more than 50 heterogeneous datasets, AOM-PLS achieves a median RMSEP reduction of 4% over conventional PLS (outperforming it in 42 cases), while AOM-Ridge yields an average improvement of 2.22% (winning in 35 cases), with training times of only a few seconds.

This work addresses the challenge of hyperspectral image restoration, which is hindered by data scarcity, sensor-specific characteristics, and the high dimensionality of spectral information, making it difficult to learn robust priors. The authors propose a lightweight transfer framework that projects hyperspectral data into a low-dimensional subspace, leverages a frozen pre-trained RGB denoiser for noise removal, and reconstructs the hyperspectral cube through a lightweight adapter coupled with constrained linear aggregation. This approach is the first to efficiently transfer large-scale RGB image priors to hyperspectral restoration tasks, achieving plug-and-play performance with minimal training. It consistently outperforms specialized hyperspectral methods across multiple datasets in denoising, deblurring, and super-resolution, demonstrating the remarkable transferability of RGB-based priors.

Existing variational models for hyperspectral image denoising lack noise adaptivity due to fixed weights in the data-fidelity term, rendering them inadequate for complex, unknown mixed noise (e.g., impulse, stripe, and coupled noise). To address this, we propose a learnable pixel-wise weighted data-fidelity term and design a Hyper-Weight Network (HWnet) that dynamically predicts spatially varying noise intensity maps. Within a bi-level optimization framework, weight prediction and denoising are decoupled. This work is the first to formulate noise intensity estimation as a hypernetwork learning problem, introduces a model-level noise knowledge transfer mechanism, and provides preliminary theoretical analysis of generalizability. Experiments demonstrate consistent PSNR improvements of 2.1–4.7 dB over mainstream model-driven frameworks (e.g., LRMR, LRTV). Moreover, HWnet exhibits strong cross-model and cross-noise-type generalization capability.

This study addresses the scarcity of annotated data and the underutilization of visual priors in hyperspectral foundation models by proposing a promptable hyperspectral foundation model. Methodologically, a physics-informed full-spectrum synthesis pipeline is constructed to generate high-quality training data. Architecturally built upon SAM3 and ViT, the model introduces a mixture-of-experts fine-tuning mechanism that integrates a frozen RGB branch with zero-initialized injection. Efficient spectral adaptation is further achieved through ControlNet-style feature injection and cross-modal sample selection. Experimental results demonstrate that the proposed model exhibits strong generalization across diverse tasks, including classification and anomaly detection, validating that high-quality synthetic data outperforms noisy supervision scaling strategies.

This work addresses the limitations of existing hyperspectral band selection methods, which are sensitive to initialization, require a preset number of bands lacking flexibility, and exhibit unstable performance under spatially disjoint evaluation protocols. To overcome these issues, the authors propose SGBR-HC, a two-stage approach that first performs supervised band ranking based on class separability and spectral diversity to provide an informative prior for learnable sparse gating. In the second stage, the sparse gating module and a spatial classifier are jointly trained to adaptively determine the optimal number of bands. By integrating differentiable sparse gating, Hard-Concrete initialization, and spatially disjoint evaluation, the method effectively prevents information leakage. Experiments on the Pavia University and Houston 2013 datasets demonstrate state-of-the-art overall accuracy and Cohen’s kappa using only approximately 20 selected bands, while ablation studies confirm the critical role of the ranking prior.

This study addresses the challenges of cross-sensor generalization and arbitrary-scale reconstruction in hyperspectral super-resolution, where existing methods require additional training for novel scenarios. We propose OmniHSR, an all-band shared operator prediction paradigm that achieves spatial-spectral decoupling by resampling inputs via cross-spectral mapping, predicting Gaussian-supported local operators, and applying them across all bands through continuous operator field reconstruction. This formulation enables zero-shot transfer to unseen sensors and arbitrary scales with minimal parameters. Extensive experiments demonstrate that OmniHSR outperforms baselines on seven datasets and achieves state-of-the-art zero-shot performance on six unseen datasets. Furthermore, it yields a 0.55 dB PSNR improvement across twelve upscaling factors while accelerating inference by 36×.

This work addresses the challenges in hyperspectral image denoising, particularly the difficulty of dynamically balancing data fidelity and noise priors, as well as inadequate modeling of mixed noise. To overcome these issues, the authors propose a spatial-spectral adaptive denoising framework that employs an adaptive weight tensor to dynamically adjust the trade-off between the fidelity term and regularization. The method integrates a lightweight noise prior, a pixel-wise robust model, and coefficient total variation regularization to effectively preserve both the low-rank structure and local smoothness of the image. An optimization strategy based on the Alternating Direction Method of Multipliers (ADMM) ensures a favorable balance between computational efficiency and reconstruction accuracy. Extensive experiments on both synthetic and real-world datasets demonstrate state-of-the-art denoising performance.

Remote sensing image fusion models suffer from limited generalizability due to scarce authentic ground-truth data and domain shifts across heterogeneous sensors. To address this, we introduce, for the first time, a foundation model paradigm into remote sensing fusion, proposing a large-scale pretraining framework grounded in spatial-spectral priors. Our method synthesizes a highly diverse dataset by applying realistic degradations—including blur, noise, and downsampling—to ImageNet and SkyScript images. This enables effective pretraining across multiple architectures, including CNNs, Transformers, and Mamba. The resulting model achieves zero-shot and few-shot pan-sharpening, outperforming state-of-the-art methods on six major satellite datasets (e.g., WorldView). Remarkably, it adapts to unseen sensor domains with fine-tuning on merely a single real-world image. Our work establishes a new benchmark for cross-domain generalization in remote sensing image fusion.