Yuhan Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yuhan Sun has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Shengyang Li
Chinese Academy of Sciences
7 shared publications - Han Wang
Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing, China
4 shared publications - Jian Yang
Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing, China
4 shared publications - Yanfeng Gu
Harbin Institute of Technology
3 shared publications - Zhuang Zhou
Technology and Engineering Center for Space Utilization and the Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing, China
3 shared publications - Manqi Zhao
Key Laboratory of Space Utilization, Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 7 April 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) , 12775-12790
CDNet: Consistency-Driven Semi-Supervised Object Detection for Remote Sensing
Remote sensing images exhibit arbitrary orientations, uneven sample distributions, and large scale variations. Existing semi-supervised detectors mainly handle rotation yet neglect how distribution and scale impair pseudo-label quality and positive–negative sample assignment. These limitations manifest in three ways: (1) Inconsistent features. Classification and regression features often mismatch, leading pseudo-labels selected based on classification scores to poorly align with ground truths. (2) Inconsistent sample distributions. A fixed selection threshold yields unstable pseudo-labels due to uneven object distributions, injecting noise into training. (3) Inconsistent discriminative regions. Objects with extreme aspect ratios often have key features at the periphery, causing them to be mislabeled as background. To tackle these issues, we propose CDNet, a novel Consistency-Driven semi-supervised object detection network for remote sensing with four designs: (i) A task-feature alignment module align classification and regression features, allowing classification scores to better reflect pseudo-label quality. (ii) A GMM-based adaptive thresholding strategy that adjusts pseudo-label selection criteria across training stages to cope with uneven distributions. (iii) A sample assignment consistency constraint method and (iv) a scale consistency loss to regularize high-aspect-ratio and scale-varied objects. Extensive experiments on DOTAv1.5 and DIOR-R demonstrate CDNet's superiority over state-of-the-art methods.
- 7 April 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) , 12775-12790
Open AccessCDNet: Consistency-Driven Semi-Supervised Object Detection for Remote Sensing
Remote sensing images exhibit arbitrary orientations, uneven sample distributions, and large scale variations. Existing semi-supervised detectors mainly handle rotation yet neglect how distribution and scale impair pseudo-label quality and positive–negative sample assignment. These limitations manifest in three ways: (1) Inconsistent features. Classification and regression features often mismatch, leading pseudo-labels selected based on classification scores to poorly align with ground truths. (2) Inconsistent sample distributions. A fixed selection threshold yields unstable pseudo-labels due to uneven object distributions, injecting noise into training. (3) Inconsistent discriminative regions. Objects with extreme aspect ratios often have key features at the periphery, causing them to be mislabeled as background. To tackle these issues, we propose CDNet, a novel Consistency-Driven semi-supervised object detection network for remote sensing with four designs: (i) A task-feature alignment module align classification and regression features, allowing classification scores to better reflect pseudo-label quality. (ii) A GMM-based adaptive thresholding strategy that adjusts pseudo-label selection criteria across training stages to cope with uneven distributions. (iii) A sample assignment consistency constraint method and (iv) a scale consistency loss to regularize high-aspect-ratio and scale-varied objects. Extensive experiments on DOTAv1.5 and DIOR-R demonstrate CDNet's superiority over state-of-the-art methods.
