Yongqi Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yongqi Sun has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Anzhu Yu
Information Engineering University, Zhengzhou, China
5 shared publications - Hang Pei
Information Engineering University, Zhengzhou, China
4 shared publications Weiqi Lian
4 shared publications- Yujun Quan
Information Engineering University, Zhengzhou, China
4 shared publications - Chenguang Dai
Institute of Geospatial Information, Information Engineering University, Zhengzhou, China
3 shared publications - Yanfei Zhong
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, China
3 shared publications
Latest Publications
- 15 October 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 334
Prompt-guided vision-language model for hyperspectral image classification with limited labeled samples
- 15 October 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 334
Prompt-guided vision-language model for hyperspectral image classification with limited labeled samples
- 26 August 2026
- Published by Elsevier in Information Fusion
- Vol. 138
Decoupled wavelet-guided frequency-spatial fusion for RGB-X semantic segmentation
- 26 August 2026
- Published by Elsevier in Information Fusion
- Vol. 138
Decoupled wavelet-guided frequency-spatial fusion for RGB-X semantic segmentation
- Open Access
- 6 July 2026
- Published by Elsevier in SSRN Electronic Journal
Prompt-Guided Vision-Language Model for Hyperspectral Image Classification with Limited Labeled Samples
Hyperspectral images (HSIs), which uniquely integrate spatial features and continuous spectral signatures, exhibit great potential for fine-grained classification. Existing deep learning methods typically learn a direct end-to-end mapping between spatial-spectral features and ground-truth labels. But it is difficult to fully exploit the spectral physical knowledge inherent in HSIs, such as absorption valleys, reflectance peaks, and overall spectral shape. With the advancement of HSI acquisition techniques, HSI classification under limited labeled samples has become an increasingly critical challenge. Despite possessing rich pre-trained priors and superior visual comprehension capabilities, large vision-language models (LVLMs) are difficult to directly apply to HSIs as they are predominantly pre-trained on RGB images. To bridge this gap and incorporate the classic physical principles of conventional HSI classification, we conduct an in-depth exploration of LVLMs for HSI classification with limited samples and propose a novel LVLM-based spatial-spectral fusion framework. Specifically, we first convert each spectral vector into a spectral curve plot. This allows the model to classify based on spectral shapes, similar to how humans perceive spectra. Also, we extract false color spatial patches from the neighborhood of the central pixel to embed local spatial context. Furthermore, we design task-aware textual prompts and incorporate a classification head to facilitate classification. These designs guide the model to exploit spatial-spectral features and restrict predictions to predefined categories. Extensive experiments on three benchmark datasets illustrate that our method achieves performance comparable with or even superior to state-of-the-art approaches under limited labeled samples. In particular, the fusion of textual prompts and spatial-spectral representations yields a clear improvement in classification performance.
- 6 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrompt-Guided Vision-Language Model for Hyperspectral Image Classification with Limited Labeled Samples
Hyperspectral images (HSIs), which uniquely integrate spatial features and continuous spectral signatures, exhibit great potential for fine-grained classification. Existing deep learning methods typically learn a direct end-to-end mapping between spatial-spectral features and ground-truth labels. But it is difficult to fully exploit the spectral physical knowledge inherent in HSIs, such as absorption valleys, reflectance peaks, and overall spectral shape. With the advancement of HSI acquisition techniques, HSI classification under limited labeled samples has become an increasingly critical challenge. Despite possessing rich pre-trained priors and superior visual comprehension capabilities, large vision-language models (LVLMs) are difficult to directly apply to HSIs as they are predominantly pre-trained on RGB images. To bridge this gap and incorporate the classic physical principles of conventional HSI classification, we conduct an in-depth exploration of LVLMs for HSI classification with limited samples and propose a novel LVLM-based spatial-spectral fusion framework. Specifically, we first convert each spectral vector into a spectral curve plot. This allows the model to classify based on spectral shapes, similar to how humans perceive spectra. Also, we extract false color spatial patches from the neighborhood of the central pixel to embed local spatial context. Furthermore, we design task-aware textual prompts and incorporate a classification head to facilitate classification. These designs guide the model to exploit spatial-spectral features and restrict predictions to predefined categories. Extensive experiments on three benchmark datasets illustrate that our method achieves performance comparable with or even superior to state-of-the-art approaches under limited labeled samples. In particular, the fusion of textual prompts and spatial-spectral representations yields a clear improvement in classification performance.
