Ziyuan Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ziyuan Liu has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Yuantao Gu
Department of Electronic Engineering, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China
9 shared publications - Jiawei Zhang
Department of Electronic Engineering, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China
4 shared publications - Wenyu Wang
Army Engineering University of PLA
4 shared publications - Lei Zhu
Army Engineering University of PLA
2 shared publications - Shaoping Wang
Tsinghua Shenzhen International Graduate School, Shenzhen, China
2 shared publications - Yufan Chen
College of Communications Engineering, Army Engineering University of PLA, Nanjing, China
2 shared publications
Latest Publications
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 2510217
JL1-CD: A New Benchmark and a Robust Multiteacher Knowledge Distillation Framework for Remote Sensing Change Detection
hange detection (CD) in remote sensing images plays a vital role in dynamic Earth measurement.hange detection (CD) in remote sensing images plays a vital role in dynamic Earth measurement.C However, the scarcity of high-resolution, comprehensive open-source datasets and the difficulty in achieving robust performance across varying change types remain major challenges. To address these issues, we introduce JL1-CD, a large-scale, sub-meter CD dataset consisting of 5,000 image pairs. We further propose a novel Origin-Partition (O-P) strategy and integrate it into a Multi-Teacher Knowledge Distillation (MTKD) framework to enhance CD performance. The O-P strategy partitions the training set by Change Area Ratio (CAR) and trains specialized teacher models on each subset. The MTKD framework then distills complementary knowledge from these teachers into a single student model, enabling improved detection results across diverse CAR scenarios without additional inference cost. Our MTKD approach demonstrated strong performance in the 2024 “Jilin-1” Cup challenge, ranking first in the preliminary and second in the final rounds. Extensive experiments on the JL1-CD and SYSU-CD datasets show that the MTKD framework consistently improves the performance of CD models with various network architectures and parameter sizes, establishing new state-of-the-art results. Code and dataset are available at https://github.com/circleLZY/MTKD-CD.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 2510217
JL1-CD: A New Benchmark and a Robust Multiteacher Knowledge Distillation Framework for Remote Sensing Change Detection
hange detection (CD) in remote sensing images plays a vital role in dynamic Earth measurement.hange detection (CD) in remote sensing images plays a vital role in dynamic Earth measurement.C However, the scarcity of high-resolution, comprehensive open-source datasets and the difficulty in achieving robust performance across varying change types remain major challenges. To address these issues, we introduce JL1-CD, a large-scale, sub-meter CD dataset consisting of 5,000 image pairs. We further propose a novel Origin-Partition (O-P) strategy and integrate it into a Multi-Teacher Knowledge Distillation (MTKD) framework to enhance CD performance. The O-P strategy partitions the training set by Change Area Ratio (CAR) and trains specialized teacher models on each subset. The MTKD framework then distills complementary knowledge from these teachers into a single student model, enabling improved detection results across diverse CAR scenarios without additional inference cost. Our MTKD approach demonstrated strong performance in the 2024 “Jilin-1” Cup challenge, ranking first in the preliminary and second in the final rounds. Extensive experiments on the JL1-CD and SYSU-CD datasets show that the MTKD framework consistently improves the performance of CD models with various network architectures and parameter sizes, establishing new state-of-the-art results. Code and dataset are available at https://github.com/circleLZY/MTKD-CD.
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Wireless Communications Letters
- Vol. 15 (21622337) , 1543-1547
Boundary-Aware Margin Shaping for Joint Open-Set Modulation Recognition and Adversarial Robustness
Robust automatic modulation recognition (AMR) under open-set conditions and adversarial attacks remains a critical challenge for secure and reliable wireless communication. This letter proposes a Boundary-aware Margin Shaping (BMS) framework that unifies open-set recognition and adversarial robustness through adaptive boundary modeling. BMS introduces a margin-aware training scheme that jointly stabilizes feature representations, calibrates confidence, and smooths decision boundaries under adversarial conditions. For inference, we employ a standard KNN detector to perform out-of-distribution (OOD) separation. Comprehensive experiments on the RML2016.10a dataset demonstrate that BMS consistently achieves superior classification and OOD detection performance across various SNRs and adversarial attacks.
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Wireless Communications Letters
- Vol. 15 (21622337) , 1543-1547
Boundary-Aware Margin Shaping for Joint Open-Set Modulation Recognition and Adversarial Robustness
Robust automatic modulation recognition (AMR) under open-set conditions and adversarial attacks remains a critical challenge for secure and reliable wireless communication. This letter proposes a Boundary-aware Margin Shaping (BMS) framework that unifies open-set recognition and adversarial robustness through adaptive boundary modeling. BMS introduces a margin-aware training scheme that jointly stabilizes feature representations, calibrates confidence, and smooths decision boundaries under adversarial conditions. For inference, we employ a standard KNN detector to perform out-of-distribution (OOD) separation. Comprehensive experiments on the RML2016.10a dataset demonstrate that BMS consistently achieves superior classification and OOD detection performance across various SNRs and adversarial attacks.