- 22 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-15
Language-Empowered Conversion for Remote Sensing Image Retrieval With Text Feedback
Remote sensing image retrieval with text feedback (RSIR-TF) presents a challenging multimodal retrieval task that leverages a reference image, modification text, and scene graph to retrieve the relevant target image from a gallery. Existing approaches rely on cross-modal combiners to integrate multimodal features extracted separately from modality-specific encoders. However, the modality-specific encoders often suffer from insufficient representational capacity and limited cross-modal alignment due to the lack of effective pretraining. Recently, vision language models (VLMs) pretrained on large-scale image-text pairs have demonstrated exceptional representation and alignment capabilities in the remote sensing (RS) domain. However, these VLMs struggle with processing structured scene graphs, limiting their applicability to tasks like RSIR-TF that require composite reasoning over structured and unstructured modalities. To address these limitations, we propose a novel pipeline, language-empowered conversion (LEmpo), which effectively migrates the large language model (LLM) and VLM to the RSIR-TF task. First, we perform pseudo caption generation (PCG) and scene graph interpretation (SGI) powered by LLM to convert structured scene graphs into natural language captions. This conversion bridges the gap between structured scene graphs and unstructured text captions, enabling unified feature extraction and alignment. Subsequently, we employ the pretrained VLM to extract robust visual and textual features within a joint visual-textual feature space. To fully utilizing the complementary information from visual, textual, and structured data, we introduce a hybrid similarity tuning strategy, which aggregates triplet similarity, language similarity, and pseudo caption similarity into a unified hybrid similarity. The hybrid similarity is optimized during training through vision-fixed tuning, which anchors visual features while refining textual features to enhance alignment with target images. Comprehensive experiments conducted on the airplane, tennis, and WHIRT datasets demonstrate that LEmpo significantly outperforms all comparison methods, achieving a substantial improvement in recall performance.
- 22 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-15
Language-Empowered Conversion for Remote Sensing Image Retrieval With Text Feedback
Remote sensing image retrieval with text feedback (RSIR-TF) presents a challenging multimodal retrieval task that leverages a reference image, modification text, and scene graph to retrieve the relevant target image from a gallery. Existing approaches rely on cross-modal combiners to integrate multimodal features extracted separately from modality-specific encoders. However, the modality-specific encoders often suffer from insufficient representational capacity and limited cross-modal alignment due to the lack of effective pretraining. Recently, vision language models (VLMs) pretrained on large-scale image-text pairs have demonstrated exceptional representation and alignment capabilities in the remote sensing (RS) domain. However, these VLMs struggle with processing structured scene graphs, limiting their applicability to tasks like RSIR-TF that require composite reasoning over structured and unstructured modalities. To address these limitations, we propose a novel pipeline, language-empowered conversion (LEmpo), which effectively migrates the large language model (LLM) and VLM to the RSIR-TF task. First, we perform pseudo caption generation (PCG) and scene graph interpretation (SGI) powered by LLM to convert structured scene graphs into natural language captions. This conversion bridges the gap between structured scene graphs and unstructured text captions, enabling unified feature extraction and alignment. Subsequently, we employ the pretrained VLM to extract robust visual and textual features within a joint visual-textual feature space. To fully utilizing the complementary information from visual, textual, and structured data, we introduce a hybrid similarity tuning strategy, which aggregates triplet similarity, language similarity, and pseudo caption similarity into a unified hybrid similarity. The hybrid similarity is optimized during training through vision-fixed tuning, which anchors visual features while refining textual features to enhance alignment with target images. Comprehensive experiments conducted on the airplane, tennis, and WHIRT datasets demonstrate that LEmpo significantly outperforms all comparison methods, achieving a substantial improvement in recall performance.
- 14 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
DFDNet: Deep Feature Decoupling for Oriented Object Detection
Objects in remote sensing images exhibit diverse orientations. Current oriented object detection (OOD) methods estimate object angle by designing different loss functions and bounding box representations. However, these approaches do not account for the effects of coupling between rotation-equivariant and rotation-invariant features on the regression of oriented bounding box (OBB) parameters. We manifest the problem in two aspects, (1) The coupling of parameters with different attributes. Current OOD methods overlook the inherent differences among features representing an object’s location, scale, and angle, making it challenging to accurately predict OBB parameters with different attributes. (2) The coupling of object and background features. Conventional OOD methods apply convolution kernels uniformly across objects and background regions, leading to feature entanglement and degradation in detection performance. To address the above issues, we propose Deep Feature Decoupling Network (DFDNet) to decouple the extracted features. Specifically, we propose Parameter Regression Decoupling (PRD) to separate feature maps based on their attributes, subsequently assigning them to distinct branches for OBB parameter regression. This approach ensures the decoupling of features related to an object’s location, shape, angle, and category. Additionally, to enhance the ability of OOD networks to differentiate between object and background features, we design the Mask Reinforcement Module (MRM), which is integrated into the PRD branches. The MRM dynamically adjusts the weights of object features, suppressing background interference and enhancing the distinction between object and background features. Extensive experiments conducted on the DOTA, HRSC2016, and UCAS-AOD datasets validate the effectiveness of DFDNet, demonstrating that it achieves state-of-the-art performance.
