Due to the inherent spatial–spectral trade-off of real imaging systems, remote sensing platforms typically capture only low-resolution hyperspectral images (LR–HSI) or high-resolution panchromatic images (HR–PAN), which introduces mixed pixels and weakens spectral separability, ultimately impairing downstream perception. While existing HSI–PAN fusion methods attempt to recover high-resolution HSI (HR–HSI), they are mainly optimized for pixel-level fidelity without alignment to task-level supervision. Conversely, classification models are often trained independently on idealized datasets, failing to exploit physical priors embedded in the sensing and fusion process. This leads to a persistent gap between reconstruction quality and practical classification effectiveness. To mitigate this gap, we propose a physics-constrained fusion and prior-guided classification framework that connects imaging and perception through structured prior transfer. A Physics-Constrained Unfolding Transformer (PUT) is developed, which unfolds the sensor degradation model through a high-order Neumann series expansion, enabling physically interpretable and spectrally consistent HR–HSI reconstruction. Within PUT, a Spatial–Spectral Synergy Block (SSSB) serves as a learnable prior to enhance spatial–spectral representation during unfolding. After being optimized with reconstruction-oriented objectives, the fusion module produces physics-driven spectral weight maps and spatial–spectral representations, which are subsequently transferred as structured priors into a Physics-Aware Dual-Prior Classification Network (PDN). The classification network further incorporates the original LR–HSI and HR–PAN as complementary data priors, allowing it to jointly leverage physics-guided reconstruction cues and authentic sensor observations for robust discrimination. Extensive experiments on four simulated datasets and two real remote sensing datasets demonstrate its superior dual-task performance and strong potential for end-to-end remote sensing perception systems.

Remote sensing multimodal large language models (MLLMs), which integrate rich visual-linguistic modal in- formation, have shown great potential in areas such as remote sensing image analysis and interpretation. However, exist- ing knowledge distillation methods primarily focus on the compression of unimodal large language models, neglecting the alignment of features across modalities, thus hindering the performance of large language models in cross-modal tasks. To address this issue, a lightweighting method for remote sensing MLLMs based on knowledge distillation is pro- posed. This method achieves effective alignment of multimodal information by aligning the outputs across modalities at the feature level. By introducing the reverse Kullback-Leibler divergence as the loss function and combining optimiza- tion strategies such as teacher mixed sampling and single-step decomposition, the generalization and stability of the stu- dent model are further enhanced. Experimental results demonstrate that the proposed method achieves higher accuracy and efficiency in four downstream tasks of remote sensing image scene classification, visual question answering, visual localization, and image description, significantly reducing the number of model parameters and the demand for computa- tional resources, thereby providing a new solution for the efficient application of MLLMs in the field of remote sensing.

Large language models (LLMs) have demonstrated exceptional reasoning capabilities, and co-evolving paradigms have shown promising results in domains such as code and math. However, in scientific reasoning tasks, these models remain fragile due to unreliable solution evaluation and limited diversity in verification strategies. In this work, we propose Sci-CoE, a two-stage scientific co-evolving framework that enables models to self-evolve as both solver and verifier through a transition from sparse supervision to unsupervised learning. In the first stage, the model uses a small set of annotated data to establish fundamental correctness judgment anchors for the Verifier. In the second stage, we introduce a geometric reward mechanism that jointly considers consensus, reliability, and diversity, driving large-scale self-iteration on unlabeled data. Experiments on several general scientific benchmarks demonstrate that Sci-CoE enhances complex reasoning capabilities and exhibits strong scalability, facilitating the construction of more robust and diverse evaluation systems. Codes are available at https://github.com/InternScience/Sci-CoE.

