ObjectiveWith the rapid development of intelligent transportation systems， vehicle multi-object tracking （MOT） has become a fundamental technique for a wide range of traffic-oriented applications， including road video surveillance， traffic event warning， vehicle behavior analysis， and intelligent traffic management. The goal of vehicle MOT is not only to detect multiple vehicles in video frames， but also to maintain their identities over time and generate reliable trajectories under dynamically changing environments. However， in complex traffic scenes， especially those involving vehicle turning， existing tracking methods often suffer from unstable cross-frame association. When a vehicle turns， its motion direction and spatial position may change abruptly within a short period， while its appearance usually evolves more slowly. This inconsistency in temporal variation makes trajectory state propagation difficult and easily leads to missed detections， identity switches， trajectory fragmentation， and inaccurate re-association. In addition， turning maneuvers frequently cause short-term occlusion or temporary target disappearance. For these short-term lost trajectories， many existing methods simply preserve historical states or rely on delayed matching， without a dedicated update mechanism. As a result， the trajectory representation gradually deviates from the current target state， reducing the robustness of subsequent association. To address these issues， this study proposes an association-enhanced vehicle multi-object tracking method based on trajectory state update， with the aim of improving association stability and continuity in vehicle turning scenarios.MethodThe proposed method is designed from the perspective of trajectory state updating and focuses on two key challenges in turning scenes： unstable state propagation for active trajectories and insufficient state maintenance for short-term lost trajectories. First， to address the mismatch between fast-changing motion cues and slow-changing appearance cues during turning， a fast-slow dual-timescale memory module is introduced. Instead of updating all trajectory information in a single stream， the proposed module models appearance information and motion information separately according to their different temporal dynamics. Specifically， appearance-related features are treated as slow-varying states， while motion-related cues are regarded as fast-varying states. These two types of information are recursively updated in parallel and then fused to obtain a more stable and discriminative trajectory representation. This design reduces the mutual interference between heterogeneous cues during single-stream propagation and improves the robustness of state updating when vehicles undergo rapid directional changes. Second， to handle the problem of short-term lost trajectories， a short-term lost trajectory state update module is proposed. When a target is not detected for a limited period but has not yet been removed from the tracker， directly preserving its previous embedding may cause the stored state to drift away from the target’s actual evolution. To alleviate this issue， the proposed module performs state propagation for short-term lost trajectories by jointly considering three types of information： the most recently updated trajectory embedding， the trend of bounding-box variation， and the duration of disappearance. By integrating these cues， the tracker can estimate a more temporally consistent trajectory state during the missing period， which is more favorable for re-association once the target reappears. In this way， the method enhances the continuity of trajectory representation and reduces the risk of association failure caused by outdated or mismatched states. Overall， the proposed framework improves tracking from the state-update level rather than relying solely on stronger detectors or post-processing association heuristics.ResultExtensive experiments were conducted on the FastTracker， CityFlow， and TrafficMOT datasets to evaluate the effectiveness of the proposed method under different traffic conditions. The experimental results demonstrate that the proposed approach achieves strong overall performance across all three datasets and shows clear advantages on association-related metrics. In particular， the method obtains the best performance on key tracking indicators such as HOTA and AssA， while also showing superior behavior on identity- and fragmentation-related measures including IDs and Frag. In addition， the method exhibits favorable trends on trajectory quality metrics such as MT and ML， indicating that it can track a larger proportion of targets reliably and reduce the number of severely lost trajectories. These improvements suggest that the proposed trajectory state update strategy is effective in mitigating association instability in turning scenarios. From a qualitative perspective， the method is better able to preserve identity consistency during abrupt maneuvers， maintain track continuity when targets temporarily disappear， and reduce re-association errors caused by state drift. Compared with methods that rely on single-stream trajectory propagation or lack dedicated handling for short-term lost tracks， the proposed approach provides a more stable and adaptive state representation for both active and temporarily missing targets， thereby improving tracking robustness in complex traffic scenes.ConclusionThis study proposes an association-enhanced vehicle multi-object tracking method based on trajectory state update for complex traffic scenes. By introducing a fast-slow dual-timescale memory module and a short-term lost trajectory state update module， the method improves the stability and continuity of trajectory association during vehicle turning. Experimental results verify its effectiveness， and the proposed method can provide reliable trajectory support for applications such as traffic monitoring， event warning， and intelligent transportation management.