- Open Access
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. PP (19391404) , 1-13
UnlearningCD: Distill to Forget in Change Detection via Perturbed Knowledge
Change detection (CD) in remote sensing plays a vital role in Earth observation. As deep learning models and associated data grow exponentially, privacy concerns, outdated or erroneous annotations have intensified. Machine unlearning offers a solution by enabling models to selectively discard specific data. However, there is limited exploration of unlearning in remote sensing. This paper introduces an Unlearning algorithm based on Perturbed Knowledge distillation (UPK) specifically designed for CD tasks. Two variants are proposed: closed-box and open-box, addressing scenarios with varying levels of model access. Our approach injects perturbations into teacher models to guide students, achieving targeted unlearning while preserving detection capabilities for other change types. Extensive experiments conducted on semantic CD (SCD) and building damage assessment (BDA) datasets demonstrate that UPK requires only 10% of the training iterations needed for original models and outperforms existing strategies across different CD tasks and model architectures, without incurring additional inference time. This study highlights UPK's potential as a scalable solution for effective model lifecycle management in the rapidly evolving remote sensing fields.
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. PP (19391404) , 1-13
Open AccessUnlearningCD: Distill to Forget in Change Detection via Perturbed Knowledge
Change detection (CD) in remote sensing plays a vital role in Earth observation. As deep learning models and associated data grow exponentially, privacy concerns, outdated or erroneous annotations have intensified. Machine unlearning offers a solution by enabling models to selectively discard specific data. However, there is limited exploration of unlearning in remote sensing. This paper introduces an Unlearning algorithm based on Perturbed Knowledge distillation (UPK) specifically designed for CD tasks. Two variants are proposed: closed-box and open-box, addressing scenarios with varying levels of model access. Our approach injects perturbations into teacher models to guide students, achieving targeted unlearning while preserving detection capabilities for other change types. Extensive experiments conducted on semantic CD (SCD) and building damage assessment (BDA) datasets demonstrate that UPK requires only 10% of the training iterations needed for original models and outperforms existing strategies across different CD tasks and model architectures, without incurring additional inference time. This study highlights UPK's potential as a scalable solution for effective model lifecycle management in the rapidly evolving remote sensing fields.
- 11 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive Communications and Networking
- Vol. 12, 4640-4655
S3AT: Self-Paced, Self-Distilled, and Self-Finetuned Adversarial Training for Robust Automatic Modulation Recognition
Deep learning-based automatic modulation recognition (AMR) models are highly susceptible to adversarial samples. Although adversarial training offers a promising defense, most existing methods rely on fixed-intensity attacks and lack mechanisms to maintain stable feature representations, limiting recognition performance under diverse adversarial conditions. To address these issues, we propose S3AT, an adversarial training framework that integrates Self-paced learning, Self-distillation, and Self-finetuning. In the self-paced phase, a progressive, difficulty-aware training paradigm adaptively schedules adversarial strength to help the model gradually adapt from weak to strong attacks. An exponential moving average (EMA) teacher provides continuous self-distillation throughout training to stabilize feature learning and mitigate forgetting of early knowledge. In the self-finetuning phase, historical checkpoints serve as auxiliary teachers, complementing the self-distillation and enhancing robustness across varying intensities of adversarial attacks. Experiments on public and self-collected datasets confirm that S3AT outperforms baseline methods under both white-box and black-box attacks. Its compatibility with external defenses further highlights its flexibility for real-world electromagnetic applications. Code is available at https://gitee.com/wwenyuu/s3-at.
- 11 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive Communications and Networking
- Vol. 12, 4640-4655
S3AT: Self-Paced, Self-Distilled, and Self-Finetuned Adversarial Training for Robust Automatic Modulation Recognition
Deep learning-based automatic modulation recognition (AMR) models are highly susceptible to adversarial samples. Although adversarial training offers a promising defense, most existing methods rely on fixed-intensity attacks and lack mechanisms to maintain stable feature representations, limiting recognition performance under diverse adversarial conditions. To address these issues, we propose S3AT, an adversarial training framework that integrates Self-paced learning, Self-distillation, and Self-finetuning. In the self-paced phase, a progressive, difficulty-aware training paradigm adaptively schedules adversarial strength to help the model gradually adapt from weak to strong attacks. An exponential moving average (EMA) teacher provides continuous self-distillation throughout training to stabilize feature learning and mitigate forgetting of early knowledge. In the self-finetuning phase, historical checkpoints serve as auxiliary teachers, complementing the self-distillation and enhancing robustness across varying intensities of adversarial attacks. Experiments on public and self-collected datasets confirm that S3AT outperforms baseline methods under both white-box and black-box attacks. Its compatibility with external defenses further highlights its flexibility for real-world electromagnetic applications. Code is available at https://gitee.com/wwenyuu/s3-at.