Blind hyperspectral unmixing is an ill-posed inverse problem that poses significant challenges. Numerous model-based and deep learning methods have been proposed to address it, achieving remarkable progress. However, these methods are predominantly based on the linear spectral mixing model, with insufficient exploration of nonlinear blind unmixing; furthermore, their results are highly sensitive to endmember initialization, and the utilization of intrinsic properties of hyperspectral images (HSIs) remains limited. To address these issues, this paper proposes a dual-stream autoencoder based on endmember kernel (called DSEK-AE), which enables blind linear/nonlinear unmixing through a knowledge- and data-driven approach. The encoder consists of two streams: the endmember extraction stream aggregates information from all pixels to directly derive endmembers, while the abundance estimation stream comprises three serial modules—spectral perception module (SPM), global-local aggregation module (GLAM), and sparsity adjustment layer (SAL)—which effectively capture both global and local spatial-spectral information from HSIs, significantly improving unmixing accuracy. The decoder is designed based on the endmember-kernel theory, explicitly modeling interactions among end-members and enabling flexible adjustment of embedded spectral mixing models without altering the underlying network structure, thereby enhancing the network’s physical interpretability and flexibility. Moreover, we propose a novel loss function specifically designed for blind unmixing to facilitate more effective training. Experimental results on real-world and physically simulated hyperspectral datasets consistently demonstrate that the proposed method exhibits superior unmixing performance compared to current state-of-the-art methods. The code will be available at https://github.com/I3ab/DSEK-AE.

Hyperspectral images (HSIs) are valuable because of their rich spectral information and broad applications. However, the inherent limitations in the spectral, spatial, and temporal dimensions of HSIs prevent the rapid acquisition of information-rich images. In addition to approaches that fuse low-resolution HSIs with high-resolution multispectral images or employ spectral compressive imaging systems, direct reconstruction of HSIs from common multispectral images (MSIs) has gained attention as a solution, but end-to-end deep learning approaches often lack physical grounding, which diminishes both interpretability and cross-domain generalizability. Conversely, physically integrated models offer better interpretability but struggle with complex real-world scenarios. To overcome these challenges, we propose an adaptive spectral degradation guidance mechanism that combines deep learning with physical principles to enable the model to adapt to different degradation patterns and improve reconstruction accuracy, stability, and interpretability. Furthermore, we designed a cross-stage double attention mechanism to mitigate information loss in existing methods to enhance information transfer and feature interaction, thereby boosting model expressiveness and learning efficiency. To validate the performance of our approach in different scenarios (known and unknown spectral degradation), we conducted experiments on four simulated datasets and two real-world datasets, and our method outperformed SOTA algorithms across most metrics. The simulation experiments demonstrated the effectiveness of our method, the real-world experiments confirmed its feasibility, and the ablation studies verified the correctness of our architecture. Code available: https://github.com/XWangbin/DASDU-Net.

In real-world applications, data streams naturally evolves and thus may exhibit a dynamic feature space, wherein new features appear and old ones disappear. Online learning under such circumstances necessitates simultaneous learning from increasing data volume and adaptation to the dynamic feature space in real time. While several methodologies have been proposed to tackle this challenge, many of them rely on strong assumptions regarding evolving interaction manners. Instead, this study adopts a broader perspective aimed at facilitating learning from arbitrarily evolving features without any strict assumptions in the previous work. Then, we present an online learning method based on a non-parametric kernel model. This model accommodates data streams with both continuous instances and evolving features through simple deduction root and addition operations. Theoretical analysis shows the sublinear regret (9( T) of the proposed method. Empirical studies show the capability to adapt not only to the previously constrained evolving features but also to the more arbitrarily evolving features.

Event cameras are increasingly gaining attention as neuromorphic vision sensors characterized by low latency, high dynamic range, and low power consumption. These features provide significant potential for visual tasks in challenging scenarios involving high motion or low lighting conditions. However, despite the advantages of event cameras, the development of neural networks that effectively process the event data they produce remains immature, leading to performance in many visual tasks that still lag behind traditional frame-based methods. To further explore neural networks for event data, we propose the filter bi-level routing attention recurrent vision transformers (FB-RVTs), a backbone for object detection with event cameras. We present three main contributions: First, it introduces a bi-level routing attention mechanism named filter bi-level routing attention and integrates it into the spatial feature extraction module of the proposed backbone to achieve adaptive attention allocation. Second, it proposes a temporal feature extraction module, named attention-enhanced long short-term memory which includes an additional gating module to further enhance the temporal generalization capability of the module. Lastly, a convolutional downsampling module, named pixel-equidistant convolution, is presented to reduce bias and information loss in the downsampling process. The proposed network is trained and tested from the beginning on the Gen1 autonomous driving dataset, achieving a mean average precision (mAP) of 48.3. Compared with the widely influential recurrent vision transformers, which achieved an mAP of 47.2, the proposed FB-RVT improves the mAP to 48.3, marking a 1.1 percentage point enhancement. Similarly, when benchmarked against the group event transformer (GET) with an mAP of 47.9, FB-RVT raises the mAP to 48.3, demonstrating a 0.4 percentage point improvement. These findings underscore the superior performance of FB-RVT in comparison with existing state-of-the-art models in event-based object detection tasks. (c) 2025 SPIE and IS&T

