Despite significant advancements, segmentation based on deep neural networks in medical and surgical imaging faces several challenges, two of which we aim to address in this work. First, acquiring complete pixel-level segmentation labels for medical images is time-consuming and requires domain expertise. Second, typical segmentation pipelines cannot detect out-of-distribution (OOD) pixels, leaving them prone to spurious outputs during deployment. In this work, we propose a novel segmentation approach which broadly falls within the positive-unlabelled (PU) learning paradigm and exploits tools from OOD detection techniques. Our framework learns only from sparsely annotated pixels from multiple positive-only classes and does not use any annotation for the background class. These multi-class positive annotations naturally fall within the in-distribution (ID) set. Unlabelled pixels may contain positive classes but also negative ones, including what is typically referred to as background in standard segmentation formulations. To the best of our knowledge, this work is the first to formulate multi-class segmentation with sparse positive-only annotations as a pixel-wise PU learning problem and to address it using OOD detection techniques. Here, we forgo the need for background annotation and consider these together with any other unseen classes as part of the OOD set. Our framework can integrate, at a pixel-level, any OOD detection approaches designed for classification tasks. To address the lack of existing OOD datasets and established evaluation metric for medical image segmentation, we propose a cross-validation strategy that treats held-out labelled classes as OOD. Extensive experiments on both multi-class hyperspectral and RGB surgical imaging datasets demonstrate the robustness and generalisation capability of our proposed framework.

Hyperspectral remote sensing image classification based on domain-level complementarity of spatial-spectral component.

Investigation of spatial distributions of components within a pyrite concretion through Raman imaging coupled with classical least squares method.

Use of Minimum Noise Fraction (MNF) denoising, previously developed for remote sensing applications, is an increasingly popular denoising technique for Infrared (IR) imaging data. The original MNF method proposed by Green et al. along with the faster 'Fast MNF' and resolution independent 'MNF2' all use a noise correlation matrix calculated based on neighbouring pixels, creating a heavy order-dependence. This approach fails when the spatial relationship between pixels is disrupted, for example, when large images cannot be loaded into memory on a standard workstation and are thus processed in patches or tissue data extracted using masking. We propose a spatially invariant MNF denoising method (iMNF) that uses a non-uniform, physically motivated noise estimation profile that removes this order-dependence, resulting in a robust, spatially invariant MNF based denoising algorithm. This allows for the application of the MNF denoising application to datasets where the spatial assumption is likely to be weakened by use of masking, or for unordered data such as randomly drawn labelled data, patch-wise segmentations of large scale images, or single-point spectral collections. This application was tested on representative prostate tissue biopsies for their spatial and chemical heterogeneity. Results indicate a robust, spatially invariant denoiser that is comparable to the Fast MNF method for structured and loosely structured data but is superior for unstructured data. This removes a critical bottleneck in the analysis pipeline for large IR images, such as those required in spectral pathology.

White blood cell (WBC) classification plays a crucial role in clinical diagnosis. However, traditional microscopic examination and existing deep learning methods primarily focus on spatial structures, resulting in the insufficient utilization of spectral information. In this study, we apply hyperspectral imaging technology and propose a 3D Swin Transformer network (SwinMDC) based on multi-scale dilated convolutions for WBC hyperspectral image classification. The network employs a 3D multi-scale dilated convolution feature extractor as its initial embedding layer, thereby enhancing low-level representations and alleviating receptive field limitations. By integrating a 3D window-based attention mechanism, the network constructs a hierarchical Transformer encoder to capture long-range dependencies. On a three-class white blood cell microscopic hyperspectral dataset, SwinMDC achieved an overall classification accuracy of 99.07%, demonstrating its potential value for clinical white blood cell analysis.

CSA-Kansformer : Cross-scale aggregation and Kansformer network for hyperspectral image classification.