- 14 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
DFDNet: Deep Feature Decoupling for Oriented Object Detection
Objects in remote sensing images exhibit diverse orientations. Current oriented object detection (OOD) methods estimate object angle by designing different loss functions and bounding box representations. However, these approaches do not account for the effects of coupling between rotation-equivariant and rotation-invariant features on the regression of oriented bounding box (OBB) parameters. We manifest the problem in two aspects, (1) The coupling of parameters with different attributes. Current OOD methods overlook the inherent differences among features representing an object’s location, scale, and angle, making it challenging to accurately predict OBB parameters with different attributes. (2) The coupling of object and background features. Conventional OOD methods apply convolution kernels uniformly across objects and background regions, leading to feature entanglement and degradation in detection performance. To address the above issues, we propose Deep Feature Decoupling Network (DFDNet) to decouple the extracted features. Specifically, we propose Parameter Regression Decoupling (PRD) to separate feature maps based on their attributes, subsequently assigning them to distinct branches for OBB parameter regression. This approach ensures the decoupling of features related to an object’s location, shape, angle, and category. Additionally, to enhance the ability of OOD networks to differentiate between object and background features, we design the Mask Reinforcement Module (MRM), which is integrated into the PRD branches. The MRM dynamically adjusts the weights of object features, suppressing background interference and enhancing the distinction between object and background features. Extensive experiments conducted on the DOTA, HRSC2016, and UCAS-AOD datasets validate the effectiveness of DFDNet, demonstrating that it achieves state-of-the-art performance.
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
Fine-Grained Object Detection of Satellite Video in the Frequency Domain
Satellite video objects often have small scales and the occlusion of their distinguishable regions. Existing fine-grained object detection methods estimate object locations and categories by enhancing object features and increasing feature differences between categories. However, they fail to account for the impact of limited feature information on fine-grained detection, specifically reflected in two aspects: (1) Limited pixels lead to limited features. The small pixel coverage of satellite video objects results in a low upper bound on the available feature information, hindering significant improvements in fine-grained detection accuracy. (2) Limited differences exacerbate limitations. Occlusion of distinguishable regions in small-scale objects exacerbates the indistinctness of features between different fine-grained categories. It prevents the network from accurately learning unique features for certain object classes, thereby degrades detector performance. To address these challenges, we propose Frequency Auxiliary Network (FANet), which integrates frequency domain feature learning into fine-grained object detection networks. Specifically, we propose the Spectral Augmented Module (SAM) to extract multispectral features from various frequency components of satellite video frames to complement spatial-domain features, enabling the network to leverage hidden semantic information from the frequency domain. Additionally, to better distinguish small-scale objects from the background and emphasize fine-grained category features, we design the Frequency Domain Attention (FDA) mechanism. FDA assigns dynamic weights to spatial and frequency domain features, suppressing background information and enhancing feature differences between fine-grained categories. Extensive experiments on SAT-MTB dataset demonstrate that FANet achieves superior performance compared to existing methods.