Hair reconstruction in three dimensions remains a fundamental challenge in computer vision and computer graphics due to the intricate geometric structure and complex appearance of hair. Its highly anisotropic geometry， fine-scale strand organization， and sensitivity to illumination variations make accurate reconstruction from images particularly difficult in real-world scenarios. These challenges are further amplified under non-ideal lighting conditions， such as underexposure or backlighting， where multi-view images often exhibit inconsistent photometric properties. Moreover， complex hairstyles—including curly， wavy， and afro-like patterns—pose additional difficulties， as existing methods frequently suffer from sparse geometric coverage， missing fine details， and unreliable directional cues， resulting in incomplete or visually implausible reconstructions. Recent progress in 3D reconstruction has introduced anisotropic Gaussian representations， most notably 3D Gaussian Splatting （3DGS）， which model scenes as collections of oriented Gaussian ellipsoids and enable efficient rasterization-based rendering with adaptive density control. While these representations provide a flexible framework for general scene reconstruction， they are insufficient for hair-specific challenges. In regions affected by poor lighting， sparse geometric signals lead to holes in reconstructed hair volumes， and conventional density control mechanisms lack the adaptability required to handle varying hairstyle complexity. Furthermore， the strong coupling between geometry and texture under complex illumination is often overlooked， limiting photorealistic hair reconstruction. To overcome these limitations， we propose a novel multi-view hair reconstruction pipeline that integrates visual-language models （VLMs） as an intelligent guidance module for 3D Gaussian-based reconstruction. The key insight is that VLMs can provide high-level semantic and photometric understanding across multiple views， enabling informed decisions on where and how Gaussian ellipsoids should be densified. Our method leverages VLMs to guide reconstruction through three core tasks. First， the VLM evaluates the lighting quality of multi-view images using a dedicated prompt， identifying views prone to geometric degradation due to unfavorable illumination. This enables targeted densification only in necessary regions， improving geometric completeness while avoiding redundant computation. Second， for views affected by underexposure or backlighting， we trigger a three-stage adaptive density control strategy that generates denser Gaussian ellipsoids， leading to more reliable 2D orientation maps and more accurate hair direction supervision. Third， the VLM estimates hairstyle complexity from multi-view inputs and dynamically determines the maximum Gaussian density required. Complex hairstyles such as tight curls or afro patterns are assigned higher density to capture fine geometric details， while simpler hairstyles are reconstructed with fewer Gaussians. In addition， we introduce a decoupled geometry–texture optimization strategy to mitigate color biases introduced by lighting enhancement. Geometry and color are jointly optimized using enhanced images to ensure stable geometric reconstruction. Geometry is fixed while color is refined using original images， allowing the recovery of natural hair color and illumination effects. This strategy ensures both geometric precision and visual realism. Our pipeline further incorporates refinement inspired by prior Gaussian-based hair reconstruction methods. Adaptive Gaussian replication and splitting are applied selectively to balance hole filling and over-densification. Training is supervised using a combination of L1， SSIM， mask， orientation， perceptual， and edge losses， enabling accurate modeling of geometry， texture， and fine structural details while maintaining training efficiency. We evaluate our method on both self-built and public datasets， comparing against state-of-the-art approaches such as Gaussian Haircut， Im2Haircut， and DiffLocks. Experimental results demonstrate that our method produces more detailed， coherent， and realistic hair reconstructions， particularly under challenging lighting conditions and for complex hairstyles. Quantitative evaluations show consistent improvements over Gaussian Haircut， with PSNR increased by 0.54dB， SSIM improved by 0.004， and LPIPS reduced by 0.01. Ablation studies further validate the effectiveness of VLM-guided densification， lighting-aware adaptation， and decoupled optimization. In summary， we present a VLM-guided 3D Gaussian hair reconstruction framework that robustly handles non-ideal lighting and complex hairstyles. By integrating semantic analysis， adaptive densification， and decoupled optimization， our method advances the state of the art in photorealistic 3D hair reconstruction and provides a promising direction for future research.