ABSTRACT Hyperspectral image (HSI) classification remains a challenging task due to the high spectral dimensionality of the data, strong spectral redundancy, and limited availability of labelled samples. Although convolutional neural networks (CNNs) and Vision Transformers (ViTs) have demonstrated strong performance by exploiting spectral-spatial information and long-range dependencies, they often suffer from high computational complexity and large parameter counts, which limit their practical applicability. To address these limitations, a unified hybrid framework, termed ConvVitMamba, is proposed for efficient hyperspectral image classification. The proposed architecture integrates three complementary components within a single model: a multiscale convolutional feature extractor for capturing local spectral, spatial, and spectral-spatial patterns; a Vision Transformer-based tokenization and encoding stage for modelling global contextual relationships; and a lightweight Mamba-inspired gated sequence mixing module for efficient content-aware sequence refinement without relying on quadratic-complexity self-attention. Principal Component Analysis (PCA) is employed as a preprocessing step to reduce spectral redundancy and improve computational efficiency. Extensive experiments are conducted on four benchmark hyperspectral datasets, including Houston and three UAV-borne QUH datasets (Pingan, Qingyun, and Tangdaowan). Quantitative results, evaluated using Overall Accuracy, Average Accuracy, and the Kappa coefficient, demonstrate that ConvVitMamba consistently outperforms state-of-the-art CNN-, Transformer-, and Mamba-based methods while maintaining a favourable balance between classification accuracy, model size, and inference efficiency. Ablation studies further confirm the complementary contributions of the multiscale convolutional, transformer, and Mamba-inspired components. These results indicate that the proposed framework provides an effective and efficient solution for hyperspectral image classification under both urban and natural scene settings. The source code is publicly available at https://github.com/mqalkhatib/ConvVitMamba

Due to the inherent limitations of both hyperspectral and multispectral imagery, balancing high spatial resolution with high spectral fidelity has become one of the fundamental challenges in remote sensing image processing. A prevailing strategy is to fuse these two types of data to reconstruct images that jointly preserve their respective advantages. However, existing reconstruction approaches still suffer from complex coupling between spatial and spectral information, and limited feature extraction capabilities. To address these issues, this study proposes PMSwinNet (Pyramid Multi-scale Swin Transformer Network), a novel architecture that integrates pyramid-based feature enhancement with Transformer mechanisms. The PMSwinNet incorporates multi-scale pyramid feature fusion and window-based self-attention. Through a progressive multi-stage design and three complementary components—feature extraction and reconstruction modules—the Transformer branch leverages window partitioning and shifting operations to capture long-range spatial dependencies and local contextual cues, while the pyramid features extract both global and local information across multiple spatial scales. In addition, a high-frequency branch is introduced, which employs lightweight convolutions to enhance edges, textures, and other high-frequency details, effectively suppressing blurring and artifacts during reconstruction. Experimental evaluations on multiple public hyperspectral datasets demonstrate that the PMSwinNet outperforms state-of-the-art methods, particularly in terms of detail preservation, spectral distortion suppression, and robustness.

Endmember extraction (EE) plays an essential role in the unmixing of hyperspectral images, and many EE algorithms have been proposed. Among them, evolutionary algorithm (EA) based EE algorithms attract much attention due to the EA's powerful global search ability. However, hyperspectral EE problem itself is a constrained and sparse large-scale optimization problem, and it is difficult to search the optimal solutions efficiently. In this paper, to address this problem, we tackle this complex optimization problem from the view of evolutionary multi-tasking. Specifically, an evolutionary multi-tasking method for hyperspectral EE, termed EMT-HEE, is suggested, where two related tasks cooperate with each other to achieve endmembers with higher quality. In EMT-HEE, the original hyperspectral EE problem is regarded as the main task, whose aim is to obtain the final accurate endmembers. Meanwhile, an unconstrained task is constructed to assist the main task, and is used to explore the sparse large-scale search space thoroughly and avoid the local optima. To implement the evolutionary multi-tasking idea, two populations (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$MP$</tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$AP$</tex-math></inline-formula>) are evolved for the two tasks, respectively. During the evolving, a learning based solution generation strategy is suggested for the main task, which can produce high-quality solutions. Besides, a global search strategy with dynamic ranges is developed for the assisting task, which generates the solutions with more diversity. To make full use of the advantage of each task, a pair of knowledge transferring strategy (including the “assisting-to-main repairing strategy” and the “main-to-assisting enhancing strategy”) is proposed, which improves the performance of each task greatly. The competitiveness of EMT-HEE is validated on different hyperspectral datasets, and EMT-HEE can extract more accurate endmembers than the state-of-the-art EE algorithms.