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
Fine-Grained Object Detection of Satellite Video in the Frequency Domain
Satellite video objects often have small scales and the occlusion of their distinguishable regions. Existing fine-grained object detection methods estimate object locations and categories by enhancing object features and increasing feature differences between categories. However, they fail to account for the impact of limited feature information on fine-grained detection, specifically reflected in two aspects: (1) Limited pixels lead to limited features. The small pixel coverage of satellite video objects results in a low upper bound on the available feature information, hindering significant improvements in fine-grained detection accuracy. (2) Limited differences exacerbate limitations. Occlusion of distinguishable regions in small-scale objects exacerbates the indistinctness of features between different fine-grained categories. It prevents the network from accurately learning unique features for certain object classes, thereby degrades detector performance. To address these challenges, we propose Frequency Auxiliary Network (FANet), which integrates frequency domain feature learning into fine-grained object detection networks. Specifically, we propose the Spectral Augmented Module (SAM) to extract multispectral features from various frequency components of satellite video frames to complement spatial-domain features, enabling the network to leverage hidden semantic information from the frequency domain. Additionally, to better distinguish small-scale objects from the background and emphasize fine-grained category features, we design the Frequency Domain Attention (FDA) mechanism. FDA assigns dynamic weights to spatial and frequency domain features, suppressing background information and enhancing feature differences between fine-grained categories. Extensive experiments on SAT-MTB dataset demonstrate that FANet achieves superior performance compared to existing methods.
- 5 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-15
MP2Net: Mask Propagation and Motion Prediction Network for Multiobject Tracking in Satellite Videos
Mainstream multi-object tracking (MOT) algorithms employ global object detection and association methods. However, when dealing with scenarios involving crowded tiny objects in satellite videos, existing global trackers often yield numerous missed detections and unstable trajectories. To address this issue, we propose a novel joint-detection-and-tracking framework, MP 2 Net, which integrates local detection enhancements for tiny targets and bridges the gap between detection and association. Specifically, our approach incorporates a mask propagation network that enhances feature representation for tiny targets by matching frame-by-frame to capture local details. Additionally, we utilize an implicit and explicit motion prediction strategy that merges tracking information into detection at both feature and instance levels, thereby improving tracking robustness. Experimental results on two large-scale datasets demonstrate the effectiveness and robustness of MP 2 Net, achieving state-of-the-art performance on typical moving objects in satellite videos, such as 66.7% MOTA and 75.9% IDF1 on the SatVideoDT challenge dataset. The code will be available at https://github.com/DonDominic/MP2Net.
- 5 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-15
MP2Net: Mask Propagation and Motion Prediction Network for Multiobject Tracking in Satellite Videos
Mainstream multi-object tracking (MOT) algorithms employ global object detection and association methods. However, when dealing with scenarios involving crowded tiny objects in satellite videos, existing global trackers often yield numerous missed detections and unstable trajectories. To address this issue, we propose a novel joint-detection-and-tracking framework, MP 2 Net, which integrates local detection enhancements for tiny targets and bridges the gap between detection and association. Specifically, our approach incorporates a mask propagation network that enhances feature representation for tiny targets by matching frame-by-frame to capture local details. Additionally, we utilize an implicit and explicit motion prediction strategy that merges tracking information into detection at both feature and instance levels, thereby improving tracking robustness. Experimental results on two large-scale datasets demonstrate the effectiveness and robustness of MP 2 Net, achieving state-of-the-art performance on typical moving objects in satellite videos, such as 66.7% MOTA and 75.9% IDF1 on the SatVideoDT challenge dataset. The code will be available at https://github.com/DonDominic/MP2Net.
- 18 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 2238-2251
Satellite Video Multi-Label Scene Classification With Spatial and Temporal Feature Cooperative Encoding: A Benchmark Dataset and Method
Satellite video multi-label scene classification predicts semantic labels of multiple ground contents to describe a given satellite observation video, which plays an important role in applications like ocean observation, smart cities, et al. However, the lack of a high-quality and large-scale dataset prevents further improvement of the task. And existing methods on general videos have the difficulty to represent the local details of ground contents when directly applied to the satellite videos. In this paper, our contributions include (1) we develop the first publicly available and large-scale satellite video multi-label scene classification dataset. It consists of 18 classes of static and dynamic ground contents, 3549 videos, and 141960 frames. (2) we propose a baseline method with the novel Spatial and Temporal Feature Cooperative Encoding (STFCE). It exploits the relations between local spatial and temporal features, and models long-term motion information hidden in inter-frame variations. In this way, it can enhance features of local details and obtain the powerful video-scene-level feature representation, which raises the classification performance effectively. Experimental results show that our proposed STFCE outperforms 13 state-of-the-art methods with a global average precision (GAP) of 0.8106 and the careful fusion and joint learning of the spatial, temporal, and motion features are beneficial to achieve a more robust and accurate model. Moreover, benchmarking results show that the proposed dataset is very challenging and we hope it could promote further development of the satellite video multi-label scene classification task.
