ZLZhe Li

Zhe Li

*This is an algorithmically generated author record based on publicly available information.

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


Zhe Li has a total of 44 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    MFE-Net: Modal-Invariant Feature Extraction Network for Unsupervised Multimodal Change Detection

    by Zhe Li, Detong Wang, Sheng Fang, Xingli Zhang, Miaolin Yang, Yue Song

    Unsupervised multimodal change detection (UMCD) has attracted significant attention due to its practical value and the main challenges it poses. Aligning cross-modal images in a common subspace via deep learning, without decoupling residual modal noise from intermediate features, leads to poor alignment and discrimination errors when comparing features directly. Therefore, learning modal-invariant features that actively remove modal specificity while preserving the essential semantics of land use and land cover(LULC) is crucial. We propose a Modal-invariant Feature Extraction Network (MFE-Net) that learns semantics of LULC through two core modules: the Hybrid CNN-KAN (HCK) encoder and the Modal Difference Learning Module (MDLM). The HCK encoder provides a robust feature foundation for learning modal-invariant features. HCK combines convolutional neural networks (CNNs)' local perception with Kolmogorov-Arnold network (KAN)'s global nonlinear modeling to suppress modal noise and capture semantics, yielding lowmode-correlated features. The MDLM uses adversarial multi-task learning to explicitly separate modal-invariant components from mode-specific noise, ensuring that the HCK encoder generates features that are indistinguishable by modality of origin. To effectively leverage the extracted modal-invariant features to create a robust change map, we designed the Differential Feature Enhancement Module (DFEM) to calculate change intensity by fusing feature-level differences via Spearman's rank correlation. Experiments on six real datasets show MFE-Net outperforms state-of-the-art methods. Code and datasets are available at https://github.com/wangdetong-ctrl/MFE-Net.

  • Open Access

    Intermediary-Focused Multi-Task Learning: Managing Heterogeneous Sources for Semantic Change Detection

    by Sheng Fang, Yue Song, Ruisen Huang, Zhe Li, Xingli Zhang, Jiye Zheng, Jianli Zhao

    Although Multi-Task Learning (MTL) is widely adopted in computer vision for learning shared representations, it struggles with inconsistencies from heterogeneous sources, especially in dense prediction tasks like Semantic Change Detection (SCD). To address this, we propose a novel intermediary-focused MTL framework, exemplified by the Prototypes-Focused SCD Network (ProF-SCD). The ProF-SCD features three key components: a Prototypes-Based Domain Alignment (PDA) module that aligns features across domains using label-supervised prototypes, effectively reducing temporal gaps and preserving true semantic changes; a Double Predictions Semantic Segmentation head (DP-SS); and a Dual Guidance Binary Change Detection head (DG-BCD). The latter two heads utilize auxiliary guidance from the PDA module to enhance their respective tasks. Extensive experiments on real-world datasets show that ProF-SCD achieves significant performance gains, improving mIoU by up to 0.33% and SeK precision by 0.42% over state-of-the-art methods, while using less than onethird of the parameters and half of the FLOPs. These results validate ProF-SCD’s effectiveness, demonstrating the robustness of our intermediary-focused approach and advancing the state of SCD. Our source code is available at https://github.com/sdustmmlab/ProF-SCD .

  • Open Access

    Reliability- and Stability-Aware Prototype Evolution for Semantic Change Detection

    by Ruisen Huang, Sheng Fang, Zhe Li, Jiye Zheng, Xingli Zhang, Yue Song

    In recent years, Semantic Change Detection (SCD) has attracted widespread attention due to its crucial role in high-resolution remote sensing scenarios. However, in complex scenarios, prevalent inter-class similarity makes features hard to distinguish, leading to unreliability, while excessive intraclass variation causes feature drift, resulting in instability. To address the above challenges, this paper proposes the RSPESCD framework, a reliability-and stability-aware prototype evolutionbased semantic change detection method, which comprises three modules. First, the APU performs Adaptive Prototype Updating by combining sample confidence and feature variance, constructing class prototypes with high reliability and stability, and guiding features to converge toward them, thereby enhancing the reliability and stability of the features. Subsequently, to further optimize the features, the PLR and D2Gate modules are designed for the semantic and change branches, respectively. PLR performs Pseudo-Label Refinement through weighting, providing supervisory signals for non-change regions; D2Gate enhances cross-temporal interaction adaptively by leveraging shallow Detail residual information and a Dual-Gating mechanism. Experimental results demonstrate that the proposed method achieves superior performance compared with existing approaches on both the SECOND and Landsat-SCD datasets, thereby validating its effectiveness for SCD. Our source code is released at https://github.com/camellia-forever/RSPESCD.