With the rapid advancement of smart grid construction and intelligent power inspection technology， power image monitoring has become a core technical means for equipment fault detection， line state assessment and power operation safety supervision. Power inspection images collected by wireless terminal devices contain key semantic information such as insulator state， line asset distribution and equipment operating conditions， which directly determines the accuracy and efficiency of subsequent intelligent recognition and fault diagnosis. However， traditional power image monitoring and transmission schemes adopt separated processing modes of signal transmission and target recognition， which lack joint optimization design at the semantic level of image features. Most conventional methods merely focus on the optimization of pixel-level image transmission quality or shallow feature extraction， failing to fully mine， retain and transmit the core semantic features unique to power images during wireless channel transmission. In complex electromagnetic operation environments of power systems， wireless transmission channels are susceptible to random noise interference， signal attenuation and channel fading， which easily cause the loss of effective power semantic information， blurring of key target features and confusion of background information. These inherent defects ultimately lead to low target recognition accuracy， poor anti-interference performance and unsatisfactory overall monitoring efficiency of existing power image monitoring technologies， severely restricting the reliable application of intelligent inspection equipment in complex power operation scenarios. Aiming at the low monitoring accuracy and poor noise resistance capability of traditional power image monitoring methods under low signal-to-noise ratio （SNR） and complex electromagnetic interference conditions， this paper proposes an end-to-end power image monitoring method based on semantic communication joint optimization， which integrates semantic feature transmission and intelligent target recognition into a unified framework to break the performance bottleneck of separated traditional schemes.In terms of methodological design， this paper constructs a novel end-to-end joint optimization framework centering on semantic communication and power image intelligent recognition. Firstly， against the traditional transmission model’s defect of ignoring semantic feature priority， a lightweight convolutional autoencoder （CAE） transmission model is constructed under the complete semantic communication framework. To solve the problem of ineffective extraction and loss of key power semantic features in the transmission process， a channel-based attention mechanism module is embedded into the CAE transmission model. This module can adaptively assign weight coefficients to different channel features， suppress the interference of invalid background noise features， and force the model to focus on extracting and retaining core power semantic features such as power line contours， insulator structures and equipment defect characteristics， thereby realizing semantic-aware optimized transmission of power images. Secondly， aiming at the insufficient local semantic feature perception capability of mainstream recognition models for power images， a vision transformer （ViT） based intelligent recognition backbone model is designed. On the basis of the standard ViT architecture， an attention-guided convolutional coding block module is innovatively proposed and embedded into the recognition network. This module combines the local feature extraction advantages of convolution and the global semantic modeling capability of transformer， effectively compensating for the deficiency of pure ViT models in capturing fine-grained local semantic features of power images， and significantly improving the model’s understanding and discrimination ability of subtle defect features and power equipment semantic information. On this basis， the CAE transmission model and the improved ViT recognition model are organically integrated to build a unified CAE-ViT power image monitoring framework. To further realize deep coupling and global optimization of transmission and recognition links， a targeted joint training strategy is designed according to the parameter space expansion characteristics and semantic feature optimization direction of the proposed framework. This training method synchronously optimizes the transmission coding parameters and recognition network parameters， eliminates the feature mismatch problem caused by separate training of transmission and recognition models， and finally achieves semantic communication joint optimization for full-process power image monitoring.To verify the feasibility， robustness and superiority of the proposed method in complex noise interference scenarios， comprehensive comparative experiments are conducted on two authoritative public power image datasets， namely the chinese power line insulator dataset （CPLID） and the inspection power line asset dataset （InsPLAD）. The experimental test scenarios cover a typical low SNR range from -10dB to 10dB， which simulates the actual complex electromagnetic noise interference environment of field power inspection. A variety of mainstream existing power image monitoring and recognition methods are selected as baseline comparison schemes for horizontal performance evaluation. The experimental results demonstrate that compared with all comparative methods， the proposed semantic communication joint optimization method achieves a significant improvement in power image monitoring and recognition accuracy， with the overall recognition rate increased by 4% to 26% under low SNR conditions. The performance improvement is more obvious in extreme low-noise environments， which fully proves that the proposed method can effectively resist wireless channel noise interference， retain key power semantic information stably， and avoid feature loss and recognition failure caused by channel distortion.In conclusion， the proposed CAE-ViT joint optimization method exhibits excellent channel robustness and semantic feature retention capability in complex electromagnetic wireless transmission environments. It fundamentally solves the problems of insufficient semantic feature extraction， serious feature loss and low recognition accuracy existing in traditional separated power image monitoring schemes， and greatly improves the overall intelligent monitoring efficiency and environmental adaptability of power image inspection. This method provides an effective and feasible technical solution for high-precision and high-reliability power image monitoring under complex working conditions， and offers a valuable theoretical reference and engineering application basis for the popularization and deployment of intelligent power inspection， power equipment fault diagnosis and smart grid visual monitoring systems in complex electromagnetic scenarios.