Although remote sensing (RS) data with multiple modalities can be used to significantly improve the accuracy of semantic segmentation in RS data, how to effectively extract multimodal information through multimodal feature fusion remains a challenging task. Specifically, existing methods for multimodal feature fusion still face two major challenges: 1) Due to the diverse imaging mechanisms of multimodal RS data, the boundaries of the same foreground may vary across different modalities, leading to the inclusion of unwanted background semantics in the fused foreground features; 2) RS data from different modalities exhibit varying discriminative abilities for different foregrounds, making it challenging to determine the proportion of semantic information for each modality in the fusion results. To address the above issues, we propose a dynamic feature fusion method based on region-wise queries, namely DF 2 RQ, for SS of multimodal RS data. This method is primarily composed of two components: the spatial reconstruction (SR) module and the dynamic fusion (DF) module. Within the SR module, we propose a spatial reconstruction scheme that samples foreground features from different modalities, achieving independent reconstruction of different unimodal features, thereby alleviating the semantic mixing between foreground and background across modalities. In the DF module, a feature fusion scheme based on unimodal feature reference positions is proposed to obtain fusion weights for each modality, thereby enabling the dynamic fusion of complementary features from multiple modalities. The performance of the proposed method has been extensively evaluated on various multimodal RS datasets for SS, and the experimental results consistently show that the proposed method achieves state-of-the-art accuracy on multiple commonly used metrics. In addition, our code is available at https://github.com/I3ab/DF2RQ.

This paper presents a novel hyperspectral anomaly detection method based on a deep prior Mamba network (DPMN), which is primarily composed of two components: abundance generation module (AGM) and background reconstruction module (BRM). Specifically, the AGM employs convolutional techniques to extract local information and bidirectional Mamba to capture global information. The BRM then uses a background dictionary to divide the background into multiple subspaces. By combining the background dictionary with the generated abundance, we can realize accurate background reconstruction while effectively suppressing the interference of anomalies on background reconstruction. Moreover, to fully leverage the intrinsic properties of hyperspectral images (spatial smoothness and low-rankness), we incorporate a regularization term into the loss function, merging total variation (TV) into low-rank representation (LRR). This approach not only utilizes spatial smoothness and low-rankness but also reduces the number of hyperparameters. Experimental results on two real hyperspectral datasets demonstrate that the proposed method significantly outperforms other state-of-the-art (SOTA) methods in detection performance.

This paper presents a high-efficiency LLC resonant bus converter with a 10:1 step-down ratio, tailored for 48V data center applications. To address power density challenges, a compact design is achieved through a single-central-leg PCB planar transformer and a 3D stacked system integration scheme, reducing the converter footprint significantly compared to conventional designs. The implemented prototype demonstrates exceptional performance, delivering a maximum output current of 85 A with a peak efficiency of 98.8% at 48V input. Under 60V input conditions, the converter achieves a current density of 0.29 A/mm2 and a power density of 3 kW/inch3.

Addressing the challenges of large-scale, diverse, and densely distributed targets, as well as the limited resources of edge computing devices in unmanned aerial vehicle (UAV) aerial imagery, this paper proposes an enhanced STV2YOLO target detection framework. First, a dynamic multi-scale feature pyramid is constructed by integrating the cross-window attention mechanism from Swin Transformer V2, along with a hierarchical window partitioning strategy to improve the feature representation of small targets. Second, a lightweight backbone network is developed using depthwise separable convolution, the C3-Inception module to enable parallel and efficient extraction of multi-scale features. Finally, a dynamic weighted WIoU loss function is introduced to optimize the bounding box localization accuracy for small targets. Experimental results demonstrate that the proposed method achieves 60.5% in detection accuracy (mAP@0.5), which is 9.2 percentage points higher than the baseline YOLOv5 model and significantly outperforms mainstream models such as TPH-YOLOv5 and YOLOv5. Additionally, the framework achieves an inference speed of 32.5 FPS, balancing detection accuracy with real-time performance requirements. The code for this work has been open-sourced at https://github.com/Crazy-Create/STV2YOLO.git.