- 9 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 5626118
Relative Depth-Guided Cross-Domain Semantic Segmentation for High-Resolution Satellite Remote Sensing Images
Satellite remote sensing images exhibit significant cross-domain discrepancies due to variations in sensors, imaging conditions, and regional characteristics. These discrepancies degrade the performance of semantic segmentation models on unseen target domains. Unsupervised Domain Adaptation (UDA) mitigates this problem but requires target domain images during training. Achieving robust cross-domain generalization with foundation models typically relies on large-scale architectures, incurring substantial computational overhead that limits practical deployment. This paper introduces a novel perspective by leveraging zero-shot relative depth maps as domain-invariant geometric priors to guide cross-domain semantic segmentation of satellite remote sensing images. The proposed framework is compatible with most encoder-decoder architectures and makes three key contributions: 1) A unified depth-guided framework for optical remote sensing cross-domain segmentation. 2) A multi-scale contrastive learning module that aligns optical and relative depth features, enhancing optical feature representation. 3) A refinement strategy using shallow relative depth features to guide semantic mask prediction. Experiments across three cross-domain scenarios, including the LoveDA benchmark, a self-constructed circumpolar cross-seasonal dataset, and a cross=-dataset generalization setting, demonstrate that relative depth serves as an effective spatial prior. Notably, the proposed method achieves comparable or superior accuracy to state-of-the-art UDA approaches that require target-domain images during training, while operating under the stricter zero-shot domain generalization setting where no target data is accessible. The codes will be released at https://github.com/ArthasMil/DGCD.
- 9 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 5626118
Relative Depth-Guided Cross-Domain Semantic Segmentation for High-Resolution Satellite Remote Sensing Images
Satellite remote sensing images exhibit significant cross-domain discrepancies due to variations in sensors, imaging conditions, and regional characteristics. These discrepancies degrade the performance of semantic segmentation models on unseen target domains. Unsupervised Domain Adaptation (UDA) mitigates this problem but requires target domain images during training. Achieving robust cross-domain generalization with foundation models typically relies on large-scale architectures, incurring substantial computational overhead that limits practical deployment. This paper introduces a novel perspective by leveraging zero-shot relative depth maps as domain-invariant geometric priors to guide cross-domain semantic segmentation of satellite remote sensing images. The proposed framework is compatible with most encoder-decoder architectures and makes three key contributions: 1) A unified depth-guided framework for optical remote sensing cross-domain segmentation. 2) A multi-scale contrastive learning module that aligns optical and relative depth features, enhancing optical feature representation. 3) A refinement strategy using shallow relative depth features to guide semantic mask prediction. Experiments across three cross-domain scenarios, including the LoveDA benchmark, a self-constructed circumpolar cross-seasonal dataset, and a cross=-dataset generalization setting, demonstrate that relative depth serves as an effective spatial prior. Notably, the proposed method achieves comparable or superior accuracy to state-of-the-art UDA approaches that require target-domain images during training, while operating under the stricter zero-shot domain generalization setting where no target data is accessible. The codes will be released at https://github.com/ArthasMil/DGCD.