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 13608-13612
Open-CD: A Comprehensive Toolbox for Change Detection
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 13608-13612
Open AccessOpen-CD: A Comprehensive Toolbox for Change Detection
- Open Access
- 11 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 18 (19391404) , 22642-22655
A Novel Multibranch Self-Distillation Framework for Optimizing Remote Sensing Change Detection
Deep learning (DL) has achieved remarkable success in the field of change detection (CD) for remote sensing images. Existing training methods for DL-based CD models are predominantly single-stage, single-stream, and end-to-end. Despite the numerous optimization techniques proposed, such as advanced network architectures, loss functions, and hyperparameter tuning, these methods still struggle to achieve consistent and satisfactory detection results across images with varying change area ratios (CARs). This raises the critical question: Is the current training paradigm truly optimal? To address this question, we propose a novel multibranch self-distillation training framework. In this framework, different partition branches learn detection patterns under diverse CAR scenarios and guide the main branch through distillation. Our approach consistently enhances the detection accuracy of CD models across various change regions without introducing additional time or computational costs during the inference phase. Extensive experiments on the JL1-CD, SYSU-CD, and CDD datasets demonstrate that the MBSD framework consistently improves the performance of CD models with diverse network architectures and parameter sizes, achieving new state-of-the-art results.
- 11 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 18 (19391404) , 22642-22655
Open AccessA Novel Multibranch Self-Distillation Framework for Optimizing Remote Sensing Change Detection
Deep learning (DL) has achieved remarkable success in the field of change detection (CD) for remote sensing images. Existing training methods for DL-based CD models are predominantly single-stage, single-stream, and end-to-end. Despite the numerous optimization techniques proposed, such as advanced network architectures, loss functions, and hyperparameter tuning, these methods still struggle to achieve consistent and satisfactory detection results across images with varying change area ratios (CARs). This raises the critical question: Is the current training paradigm truly optimal? To address this question, we propose a novel multibranch self-distillation training framework. In this framework, different partition branches learn detection patterns under diverse CAR scenarios and guide the main branch through distillation. Our approach consistently enhances the detection accuracy of CD models across various change regions without introducing additional time or computational costs during the inference phase. Extensive experiments on the JL1-CD, SYSU-CD, and CDD datasets demonstrate that the MBSD framework consistently improves the performance of CD models with diverse network architectures and parameter sizes, achieving new state-of-the-art results.
- 21 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
M 2 CD: A Unified MultiModal Framework for Optical-SAR Change Detection With Mixture of Experts and Self-Distillation
Most existing change detection (CD) methods focus on optical images captured at different times, and deep learning (DL) has achieved remarkable success in this domain. However, in extreme scenarios such as disaster response, synthetic aperture radar (SAR), with its active imaging capability, is more suitable for providing post-event data. This introduces new challenges for CD methods, as existing weight-sharing Siamese networks struggle to effectively learn the cross-modal data distribution between optical and SAR images. To address this challenge, we propose a unified MultiModal CD framework, M2CD. We integrate Mixture of Experts (MoE) modules into the backbone to explicitly handle diverse modalities, thereby enhancing the model’s ability to learn multimodal data distributions. Additionally, we innovatively propose an Optical-to-SAR path (O2SP) and implement self-distillation during training to reduce the feature space discrepancy between different modalities, further alleviating the model’s learning burden. We design multiple variants of M2CD based on both CNN and Transformer backbones. Extensive experiments validate the effectiveness of the proposed framework, with the MiT-b1 version of M2CD outperforming all state-of-the-art (SOTA) methods in optical-SAR CD tasks. Code is available at https://github.com/circleLZY/M2CD.
- 21 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 22 (1545598X) , 1-5
M 2 CD: A Unified MultiModal Framework for Optical-SAR Change Detection With Mixture of Experts and Self-Distillation
Most existing change detection (CD) methods focus on optical images captured at different times, and deep learning (DL) has achieved remarkable success in this domain. However, in extreme scenarios such as disaster response, synthetic aperture radar (SAR), with its active imaging capability, is more suitable for providing post-event data. This introduces new challenges for CD methods, as existing weight-sharing Siamese networks struggle to effectively learn the cross-modal data distribution between optical and SAR images. To address this challenge, we propose a unified MultiModal CD framework, M2CD. We integrate Mixture of Experts (MoE) modules into the backbone to explicitly handle diverse modalities, thereby enhancing the model’s ability to learn multimodal data distributions. Additionally, we innovatively propose an Optical-to-SAR path (O2SP) and implement self-distillation during training to reduce the feature space discrepancy between different modalities, further alleviating the model’s learning burden. We design multiple variants of M2CD based on both CNN and Transformer backbones. Extensive experiments validate the effectiveness of the proposed framework, with the MiT-b1 version of M2CD outperforming all state-of-the-art (SOTA) methods in optical-SAR CD tasks. Code is available at https://github.com/circleLZY/M2CD.