The Arctic is warming more than four times as fast as any other region of our planet. This warming has led a dramatic decline in seasonal sea-ice coverage and thickness and to a poleward extent of the Atlantic water. The Fram Strait is the only deep gateway of the Arctic Ocean where the West Spitsbergen Current transports the Atlantic water northward and the East Greenland Current the polar water southward. This is a highly dynamic oceanographic regime in the zone (central Fram Strait) between the two currents with its mixed water in combination with the semi-permanent sea-ice edge and large horizontal density gradients observed in the marginal ice zone. Long-term high-spatial resolution data on phytoplankton distribution and its community composition help to assess the impact of these physical changes on the biological processes in the surface waters of Fram Strait. In this study, we resolve the implications of changes in sea-ice conditions and physical hydrographic properties in Fram Strait to the chlorophyll-a concentration (Chl-a) of the whole phytoplankton community and of the major phytoplankton groups (PGs) contributing to its biomass. We extend the formerly collected total Chl-a (TChl-a) and PG Chl-a data sets of High-Performance Liquid Chromatography and of hyperspectral particulate absorption data sets from high-frequency underway spectrophotometry measurements to include the five recent East Greenland Sea expeditions from 2019 and 2021 to 2024. We adapt well-established methods to retrieve from the hyperspectral data as baseline for more high-spatial resolution TChl-a and PG Chl-a retrievals. Cross validation and independent validation confirm the robustness of our 2015 to 2024 data sets. We analyse the ship-based high-spatial resolution (around 300 m) continuous Fram Strait time series data along the ship transect of the eight expeditions. We observe diatoms and haptophytes, the key functional groups of the marine ecosystem, together with chlorophytes to represent most of the phytoplankton community. We identify a shift in spring blooms’ phytoplankton community composition responding to the distinct sea-ice thickness conditions in different years. We observe in the West Spitsbergen Current that the mid-summer TChl-a and the contribution of haptophytes to the phytoplankton community are much lower during the latest years (2022 and 2024) identified as cold as opposed to the warm (2015 to 2017) temperature anomaly years. We recommend continuing the underway spectrophotometry measurements in future and combine these long-term high-spatial resolution time series on phytoplankton community composition with satellite measurements to track the effects of the changes of the physical environment to global warming on the main primary producer in the Arctic Ocean. • High-spatial resolution Chl-a data, including phytoplankton groups’ contribution. • Spring bloom phytoplankton composition responds to sea-ice thickness conditions. • Chl-a is lower in mid-summer during cold opposed to warm temperature anomaly years. • Haptophytes contribute less during cold as opposed to warm anomaly years.

Coastal water color, as captured by the hue angle from satellite imagery, provides a comprehensive indicator of ecosystem dynamics influenced by environmental and human factors. This study optimizes a bias-correction model for deriving hue angle from MODIS data using a large in situ hyperspectral dataset, achieving improved accuracy ( R ² = 0.71, MAPE = 5.67%, RMSE = 12.05°). Applied to Qinhuangdao coastal waters over 2003–2024, the model uncovers a nearshore-to-offshore gradient, a V-shaped seasonal pattern, and a triphasic interannual trend: degradation (2003–2011), recovery (2012–2023), and a 2024 reversal. Phytoplankton chlorophyll-a (Chl-a) emerges as the dominant driver of color variability ( R ² = 0.67), with dissolved inorganic nitrogen (DIN) as the key environmental influence ( R ² = 0.54), underscoring nutrient-driven eutrophication. Policy measures, including 2008 Olympic pollution controls and the 2018 Bohai Sea remediation, correlate with post-2012 improvements. The hue angle demonstrates greater sensitivity than the Forel-Ule Index, offering a valuable tool for coastal management.

This study presents an annotated multi-sensor, multimodal, and hyperspectral dataset designed to support deep learning-based classification and segmentation of bulky waste. The dataset comprises four distinct sensor modalities: high-resolution visible RGB images (VIS), hyperspectral near-infrared (NIR), temporally resolved thermal infrared (IR), and terahertz (THz) imaging with depth information, providing complementary multimodal information. An image registration process aligns all modalities to a common reference frame, enabling near pixel-precise fusion across sensors. WoodVIT contains 56 registered multi-sensor scenes, partitioned into 22,659 annotated patches with two main classes (wood and non-wood) and 16 subclass labels. It includes pixel-masks and patch-wise annotations to facilitate both segmentation and classification tasks. The primary benchmark task is binary discrimination of wood versus non-wood. The dataset also includes challenging scenarios involving occlusion and concealed contaminants (e.g., embedded metals) to motivate robust multimodal fusion approaches. We provide predefined train/validation/test splits and report baseline results using convolutional neural networks and fusion architectures to establish reference performance. WoodVIT is publicly available to support research on multi-sensor learning for waste sorting.