Dynamic magnetic resonance imaging (MRI) requires accurate reconstruction from undersampled k-space data to achieve high temporal resolution within clinically acceptable scan times. Deep unrolling architectures have recently emerged as effective solutions by integrating physics-based data consistency with learned priors. However, their ability to exploit temporal relationships remains limited, as many approaches rely on independent stage-wise processing with only final-stage outputs propagated across iterations, which restricts feature interaction and often leads to performance degradation when acceleration factors increase. To enhance temporal prior learning, we introduce a bidirectional recurrent convolutional unit within the sparse prior update module. Our approach strengthens temporal dependency modeling by recurrently aggregating contextual information from both past and future frames, thereby improving stability and representation capacity under highly undersampled conditions. Furthermore, we incorporate inter-stage feature transmission that forwards intermediate representations instead of only single-stage outputs. This design substantially improves multi-stage collaboration, enabling more effective refinement across iterations. Experimental results on accelerated dynamic MRI datasets (6×, 12×, and 24×) demonstrate that the proposed method consistently outperforms state-of-the-art unrolling and deep learning strategies in reconstruction accuracy and temporal fidelity. Ablation studies further validate the contributions of recurrent temporal learning and inter-stage feature transmission.

The data scaling law has been shown to significantly enhance the performance of large multi-modal models (LMMs) across various downstream tasks. However, in the domain of perceptual video quality assessment (VQA), the potential of scaling law remains unprecedented due to the scarcity of labeled resources and the insufficient scale of datasets. To address this, we propose OmniVQA, an efficient framework designed to efficiently build high-quality, human-in-the-loop VQA multi-modal instruction databases (MIDBs). We then scale up to create OmniVQA-Chat-400K, the largest MIDB in the VQA field concurrently. Our focus is on the technical and aesthetic quality dimensions, with abundant in-context instruction data to provide fine-grained VQA knowledge. Additionally, we have built the OmniVQA-MOS-20K dataset to enhance the model's quantitative quality rating capabilities. We then introduce a complementary training strategy that effectively leverages the knowledge from datasets for quality understanding and quality rating tasks. Furthermore, we propose the OmniVQA-FG (fine-grain)-Benchmark to evaluate the fine-grained performance of the models. Our results demonstrate that our models achieve state-of-the-art performance in both quality understanding and rating tasks.

Recently, the classification of multimodal remote sensing images (RSIs) has garnered significant attention due to its ability to provide rich information for various scenes on Earth. Compared to traditional feature fusion methods used for the classification of multimodal RSIs, neural architecture search (NAS) is capable of identifying the optimal network structure for multimodal RSIs and downstream tasks. However, due to the diverse spatial resolutions, complex channel dimensions, and drastic foreground scale variations of multimodal RSIs, challenges arise when employing NAS methods for precise classification: 1) Due to the complementary and redundant nature between different modalities in RSIs, determining the features within each modality for fusion becomes quite challenging; 2) the design of fusion operators does not take into account the spatial positions and channel relationships between different modalities of RSIs, making it difficult for the fused features to match downstream tasks. To address these issues, we propose a dual-stage feature fusion framework based on NAS, termed DSF2-NAS, for the classification of multimodal RSIs. It primarily consists of two components: the feature candidate search (FCS) module and the fusion operator search (FOS) module, which execute sequentially. In the FCS module, a feature distance-based regularization approach is proposed to ensure fusion using multimodal features with the highest complementarity. Meanwhile, in the FOS module, a series of fusion operators are designed, which are based on spatial positions, channel relationships, and self-attention mechanisms, aiming to better integrate multimodal features with complex spatial and channel information. The proposed method has been evaluated on various datasets of multimodal RSIs, and experimental results consistently show that this method achieves state-of-the-art performance across multiple classification metrics.