- Open Access
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 19374-19391
VFM-CAKD: Category-Aware Knowledge Distillation From Visual Foundation Model for Aerial Imagery Semantic Segmentation
Aerial image semantic segmentation is a cornerstone technology for low-altitude economy applications, but lightweight models-indispensable for real-time inference on resource-constrained platforms like drones-suffer from limited feature representation capacity. Visual Foundation Models (VFMs), with robust and generalizable visual semantics learned from massive data, offer great potential for enhancing lightweight models via knowledge distillation (KD). However, direct application of VFM-based KD faces three unaddressed challenges: task-irrelevant knowledge interference in VFMs, inherent intra-class variations in aerial imagery, and inadequate long-range dependency modeling of lightweight networks. To tackle these issues, we propose VFM-CAKD, a category-aware knowledge distillation method tailored for lightweight aerial image semantic segmentation. First, we introduce learnable category-aware vectors to filter task-irrelevant features from VFMs, generating Category-Aware Guidance (CAG) to prioritize task-critical semantics. Second, a Category-Aware-Guided Dual-Scale Feature Fusion (CAG-DSFusion) module is designed to activate category-aware representations, extracting stable intra-class features and robust prototypes to mitigate intra-class variations. Third, a Category-Aware-Guided Global Attention (CAG-Global) module aligns teacher-student global attention maps, transferring VFMs' long-range dependency modeling capability. Experiments on drone-view datasets (Aeroscapes, UAVid), aerial remote sensing datasets (Potsdam, Vaihingen), and a disaster-scenario dataset (RescueNet) demonstrate that VFM-CAKD effectively transfers task-relevant VFM knowledge, significantly improving the performance of lightweight segmentation models. The code is avaliable at https://github.com/aresdrw/VFM-CAKD.
- 3 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 19374-19391
Open AccessVFM-CAKD: Category-Aware Knowledge Distillation From Visual Foundation Model for Aerial Imagery Semantic Segmentation
Aerial image semantic segmentation is a cornerstone technology for low-altitude economy applications, but lightweight models-indispensable for real-time inference on resource-constrained platforms like drones-suffer from limited feature representation capacity. Visual Foundation Models (VFMs), with robust and generalizable visual semantics learned from massive data, offer great potential for enhancing lightweight models via knowledge distillation (KD). However, direct application of VFM-based KD faces three unaddressed challenges: task-irrelevant knowledge interference in VFMs, inherent intra-class variations in aerial imagery, and inadequate long-range dependency modeling of lightweight networks. To tackle these issues, we propose VFM-CAKD, a category-aware knowledge distillation method tailored for lightweight aerial image semantic segmentation. First, we introduce learnable category-aware vectors to filter task-irrelevant features from VFMs, generating Category-Aware Guidance (CAG) to prioritize task-critical semantics. Second, a Category-Aware-Guided Dual-Scale Feature Fusion (CAG-DSFusion) module is designed to activate category-aware representations, extracting stable intra-class features and robust prototypes to mitigate intra-class variations. Third, a Category-Aware-Guided Global Attention (CAG-Global) module aligns teacher-student global attention maps, transferring VFMs' long-range dependency modeling capability. Experiments on drone-view datasets (Aeroscapes, UAVid), aerial remote sensing datasets (Potsdam, Vaihingen), and a disaster-scenario dataset (RescueNet) demonstrate that VFM-CAKD effectively transfers task-relevant VFM knowledge, significantly improving the performance of lightweight segmentation models. The code is avaliable at https://github.com/aresdrw/VFM-CAKD.
- Open Access
- 23 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 13877-13894
Instance Prompt Meets Dynamic Adapter: Continual Test-Time Adaptation for Urban Scene Image Segmentation
Recently, to enhance the ability of deployed models to adapt to continually changing environments, the continual test time adaptation (CTTA) for urban scene images (USIs) segmentation has been preliminarily studied. CTTA aims to improve the performance of models while continually adapting to multiple different USIs domains that arrive online in sequence. However, due to the lack of effective solutions for error accumulation and catastrophic forgetting, existing methods fail to achieve satisfactory performance. To this end, this paper proposes a novel CTTA method featuring instance prompts and dynamic adapters. The key innovations include: (1) An instance-wise prompt generator based on the depth anything model is proposed, which can provided more precise prompt information at different levels and effectively alleviate error accumulation. (2) A dynamic adapter module is proposed, where the short-term and long-term adapters learn domain-specific and domain-invariant knowledge respectively, to quickly adapt to new domains while avoiding catastrophic forgetting. (3) An uncertainty estimation strategy with multiple prediction is designed to further optimize the CTTA of USIs. Two typical CTTA scenarios, the continual discrepancies in geographical regions and weather conditions, are established using unmanned aerial vehicle images and street-view images respectively. Extensive experiments indicate that the proposed method outperforms existing methods by a sizeable margin, with an improvement of at least 1.8%∼2.1%.