- 18 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 2238-2251
Satellite Video Multi-Label Scene Classification With Spatial and Temporal Feature Cooperative Encoding: A Benchmark Dataset and Method
Satellite video multi-label scene classification predicts semantic labels of multiple ground contents to describe a given satellite observation video, which plays an important role in applications like ocean observation, smart cities, et al. However, the lack of a high-quality and large-scale dataset prevents further improvement of the task. And existing methods on general videos have the difficulty to represent the local details of ground contents when directly applied to the satellite videos. In this paper, our contributions include (1) we develop the first publicly available and large-scale satellite video multi-label scene classification dataset. It consists of 18 classes of static and dynamic ground contents, 3549 videos, and 141960 frames. (2) we propose a baseline method with the novel Spatial and Temporal Feature Cooperative Encoding (STFCE). It exploits the relations between local spatial and temporal features, and models long-term motion information hidden in inter-frame variations. In this way, it can enhance features of local details and obtain the powerful video-scene-level feature representation, which raises the classification performance effectively. Experimental results show that our proposed STFCE outperforms 13 state-of-the-art methods with a global average precision (GAP) of 0.8106 and the careful fusion and joint learning of the spatial, temporal, and motion features are beneficial to achieve a more robust and accurate model. Moreover, benchmarking results show that the proposed dataset is very challenging and we hope it could promote further development of the satellite video multi-label scene classification task.
- Open Access
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 16 (19391404) , 6776-6798
Recent Advances in Intelligent Processing of Satellite Video: Challenges, Methods, and Applications
Intelligent processing of satellite video focuses on extracting specific information of ground objects and scenes from Earth observation videos through intelligent image/video processing technology, which has important applications in fields such as traffic monitoring, resource and environmental monitoring. The integration of deep learning technology in satellite video processing has led to significant advancements in tasks like object detection and object tracking, expanding into emerging research areas such as satellite video scene classification, object segmentation. However, there is no comprehensive review and summary in the intelligent processing of satellite video. This paper presents a systematic review and quantitative analysis of the results published over the last decade, intending to further promote the development of various intelligent processing tasks for satellite video. It analyzes the current difficulties, challenges and the methodological system for each task. Additionally, it provides an in-depth analysis and summary of publicly available datasets and evaluation benchmarks for each task, as well as classic algorithm performance and application scenarios. Finally, this paper summarizes the current research status and looks forward to the future development trend, hoping to inspire researchers in related fields and jointly promote the development of intelligent processing of satellite video.
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 16 (19391404) , 6776-6798
Open AccessRecent Advances in Intelligent Processing of Satellite Video: Challenges, Methods, and Applications
Intelligent processing of satellite video focuses on extracting specific information of ground objects and scenes from Earth observation videos through intelligent image/video processing technology, which has important applications in fields such as traffic monitoring, resource and environmental monitoring. The integration of deep learning technology in satellite video processing has led to significant advancements in tasks like object detection and object tracking, expanding into emerging research areas such as satellite video scene classification, object segmentation. However, there is no comprehensive review and summary in the intelligent processing of satellite video. This paper presents a systematic review and quantitative analysis of the results published over the last decade, intending to further promote the development of various intelligent processing tasks for satellite video. It analyzes the current difficulties, challenges and the methodological system for each task. Additionally, it provides an in-depth analysis and summary of publicly available datasets and evaluation benchmarks for each task, as well as classic algorithm performance and application scenarios. Finally, this paper summarizes the current research status and looks forward to the future development trend, hoping to inspire researchers in related fields and jointly promote the development of intelligent processing of satellite video.