Object detection in foggy scenes is an important research topic in computer vision， aiming to improve the perception capability of intelligent systems under low-visibility conditions. However， fog-induced degradation， including low contrast， blurred boundaries， weakened texture details， and severe background interference， significantly reduces feature discriminability and often leads to missed detections and false alarms. In practical applications such as autonomous driving， intelligent transportation systems， unmanned aerial vehicle surveillance， and security monitoring， reliable object detection under adverse weather conditions is essential for ensuring environmental awareness and decision-making safety. Nevertheless， atmospheric scattering not only obscures object appearance but also weakens the distinction between foreground targets and background regions， making feature extraction and target localization considerably more challenging. Existing approaches mainly improve detection performance through image dehazing or network architecture optimization. However， dehazing-based methods usually introduce additional computational burden and may generate artifacts that negatively affect subsequent detection， while many detection-oriented methods still suffer from insufficient feature representation， inadequate multi-scale information fusion， and weak robustness to small or low-contrast objects in complex foggy environments. To address these issues， this paper proposes a feature-enhanced object detection network named SD-YOLO11 based on YOLO11. The proposed method improves the network from both the backbone and neck components to enhance the representation and fusion of degraded features under foggy conditions. Specifically， a Spatial Interaction Block （SIBlock） is introduced into the backbone network to strengthen the modeling capability for blurred targets and weak texture regions. By recursively aggregating spatial contextual information， SIBlock enhances high-order spatial interactions and improves the network's ability to capture structural continuity and semantic correlations in degraded scenes. Compared with conventional convolution operations that mainly focus on local receptive fields， the recursive interaction mechanism enables the network to establish richer contextual dependencies and improves the representation of degraded object features. To further improve computational efficiency while preserving feature quality， Ghost Convolution （GhostConv） is adopted to reduce redundant computation through inexpensive feature generation operations， and an Adaptive Downsampling （ADown） module is employed to enhance feature preservation during the downsampling process. These two modules jointly improve the balance between detection accuracy and computational complexity by reducing parameter redundancy while retaining essential semantic information. In the neck network， an improved feature fusion structure named BifpnCat is designed to enhance information exchange among feature maps of different resolutions through bidirectional feature aggregation and channel-wise fusion. Compared with conventional feature concatenation strategies， the proposed structure facilitates more effective integration of semantic information and spatial details， thereby improving multi-scale feature representation. In addition， a simplified DASI attention mechanism is incorporated into the feature fusion stage to adaptively emphasize target-related responses while suppressing background interference， which enhances the discriminative capability of the network in complex foggy scenes. Furthermore， the feature concatenation pathway is redesigned to retain richer channel information during feature aggregation， and a high-resolution P2 detection head is introduced to improve the detection performance for small and distant objects. The additional detection head allows the network to utilize high-resolution feature maps more effectively， thereby alleviating information loss caused by repeated downsampling operations. Unlike methods that rely on explicit image restoration， the proposed framework focuses on enhancing feature learning directly within the detector， reducing computational overhead while avoiding error propagation from image enhancement to object detection. Through the collaborative optimization of feature extraction， feature preservation， attention enhancement， and multi-scale feature fusion， SD-YOLO11 establishes a more robust detection framework for low-visibility environments. To verify the effectiveness of the proposed method， extensive experiments are conducted on three representative fog-related datasets， namely HazyDet， Foggy Cityscapes， and RTTS. Experimental results demonstrate that SD-YOLO11 achieves 78.8%<inline-formula><alternatives><math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M3"><mi mathvariant="normal">m</mi><mi mathvariant="normal">A</mi><msub><mrow><mi mathvariant="normal">P</mi></mrow><mrow><mn mathvariant="normal">50</mn></mrow></msub></math><graphic specific-use="big" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/E4B73080-F457-4f21-9203-5355D3FB1A35-M001.jpg"><?fx-imagestate width="8.12800026" height="3.21733332"?></graphic><graphic specific-use="small" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/E4B73080-F457-4f21-9203-5355D3FB1A35-M001c.jpg"><?fx-imagestate width="8.12800026" height="3.21733332"?