- 23 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 13877-13894
Open AccessInstance Prompt Meets Dynamic Adapter: Continual Test-Time Adaptation for Urban Scene Image Segmentation
Recently, to enhance the ability of deployed models to adapt to continually changing environments, the continual test time adaptation (CTTA) for urban scene images (USIs) segmentation has been preliminarily studied. CTTA aims to improve the performance of models while continually adapting to multiple different USIs domains that arrive online in sequence. However, due to the lack of effective solutions for error accumulation and catastrophic forgetting, existing methods fail to achieve satisfactory performance. To this end, this paper proposes a novel CTTA method featuring instance prompts and dynamic adapters. The key innovations include: (1) An instance-wise prompt generator based on the depth anything model is proposed, which can provided more precise prompt information at different levels and effectively alleviate error accumulation. (2) A dynamic adapter module is proposed, where the short-term and long-term adapters learn domain-specific and domain-invariant knowledge respectively, to quickly adapt to new domains while avoiding catastrophic forgetting. (3) An uncertainty estimation strategy with multiple prediction is designed to further optimize the CTTA of USIs. Two typical CTTA scenarios, the continual discrepancies in geographical regions and weather conditions, are established using unmanned aerial vehicle images and street-view images respectively. Extensive experiments indicate that the proposed method outperforms existing methods by a sizeable margin, with an improvement of at least 1.8%∼2.1%.
- Open Access
- 6 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 178
P3D: Plug-and-play prompt-driven framework for RGB-thermal semantic segmentation
- 6 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 178
Open AccessP3D: Plug-and-play prompt-driven framework for RGB-thermal semantic segmentation
- Open Access
- 13 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 7543-7559
Toward More Adaptive Fusion: Sample- and Region-Aware Dynamic Fusion Network for Full- and Missing-Modalities Land Cover Classification
In recent years, research on land cover classification (LCC) utilizing optical (OPT) and synthetic aperture radar (SAR) images under full- and missing-modalities scenarios has garnered considerable attention. Nevertheless, existing methods face challenges in sufficiently fusing multimodal features and re constructing missing-modality features. Specifically, conventional fusion methods typically utilize fixed, uniform, and inflexible strategies across all sample pairs, thereby lacking the capacity to select the most targeted fusion method for distinct pairs. Furthermore, these methods rely unconditionally on generated features without considering variations in generation quality across different regions. To address these limitations, this pa per proposes the Sample- and Region-Aware Dynamic Fusion Network (SRDFNet), an architecture that enables more adaptive fusion of features at both the sample and region levels. First, the Multi-Expert Adaptive Fusion (MAF) module is proposed, where each expert functions as an independent fusion unit, allowing dynamic activation of expert combinations tailored to individual sample pairs and thus facilitating adaptive feature fusion at the sample level. Second, the Confidence-based Adaptive Feature Generation (CAFG) module is proposed, which simultaneously produces confidence maps during feature generation to evaluate the reliability of generated features across regions, thereby enabling their refinement. Ultimately, a fusion mechanism in tegrates the refined features with regular modality features, enabling adaptive fusion at the region level. Experiments on the WHU-OPT-SAR and DFC2020 datasets demonstrate that SRDFNet achieves an improvement of 2.30% to 2.90% in mean Intersection over Union (mIoU) under full-modality scenarios, and an enhancement of 0.90% to 3.30% in scenarios with missing-modality. These results substantiate the robustness and effectiveness of the SRDFNet.