Accurate crop classification provides fundamental data for agricultural resource management and ecological research. Hyperspectral image (HSI) classification is the core technique for achieving precise crop mapping. However, existing models often suffer from excessive parameters, limited robustness under few-shot conditions, and a trade-off between efficiency and robustness. To address these issues, this paper proposes a Multi-Scale Wavelet Kolmogorov–Arnold Network (MSWKN). The model employs a Two-Branch Feature Extractor (TBFE) to capture both spectral correlations and spatial textures. a Channel Cross-Spatial (CCS) module to suppress background clutter and highlight discriminative regions. A group convolution-based Fixed Wavelet Multi-Scale Convolutional Layer (FW-MSCL) that leverages the time–frequency localization of wavelets and learnable linear combinations to enhance robustness against spectral distortion while reducing parameters. And a Fourier-based Transformer encoder to enable global frequency–space modeling. Experiments on the WHU-Hi-HanChuan and WHU-Hi-HongHu hyperspectral crop datasets show that MSWKN achieves high overall accuracy and performs favorably on few-shot categories. Under lower parameter counts and fast inference conditions, the model demonstrates a reasonable trade-off between accuracy and computational efficiency. Ablation studies and wavelet kernel comparisons further confirm the contribution of each module and the advantage of the wavelet. The proposed framework provides an efficient and robust solution for fine-grained hyperspectral crop classification.

The classification of hyperspectral images (HSI) is fundamental to applications such as resource exploration, environmental monitoring, and precision agriculture. However, the scarcity of labeled samples caused by the high cost of manual annotation severely limits the performance of supervised HSI classification models. To address this issue, self-supervised learning (SSL), particularly contrastive learning (CL), has attracted increasing attention due to its ability to leverage large amounts of unlabeled data. Nevertheless, existing CL methods typically rely on data augmentation to construct positive pairs, which is often ineffective or inappropriate for hyperspectral pixel-level samples. To overcome these limitations, we propose a novel Multi-view Contrastive Learning (MVCL) framework for HSI classification, which performs contrastive learning by explicitly exploiting multiple complementary views of the same hyperspectral sample, rather than relying on handcrafted data augmentations. By maximizing mutual information across different views during pre-training, MVCL learns more discriminative and robust spectral–spatial representations. Furthermore, we design a Spatial Aggregation-based Attention (SPAA) module to enhance the transformer backbone within the contrastive learning framework. The SPAA module integrates convolutional spatial aggregation into the self-attention mechanism, enabling more effective local feature modeling while significantly reducing computational complexity. In addition, a supervised classification loss is jointly incorporated to further enforce intra-class compactness and inter-class separability during fine-tuning. Extensive experiments on three widely used hyperspectral benchmark datasets demonstrate that the proposed MVCL framework consistently maintains competitive performance. Specifically, the proposed approach achieves Overall Accuracies of 93.58%, 94.87% and 99.15% compared to the ViT baseline, the improvements are 15.63%, 3.38% and 18.08% respectively, confirming the effectiveness and generalization capability of the proposed multi-view contrastive learning strategy. The source code is publicly available at: https://github.com/yangzhengrui732/MVCL.git .

The application of hyperspectral image (HSI) processing techniques has become increasingly important in many fields such as agriculture, environmental detection, and mining. However, the number of annotated samples in existing hyperspectral datasets is limited, and most hyperspectral classification models typically use patch data for model training. There, the pixels to be classified are often taken as the center, and then a mean pixel length is calculated from the surrounding pixel neighborhood to form patch data. However, during model training, the researchers found that there was an area of overlapping pixels between most of the training data and the test data. This inevitably led to data leakage, resulting in excellent classification performance of the model. To solve this problem, we develop a method of replacing overlapping pixels (ROP) in patch data, which means the training pixel points of the same class are used to replace the test pixel points that appear in the overlapping region of the training patch data. Furthermore, a multiple feature extraction and fusion (MFEF) module is also proposed to enhance the capacity of the HSI model to extract spectral–spatial feature information from new patch data. The results on five publicly available HSI datasets demonstrate that the proposed resolving data leakage network (RDLNet) can provide competitive classification results on the patch data reconstructed with the ROP strategy, which outperforms existing state-of-the-art (SOTA) classification methods as well.