></graphic></alternatives></inline-formula>and 61.4%<inline-formula><alternatives><math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M4"><mi mathvariant="normal">m</mi><mi mathvariant="normal">A</mi><msub><mrow><mi mathvariant="normal">P</mi></mrow><mrow><mn mathvariant="normal">50</mn><mo>-</mo><mn mathvariant="normal">95</mn></mrow></msub></math><graphic specific-use="big" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/E4B73080-F457-4f21-9203-5355D3FB1A35-M002.jpg"><?fx-imagestate width="11.93799973" height="3.21733332"?></graphic><graphic specific-use="small" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/E4B73080-F457-4f21-9203-5355D3FB1A35-M002c.jpg"><?fx-imagestate width="11.93799973" height="3.21733332"?></graphic></alternatives></inline-formula>on the HazyDet dataset， improving the original YOLO11 by 1.4 and 2.0 percentage points， respectively. The ablation experiments further confirm the effectiveness of each proposed component， showing that SIBlock， GhostConv， ADown， BifpnCat， and the DASI attention mechanism all contribute positively to the final detection performance. In addition， the proposed method achieves superior results on the Foggy Cityscapes dataset compared with mainstream one-stage detectors， fog-oriented detection models， and transformer-based detectors. It also demonstrates competitive performance in categories closely related to small and low-contrast objects， indicating strong robustness under controlled fog degradation. Detailed comparative analyses show that the proposed improvements consistently enhance feature representation and multi-scale perception capability， leading to more accurate localization and classification of degraded targets. To further evaluate cross-dataset generalization capability， experiments are performed on the RTTS dataset， which contains real-world foggy traffic scenes. The results show that SD-YOLO11 maintains stable detection accuracy across different visibility conditions and achieves better overall performance than other competing methods. The consistent improvements obtained on both synthetic and real-world datasets demonstrate that the proposed method possesses strong adaptability and robustness under diverse foggy conditions. Moreover， despite the introduction of multiple enhancement modules， the proposed network still preserves real-time inference capability， making it suitable for practical deployment in intelligent transportation systems， autonomous driving platforms， and video surveillance applications operating in adverse weather environments. Overall， the proposed SD-YOLO11 effectively alleviates feature degradation caused by atmospheric scattering through enhanced spatial interaction modeling， lightweight feature extraction， adaptive feature preservation， attention-guided feature enhancement， and improved multi-scale feature fusion. Experimental results on multiple benchmark datasets demonstrate that the proposed method consistently improves detection accuracy， robustness， and generalization ability while maintaining favorable computational efficiency. These findings indicate that SD-YOLO11 provides an effective solution for object detection in foggy scenes and offers a practical reference for visual perception tasks in complex low-visibility environments. The proposed framework also demonstrates the feasibility of improving adverse-weather object detection through detector-oriented feature enhancement rather than relying solely on image restoration， which may provide valuable insights for future research on robust visual perception and intelligent scene understanding in challenging environments.

ObjectiveSegment Anything Model 2 （SAM 2） has demonstrated strong potential for object tracking， benefiting from its highly generalizable segmentation priors and long-term temporal modeling capability. These properties provide reliable spatial support for tracking and help mitigate key challenges such as background confusion and target loss after occlusion and reappearance. However， SAM 2 is not able to explicitly exploit the discriminative spectral information inherent in hyperspectral data. As a result， when the target and background exhibit highly similar spatial textures， mask prediction driven mainly by spatial semantic priors may drift. To address the insufficient target–background separation in hyperspectral object tracking， we propose a spectral-prompt and hybrid-attention method tailored for hyperspectral tracking， which enhances the model’s ability to leverage spectral cues and improves robustness in visually ambiguous scenes.MethodThe proposed HOT framework is built upon Segment Anything Model 2 （SAM 2）， leveraging its high-precision mask generation to accurately delineate objects at the pixel level and thus achieve clean object–background separation in complex scenes. However， since SAM 2 is not designed to explicitly perceive hyperspectral characteristics， it struggles to capture discriminative differences across high-dimensional spectral bands. To address this limitation， we propose a hybrid embedding network that efficiently adapts the model to hyperspectral data distributions， and further introduce a hybrid attention mechanism together with a spectral embedding module. These components transform object spectral signatures into learnable