- 22 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-13
DSRKD: Joint Despecking and Super-Resolution of SAR Images via Knowledge Distillation
Deep learning has achieved success in optical image super-resolution (SR). However, few methods have been proposed for synthetic aperture radar (SAR) images. This is because SAR images inherently suffer from severe speckle noise and their resolution is typically much lower compared to optical images, making SR more challenging. To simultaneously denoise and preserve texture details in SAR image SR tasks, we propose a joint despeckling and super-resolution network via knowledge distillation (DSRKD). By implementing feature distillation in the encoding stage, we enable the student network to obtain noise-free latent variables, thereby restoring clean SR images, and simultaneously enhancing effective supervision for the student using target distillation. This distillation strategy does not incur additional computational costs during inference. Experimental results demonstrate that our algorithm effectively suppresses speckle noise while preserving texture details of images on both synthetic and SAR datasets. Compared to other state-of-the-art algorithms, the proposed method achieves superior performance on both visual results and quantitative metrics like PSNR, SSIM, and other non-reference indicators under various SR scales and speckle intensities.
- 22 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 62 (01962892) , 1-13
DSRKD: Joint Despecking and Super-Resolution of SAR Images via Knowledge Distillation
Deep learning has achieved success in optical image super-resolution (SR). However, few methods have been proposed for synthetic aperture radar (SAR) images. This is because SAR images inherently suffer from severe speckle noise and their resolution is typically much lower compared to optical images, making SR more challenging. To simultaneously denoise and preserve texture details in SAR image SR tasks, we propose a joint despeckling and super-resolution network via knowledge distillation (DSRKD). By implementing feature distillation in the encoding stage, we enable the student network to obtain noise-free latent variables, thereby restoring clean SR images, and simultaneously enhancing effective supervision for the student using target distillation. This distillation strategy does not incur additional computational costs during inference. Experimental results demonstrate that our algorithm effectively suppresses speckle noise while preserving texture details of images on both synthetic and SAR datasets. Compared to other state-of-the-art algorithms, the proposed method achieves superior performance on both visual results and quantitative metrics like PSNR, SSIM, and other non-reference indicators under various SR scales and speckle intensities.
- 9 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 21 (1545598X) , 1-5
SAR Image Compression With Inherent Denoising Capability Through Knowledge Distillation
Due to its inherent characteristics, the synthetic aperture radar (SAR) image is mainly corrupted by speckle noise, posing additional challenges to lossy image compression algorithms. Traditional optical image compression techniques lack the ability to distinguish between image details and noise, which increases storage costs and restores images that still contain noise. Inspired by these observations, we optimize image compression algorithms to incorporate denoising capabilities, enabling joint denoising and compression of SAR images. Specifically, we transform the raw speckled images into noise-free bitstreams, allowing the subsequent decompression to produce clean images. To achieve this objective efficiently, we introduce a novel knowledge distillation (KD) strategy that incurs no additional computational cost. Furthermore, this distillation mechanism yields statistically significant performance improvements across various image compression algorithms. Experimental results demonstrate that when evaluated on both synthetic and real-world datasets, the proposed method not only achieves the best visual effects but also outperforms existing methods in terms of rate-distortion performance, equivalent number of looks, and other quantitative indicators.
- 9 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 21 (1545598X) , 1-5
SAR Image Compression With Inherent Denoising Capability Through Knowledge Distillation
Due to its inherent characteristics, the synthetic aperture radar (SAR) image is mainly corrupted by speckle noise, posing additional challenges to lossy image compression algorithms. Traditional optical image compression techniques lack the ability to distinguish between image details and noise, which increases storage costs and restores images that still contain noise. Inspired by these observations, we optimize image compression algorithms to incorporate denoising capabilities, enabling joint denoising and compression of SAR images. Specifically, we transform the raw speckled images into noise-free bitstreams, allowing the subsequent decompression to produce clean images. To achieve this objective efficiently, we introduce a novel knowledge distillation (KD) strategy that incurs no additional computational cost. Furthermore, this distillation mechanism yields statistically significant performance improvements across various image compression algorithms. Experimental results demonstrate that when evaluated on both synthetic and real-world datasets, the proposed method not only achieves the best visual effects but also outperforms existing methods in terms of rate-distortion performance, equivalent number of looks, and other quantitative indicators.