Unsupervised domain adaptation uses labeled data from a source domain to train a robust classifier for an unlabeled target domain with a distinct distribution. The Broad Learning System (BLS), known for its efficiency and effectiveness, is widely applied in classification and regression problems. This paper introduces a novel method named TD-BLS for unsupervised domain adaptation. TD-BLS combines UDA-BLSAE and discriminative BLS into an iterative network. UDA-BLSAE performs domain adaptation and data reconstruction simultaneously, balancing the preservation of intrinsic structure with the reduction of distribution discrepancy. Additionally, the discriminative BLS used in TD-BLS employs pseudo-labeling and manifold learning in the classifier stage to leverage high-confidence predictions and data geometric information. Finally, experiments on multiple public domain adaptation datasets demonstrate that our approach achieves rapid domain adaptation with higher accuracy compared to existing methods.

Recent advancements in hyperspectral anomaly detection (HAD) utilizing deep learning have garnered significant attention due to their superior performance. However, most existing methods based on convolutional neural networks (CNNs) and Transformer focus on extracting local and global features separately and assume that the background resides in a single subspace for reconstruction, thereby reducing the quality of the reconstructed background and decreasing the accuracy of HAD. Moreover, although incorporating the prior physical knowledge into the loss function can enhance the performance of the deep learning networks, it also increases the number of hyperparameters and complicates the tuning process. To address these issues, we propose a deep prior Mamba network (DPMN) for HAD, which primarily consists of two components: the abundance generation module (AGM) and the background reconstruction module (BRM). Specifically, AGM employs convolutional layers to extract local information and introduces Mamba to capture long-range dependencies, achieving feature extraction from local to global. Subsequently, BRM utilizes a learnable background dictionary to divide the background into multiple subspaces for reconstruction, realizing accurate background reconstruction while effectively suppressing the interference of anomalies on background reconstruction. Furthermore, to fully leverage the intrinsic properties of hyperspectral images (HSIs), we incorporate a regularization term into the loss function, merging the total variation (TV) with the low-rank representation (LRR), which not only exploits spatial smoothness and low ranknes but also reduces the number of hyperparameters. Experimental results on eight publicly available real datasets demonstrate that our method significantly outperforms other state-of-the-art (SOTA) methods. In addition, our code is available at: https://github.com/I3ab/DPMN

Remote sensing change detection (RSCD), a complex multi-image inference task, traditionally uses pixel-based operators or encoder-decoder networks that inadequately capture high-level semantics and are vulnerable to non-semantic perturbations. Although recent multimodal and vision-language model (VLM)-based approaches enhance semantic understanding of change regions by incorporating textual descriptions, they still suffer from challenges such as inaccurate spatial localization, imprecise pixel-level boundary delineation, and limited interpretability. To address these issues, we propose ViLaCD-R1, a two-stage framework comprising a Multi-Image Reasoner (MIR) and a Mask-Guided Decoder (MGD). Specifically, the VLM is trained through supervised fine-tuning (SFT) and reinforcement learning (RL) on block-level dual-temporal inference tasks, taking dual-temporal image patches as input and outputting a coarse change mask. Then, the decoder integrates dual-temporal image features with this coarse mask to predict a precise binary change map. Comprehensive evaluations on multiple RSCD benchmarks demonstrate that ViLaCD-R1 substantially improves true semantic change recognition and localization, robustly suppresses non-semantic variations, and achieves state-of-the-art accuracy in complex real-world scenarios.

This paper proposes a kernel-based autoencoder network to solve the problem of nonlinear spectral unmixing in hyperspectral images (HSIs), facilitating an unsupervised nonlinear unmixing process driven by both knowledge and data. Specifically, the encoder is composed of two serially connected components: the spectral perception module (SPM) constructed with a convolutional neural network (CNN), and the global-local aggregation module (GLAM) simultaneously incorporating CNN and differential transformer, so as to effectively extract both local and global spatial-spectral information from the input HSIs. The decoder is designed based on endmember-kernel theory, explicitly characterizing interactions among endmembers and allowing flexible adjustments to the embedded nonlinear spectral mixing model (NSMM) without altering the basic network architecture. Unlike existing methods, the proposed method not only integrates the strengths of model-based approaches and data-driven methodologies but also allows for the flexible incorporation of various NSMMs to address nonlinear blind unmixing problems in different scenarios. Experimental results from both the physical-based simulated and real hyperspectral datasets demonstrate that the proposed method exhibits significant advantages over the state-of-the-art (SOTA) methods.