  • Rethinking Semantic Change Detection From a Semantic Alignment Perspective

    by Sheng Fang, Wen Li, Yue Song, Zhe Li, Jianli Zhao

    Recent advancements in semantic change detection (SCD) often divide the task into two subtasks: semantic segmentation (SS) and binary change detection (BCD), typically employing three decoding heads optimized jointly based on their respective losses. However, this approach faces challenges in aligning semantic feature spaces across temporal domains, complicating the construction of change features. In this article, we propose the novel PRO-HRSCD framework, which utilizes two decoding heads: one dedicated to BCD and another for the SS subtask. The SS subtasks for bi-temporal images are unified within a prototype-based SS (PRO-SS) decoding head, which aligns semantic features across temporal domains using online prototype learning rather than relying solely on joint optimization. The prototype-based contrastive loss targets the semantic features themselves, effectively addressing large intraclass variance and bridging the feature gap between temporal domains. Additionally, a temporal correlation exploration module (TCEM) is integrated into the BCD head, leveraging the highest resolution features of bi-temporal images encoded by HRNet to enhance temporal dependency modeling, leading to mutual reinforcement with PRO-SS. The experimental results on two benchmark datasets demonstrate that our method surpasses state-of-the-art (SOTA) SCD approaches. Our source code is available at https://github.com/sdust-mmlab/PRO-HRSCD


  • 3D-HRSCD: Exploiting the Potential of Multiscale Features by 3-D Convolution

    by Yue Song, Sheng Fang, Zhe Li, Su Wang, Enyi Zhao

    Semantic change detection (SCD) in remote sensing image (RSI) is critical for monitoring land cover and land use transformations. Although existing SCD methods have made progress in modeling temporal dependency, they still struggle to effectively capture multi-scale features and make interaction among them. To address these issues, we propose 3D-HRSCD, a novel architecture that utilizes 3D convolution to model temporal dependency across HRNet’s multi-resolution features. The core of this architecture is 3D Convolution Fusion Oriented to Multi-scale Features (3DFOM) module, which makes adequate interaction in channel, spatial and temporal dimensions across multi-scale features. To support more efficient temporal dependency modeling in 3DFOM, Cosine Similarity-based Temporal Multi-Scales Attention (CTMA) module serves as a preprocessing stage by enhancing features in change regions. Additionally, Comprehensive Semantic Consistency (CSC) loss function is introduced to further suppress pseudo-changes and reduce semantic recognition errors. Experimental results reveal that our method outperforms state-of-the-art performances relative to previous SCD efforts.

  • Open Access

    Rethinking Semantic Change Detection from a Semantic Alignment Perspective

    by Sheng Fang, Wen Li, Yue Song, Zhe Li, Jianli Zhao

    Recent advancements in semantic change detection (SCD) often divide the task into two subtasks: semantic segmentation (SS) and binary change detection (BCD), typically employing three decoding heads optimized jointly based on their respective losses. However, this approach faces challenges in aligning semantic feature spaces across temporal domains, complicating the construction of change features. In this paper, we propose the novel PRO-HRSCD framework, which utilizes two decoding heads: one dedicated to BCD and another for the SS subtask. The SS subtasks for bi-temporal images are unified within a prototype-based SS (PRO-SS) decoding head, which aligns semantic features across temporal domains using online prototype learning rather than relying solely on joint optimization. The prototype-based contrastive loss targets the semantic features themselves, effectively addressing large intraclass variance and bridging the feature gap between temporal domains. Additionally, a temporal correlation exploration module (TCEM) is integrated into the BCD head, leveraging the highest resolution features of bi-temporal images encoded by HRNet to enhance temporal dependency modeling, leading to mutual reinforcement with PRO-SS. Experimental results on a benchmark dataset demonstrate that our method surpasses state-of-the-art SCD approaches. Our source code is available at https://github.com/sdust-mmlab/PRO-HRSCD.

  • Open Access

    Semantically Robust Unsupervised Image Translation for Paired Remote Sensing Images

    by Sheng Fang, Kaiyu Li, Zhe Li, Jianli Zhao, Xingli Zhang

    Image translation for change detection or classification in bi-temporal remote sensing images is unique. Although it can acquire paired images, it is still unsupervised. Moreover, strict semantic preservation in translation is always needed instead of multimodal outputs. In response to these problems, this paper proposes a new method, SRUIT (Semantically Robust Unsupervised Image-to-image Translation), which ensures semantically robust translation and produces deterministic output. Inspired by previous works, the method explores the underlying characteristics of bi-temporal Remote Sensing images and designs the corresponding networks. Firstly, we assume that bi-temporal Remote Sensing images share the same latent space, for they are always acquired from the same land location. So SRUIT makes the generators share their high-level layers, and this constraint will compel two domain mapping to fall into the same latent space. Secondly, considering land covers of bi-temporal images could evolve into each other, SRUIT exploits the cross-cycle-consistent adversarial networks to translate from one to the other and recover them. Experimental results show that constraints of sharing weights and cross-cycle consistency enable translated images with both good perceptual image quality and semantic preservation for significant differences.