prompt vectors and learn deep inter-band semantic representations， respectively， facilitating a collaborative fusion of spatial， temporal， and spectral information for robust hyperspectral object tracking. Furthermore， to foster research and provide a standardized benchmark for this emerging task， we construct AHOT， the first hyperspectral UAV object tracking dataset. AHOT is collected from real UAV flights and is specifically designed to reflect the unique challenges of low-altitude aerial tracking with hyperspectral imagery， where strong viewpoint changes， platform motion， and complex backgrounds are common.ResultComprehensive experiments against six recent trackers on three benchmarks further validate the overall effectiveness and generalization ability of our approach. We report both AUC and DP_20 to reflect complementary aspects of tracking quality， where AUC measures overall success across overlap thresholds and DP_20 evaluates localization accuracy under a 20-pixel criterion. Across different spectral configurations and scene complexities， our method exhibits consistently strong performance and stable ranking. In particular， it achieves the top DP_20 on HOTC-2024-RedNIR of 0.8416， outperforming the second-best competitor by 4.08 %， while also attaining the highest AUC of 0.6034， indicating improvements not only in point-wise precision but also in overall tracking robustness. On the more challenging AHOT benchmark， the advantage becomes more pronounced： our method reaches DP_20 of 0.8779， exceeding the runner-up by 15.57%， and obtains the best AUC of 0.3885， suggesting enhanced resilience to severe appearance changes， background clutter， and long-term interruptions such as occlusion and object re-emergence. Moreover， on HOT-2024-NIR， our method remains highly competitive with DP_20 of 0.9740 and a better AUC of 0.7648， demonstrating reliable performance in near-infrared scenarios. Overall， these cross-dataset results consistently show that the proposed tracker delivers accurate localization and robust temporal consistency， making it well suited for practical deployment under diverse spectral modalities and challenging real-world conditions. In addition， the inference speed of the method proposed in this paper reaches 15.4 FPS， achieving near real-time engineering efficiency while maintaining high accuracy.ConclusionThis paper presents a hyperspectral object tracking model that effectively integrates the complementary strengths of spectral representation learning and segmentation-based foundation models. The incorporation of spectral embedding enhances the model’s ability to perceive and exploit high-dimensional inter-band differences， improving discrimination against visually similar but materially different distractors. Coupled with long-term temporal association modeling， the tracker demonstrates stronger robustness to severe appearance changes， deformation， and long-duration occlusions， and is able to re-localize targets more reliably after disappearance and re-emergence. Overall， the proposed method delivers more accurate and stable tracking performance across diverse hyperspectral scenarios， highlighting the effectiveness of jointly modeling spatial， temporal， and spectral cues for HOT.

ObjectiveHyperspectral image （HSI） classification is an important task in remote sensing image interpretation because HSI data contain continuous spectral responses that are useful for fine-grained land-cover recognition. Existing deep learning methods， especially convolutional neural networks （CNNs）， have achieved high classification accuracy under the closed-set assumption. However， this assumption requires that all test classes have been observed during training， which is often inconsistent with real remote sensing applications. In large-scale and dynamic scenes， a trained model may encounter undefined or newly emerging land-cover categories. Conventional closed-set classifiers usually force such unknown samples into one of the known classes and may produce overconfident predictions. In addition， HSI data are affected by illumination variation， atmospheric scattering， sensor noise， and mixed pixels， resulting in the phenomena of same material with different spectra and different materials with similar spectra. These factors make hyperspectral open-set recognition （OSR） more challenging than ordinary closed-set classification. To address these problems， this paper proposes a frequency knowledge-guided contrastive learning method， called FKCL， for hyperspectral open-set recognition. The main purpose is to improve the rejection ability for unknown land-cover classes while maintaining reliable classification performance for known classes.MethodThe proposed FKCL framework integrates data-driven spatial-spectral representation learning with frequency-domain physical prior knowledge. The whole network consists of two collaborative branches. The first branch is a spatial-spectral feature extraction branch based on a 3D-2D hybrid convolutional neural network. It uses 3D convolution to capture local spectral correlations and then uses 2D convolution to extract spatial contextual information from hyperspectral image patches. The extracted high-level representation is projected into a compact feature space， where class-wise metric learning is performed. The second branch is a frequency-domain