An accurate assessment of how spectral and spatial resolution influence coastal mapping remains a critical challenge for shallow-water monitoring. This study evaluates and compares hyperspectral (97 bands), multispectral (8 bands), and RGB (3 bands) data, with different spatial resolutions (10cm and 2m), to determine the most suitable spectral-spatial configuration for shallow-water mapping. For this, the effects of spectral and spatial resolution are isolated by simulating 8-band multispectral and 3-band RGB configurations at 10cm and 2m from a single 97-band hyperspectral drone dataset. This allows for comparisons between different resolutions without considering changes in temporal, atmospheric, water column, or image capture, among others acquisition-related factors. A comprehensive methodology for processing was developed using empirical and machine learning models for bathymetry estimation (Sigmoid, Subspace-KNN) and benthic mapping (SVM, FNN). The developed framework was applied at an urban sandy beach sheltered by a natural reef with rich marine biodiversity (Las Canteras beach, Gran Canaria, Spain). Results show that hyperspectral data achieved the highest accuracy (MAE, 0.15m; accuracy, 94%), while multispectral data offered an excellent balance between resolution and performance (MAE, 0.16m; accuracy, 93%). RGB data was acceptable for bathymetry but unreliable for benthic classification in complex habitats (MAE, 0.24m; accuracy, 83%). Subspace-KNN outperformed empirical models for bathymetry, and FNN improved substrate discrimination. In addition, a comparative analysis between 2016 and 2023 imagery, comparing real WorldView-2 imagery (2016; 2m and 8 bands), and drone imagery with the same resolutions emulated (2023; 2m and 8 bands), suggests an approximate 7,200m² reduction in marine vegetation that may be influenced with anthropogenic pressures and thermal increase. This approach provides a reproducible and adaptable tool for sustainable coastal management.

ABSTRACT Graph convolutional networks (GCN) excel in rapidly mining semantic information from data with superior capabilities, enabling deep feature extraction and classification for hyperspectral images (HSI). However, supervised classification methods heavily rely on many high-quality labeled samples, which is often a significant limitation in real-world applications. Unsupervised learning, with its unique ability to automatically discover potential structures in data without labeled samples, provides a new approach to address this issue. Thus, this paper proposes an unsupervised dual-masked graph autoencoder feature extraction network (UDMG) for HSI classification. This method first utilizes a masked autoencoder to mask and reconstruct the edges and node features of the graph structure. Meanwhile, the encoder combines GCN and Transformer to capture local and global features in HSI. To validate the performance of UDMG, we employ various evaluation methods and conduct experiments on three publicly available hyperspectral datasets. The results demonstrate that UDMG exhibits superior classification performance compared to existing state-of-the-art unsupervised feature extraction methods.

Abstract Developed by the King Abdullah University of Science and Technology (KAUST) to support research in atmospheric science and remote sensing, KAUSTSat represented Saudi Arabia's (and the Middle East's) first research‐focused hyperspectral CubeSat mission for Earth observation. The primary payload consisted of the Simera HyperScape50, a miniaturized hyperspectral sensor operating in the visible to near‐infrared range. The sensor was equipped with a custom continuous variable filter that collected imagery at 30 m spatial resolution with a 120 km swath. The HyperScape50 allowed for up to 32 user‐defined spectral bands to be selected per acquisition from a total of 442 programmable channels between 442 and 884 nm, including a panchromatic band. These capabilities enable detailed observations of vegetation, soil, coastal zones, and other surface features relevant to applications in agriculture, biodiversity, resource management, and disaster response. In this paper, we provide an overview of the mission architecture, sensor design, acquisition strategy, and data structure. The hyperspectral data sets acquired over the 14‐month mission lifetime will also be presented, with a particular focus on the Arabian Peninsula and RadCalNet calibration sites. The KAUSTSat mission serves as a demonstration case of the viability of academic‐driven CubeSat platforms for delivering targeted, high‐quality environmental data, and represents a valuable reference for future small satellite Earth observation programs.

Transformer-based and band-selected models for UAV hyperspectral wheat disease classification

This paper introduces Generalized Nonnegative Structured Kruskal Tensor Regression (NS-KTR), a novel tensor regression framework that enhances interpretability and performance through mode-specific hybrid regularization and nonnegativity constraints. Our approach accommodates both linear and logistic regression formulations for diverse response variables while addressing the structural heterogeneity inherent in multidimensional tensor data. We integrate fused LASSO, total variation, and ridge regularizers — each tailored to specific tensor modes — and develop an efficient alternating direction method of multipliers (ADMM)-based algorithm for parameter estimation. Comprehensive experiments on synthetic signals and real hyperspectral datasets demonstrate that NS-KTR consistently outperforms conventional tensor regression methods. The framework’s ability to preserve distinct structural characteristics across tensor dimensions while ensuring physical interpretability makes it especially suitable for applications in signal processing and hyperspectral image analysis. • NS-KTR: nonnegative structured Kruskal tensor regression with hybrid regularization. • Mode-specific regularization: LASSO, total variation, and ridge across tensor modes. • Unified framework supports linear and logistic regression for diverse responses. • ADMM-based optimization achieves superior accuracy with significant speedups.