- 13 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 7543-7559
Open AccessToward More Adaptive Fusion: Sample- and Region-Aware Dynamic Fusion Network for Full- and Missing-Modalities Land Cover Classification
In recent years, research on land cover classification (LCC) utilizing optical (OPT) and synthetic aperture radar (SAR) images under full- and missing-modalities scenarios has garnered considerable attention. Nevertheless, existing methods face challenges in sufficiently fusing multimodal features and re constructing missing-modality features. Specifically, conventional fusion methods typically utilize fixed, uniform, and inflexible strategies across all sample pairs, thereby lacking the capacity to select the most targeted fusion method for distinct pairs. Furthermore, these methods rely unconditionally on generated features without considering variations in generation quality across different regions. To address these limitations, this pa per proposes the Sample- and Region-Aware Dynamic Fusion Network (SRDFNet), an architecture that enables more adaptive fusion of features at both the sample and region levels. First, the Multi-Expert Adaptive Fusion (MAF) module is proposed, where each expert functions as an independent fusion unit, allowing dynamic activation of expert combinations tailored to individual sample pairs and thus facilitating adaptive feature fusion at the sample level. Second, the Confidence-based Adaptive Feature Generation (CAFG) module is proposed, which simultaneously produces confidence maps during feature generation to evaluate the reliability of generated features across regions, thereby enabling their refinement. Ultimately, a fusion mechanism in tegrates the refined features with regular modality features, enabling adaptive fusion at the region level. Experiments on the WHU-OPT-SAR and DFC2020 datasets demonstrate that SRDFNet achieves an improvement of 2.30% to 2.90% in mean Intersection over Union (mIoU) under full-modality scenarios, and an enhancement of 0.90% to 3.30% in scenarios with missing-modality. These results substantiate the robustness and effectiveness of the SRDFNet.
- Open Access
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
Few-View DSM Generation of Stereo Satellite Imagery via NeRF
Digital surface model (DSM) generation from stereo satellite imagery is a crucial technique for acquiring Earth’s surface information. Traditional stereo matching methods struggle to capture fine details in reconstructing complex geometric scenes. While neural radiance fields (NeRFs) and 3-D Gaussian splatting (3DGS) achieve high-quality reconstructions, they rely on the number of views and are primarily designed for RGB imagery. For panchromatic satellite imagery with few views and weak textures, we propose an NeRF method for few-view stereo satellite reconstruction. We introduce high-precision dense point clouds as geometric prior information to enhance geometric constraint capabilities from few-view images. Then, a continuous implicit function is used to jointly model the geometry and appearance of the scene. In addition, we use a multitask joint optimization strategy for enhancing DSM accuracy. To evaluate the performance of our method, we conducted qualitative and quantitative experiments on GF7 and WorldView-3 images under few-view conditions. Experimental results demonstrate that our method achieves the best results. Furthermore, we analyzed the impact of the number of views on DSM accuracy, confirming that our method maintains higher accuracy and completeness as the number of views decreases.
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
Open AccessFew-View DSM Generation of Stereo Satellite Imagery via NeRF
Digital surface model (DSM) generation from stereo satellite imagery is a crucial technique for acquiring Earth’s surface information. Traditional stereo matching methods struggle to capture fine details in reconstructing complex geometric scenes. While neural radiance fields (NeRFs) and 3-D Gaussian splatting (3DGS) achieve high-quality reconstructions, they rely on the number of views and are primarily designed for RGB imagery. For panchromatic satellite imagery with few views and weak textures, we propose an NeRF method for few-view stereo satellite reconstruction. We introduce high-precision dense point clouds as geometric prior information to enhance geometric constraint capabilities from few-view images. Then, a continuous implicit function is used to jointly model the geometry and appearance of the scene. In addition, we use a multitask joint optimization strategy for enhancing DSM accuracy. To evaluate the performance of our method, we conducted qualitative and quantitative experiments on GF7 and WorldView-3 images under few-view conditions. Experimental results demonstrate that our method achieves the best results. Furthermore, we analyzed the impact of the number of views on DSM accuracy, confirming that our method maintains higher accuracy and completeness as the number of views decreases.