physical attribute extraction branch. Instead of using the average spectrum of the whole patch， this branch uses the central pixel spectrum because the category label of a hyperspectral patch usually corresponds to the central pixel. The central spectral sequence is transformed into the frequency domain by real fast Fourier transform （RFFT）， and its magnitude spectrum is used to describe the frequency fluctuation pattern of the ground object. This design preserves the correspondence between the original spectral reflectance curve and its physical frequency representation， and avoids the interference caused by heterogeneous neighboring pixels. To obtain stable class-level frequency representations， an exponential moving average （EMA） mechanism is introduced to update the frequency prototype of each known class during training. The EMA update reduces prototype fluctuation caused by small-batch sampling and gradually accumulates the stable frequency pattern of each known category. For feature-space discrimination， the class anchor clustering （CAC） loss is adopted. The anchor loss pulls samples toward their corresponding class anchors， whereas the tuplet loss pushes samples away from non-corresponding anchors， thereby improving intra-class compactness and inter-class separability. During inference， FKCL uses a dual-distance decision strategy. The semantic feature distance and the frequency-domain prototype distance are fused by a frequency weight selected on the validation set， and the open-set rejection threshold is also determined on the validation set. The test set is used only for final evaluation， so no test information is involved in threshold selection.ResultExperiments were conducted on three public HSI benchmark datasets， namely WHU-Hi-LongKou， Pavia University， and Salinas. A class-holdout protocol was adopted to construct pseudo-unknown classes. In this protocol， selected categories are completely removed from training and appear only during validation and testing for unknown rejection evaluation. For WHU-Hi-LongKou， the Mixed weed class was held out as the pseudo-unknown class. For Pavia University， the Shadows class was held out. For Salinas， Stubble， Lettuce_romaine_6wk， Lettuce_romaine_7wk， Vinyard_untrained， and Vinyard_vertical_trellis were treated as pseudo-unknown classes. The proposed FKCL method was compared with six representative closed-set and open-set methods， including DBDA， SSFTT， CACL， SSMTL， HyperCASR， and SS-MTr. The evaluation metrics included overall accuracy （OA）， average accuracy （AA）， Kappa coefficient， area under the receiver operating characteristic curve （AUROC）， and unknown class accuracy. Experimental results show that FKCL achieves a better balance between known-class classification and unknown-class rejection. On WHU-Hi-LongKou， FKCL achieved an OA of 92.38%， an AA of 92.58%， and a Kappa coefficient of 90.08%， while its unknown class accuracy reached 25.20%， which was higher than most deep closed-set comparison models. On Pavia University， FKCL achieved an OA of 92.02%， an AA of 89.13%， a Kappa coefficient of 89.48%， an AUROC of 99.96%， and an unknown class accuracy of 100.00%， indicating that the frequency-guided dual-distance strategy can effectively identify the held-out Shadows class. On Salinas， the proposed method maintained high classification performance for known classes， but the rejection of pseudo-unknown samples remained more difficult because several unknown classes had very high spectral similarity to known classes. This result indicates that the open-set difficulty is closely related to the physical and spectral separability between known and unknown categories. The ablation analysis further shows that the frequency-domain branch， frequency prototype learning， and dual-distance decision mechanism contribute to the improvement of open-set detection performance. In particular， removing the frequency prototype EMA weakens unknown rejection， demonstrating that stable frequency prototypes are important for modeling class-level physical attributes.ConclusionThe proposed FKCL method introduces frequency-domain knowledge into hyperspectral open-set recognition and constructs a collaborative framework that combines deep spatial-spectral features with physical spectral priors. Compared with single-branch data-driven models， FKCL uses RFFT-based frequency representations and EMA-updated frequency prototypes to provide complementary information for distinguishing spectrally similar but physically different land-cover categories. The CAC-based metric learning module constructs a compact and separable feature space for known classes， and the dual-distance decision strategy further improves the ability to reject samples outside the known-class distribution. Under the class-holdout pseudo-unknown evaluation protocol， FKCL reduces the overconfidence problem of conventional deep models and improves the discrimination of unseen categories. Nevertheless， when unknown classes are extremely similar to known classes in both the original spectral domain and the frequency domain， open-set rejection remains challenging. Future work will focus on adaptive frequency-weight learning， more realistic open-world remote sensing scenarios， cross-region generalization， dynamic newly emerging categories， and open-set recognition with multimodal remote sensing data.