The loss of high-frequency details and the insufficient fusion of global and local features remain key challenges in bi-temporal remote sensing image change captioning. To address these issues, a change captioning method based on a frequency-spatial dual-domain mamba network was proposed. A frequency-spatial dual-domain perception module was first constructed, through which global and local features are efficiently integrated via a collaborative mechanism operating across the frequency and spatial domains. Subsequently, a fourier convolution enhancement module was developed, in which fast Fourier convolution is employed to dynamically select critical frequency components, thereby improving the representation of high-frequency details. By concatenating frequency-and spatial-domain features, both high-frequency detail preservation and enhanced global perception were achieved. Finally, a spatial-adaptive gating module was introduced to regulate convolutional kernel weight distributions through spatially adaptive coordinate transformations. Connected in parallel with the Fourier enhancement branch, this module further strengthens the dynamic fusion between local features and global contextual information. Experimental results on the LEVIR-CC dataset show that BLEU-4 and CIDEr-D scores of 66.06% and 139.08% are achieved, demonstrating clear advantages in accuracy and fine-grained description quality for remote sensing image change captioning.
LIU Zhong-min, ZHOU Yun-long, HU Wen-jin.
Remote Sensing Image Change Captioning Based on Frequency-spatial Dual-domain Mamba Network[J].
Science Technology and Engineering, 2026, 26(18): 7809-7821 DOI:10.12404/j.issn.1671-1815.2504468
状态空间模型(state space model, SSM)起源于控制系统[18],用于描述动态系统中状态随时间的演变关系[19]。近年来,SSM被逐步引入深度学习领域,作为一种高效的序列建模方法,其通过状态变量与线性或非线性变换能够在理论上以线性复杂度捕捉长程依赖[20]。Gu等[21]提出了Mamba架构,该架构通过选择性状态空间机制使模型参数能够根据输入自适应调整,从而增强上下文建模能力,同时保持较低的计算复杂度。凭借这一特性,Mamba架构在自然语言处理等序列建模任务中受到广泛关注,并开始被探索应用于计算机视觉任务[22-23]。中国研究者已将Mamba架构引入遥感领域,陈辉等[24]提出一种结合注意力机制与Mamba架构的遥感图像分类模型,通过引入空间分组坐标注意力模块和双流多向状态空间编码器,增强了模型对遥感图像空间信息的感知能力,提升了分类性能并降低了计算复杂度。马文浩等[25]将Mamba架构拓展至遥感图像配准任务,提出双向多尺度Mamba网络。该方法通过多尺度选择性扫描捕获多级特征,结合跨空间交互模块建模图像间依赖,并采用互相关双向匹配提升几何一致性,有效缓解了多尺度信息利用不足、配准精度受限及计算效率低下的问题。针对遥感变化检测任务,杨瑞等[26]提出了一种基于特征交互和差分对齐的遥感影像建筑物变化检测方法。该方法采用Mamba状态空间模型构建编码器-解码器架构,通过特征交互模块增强双时相特征的融合能力,并利用差分对齐模块校正图像间的空间偏移,提升了在复杂场景下的变化检测精度。针对遥感影像变化检测中多尺度信息利用不足、微小变化难以捕捉等问题,李雪等[27]提出了基于变化引导的双向Mamba网络。该模型在编码阶段通过时域融合模块实现跨时间与跨尺度特征融合,增强双时相变化感知;在解码阶段引入双向Mamba模块,通过前向与反向建模捕获长程依赖,促进局部与全局特征协同建模;同时结合深度变化引导特征,以先验信息引导多尺度特征整合,从而保留关键变化信息。为进一步挖掘状态空间模型在遥感变化检测中的潜力并缓解其高计算开销,厉芳婷等[28]提出了双流并行全向扫描Mamba网络,通过深度分析参数量关键因素,设计双流并行结构,在降低模型参数量与计算复杂度的同时,保持优异的检测性能。为缓解伪标签质量不足与特征利用单一的问题,张宇等[29]提出了一种基于Mamba的空间-频率双域特征融合U-Net模型,借助Mamba模块捕获全局上下文并通过小波卷积增强频率域表征,从而缓解了双域信息互补不足的问题。
在RSICC任务中,生成准确且语义一致的文本不仅依赖于对显著变化区域的细粒度局部建模,还要求模型具备对跨时相、跨空间尺度的全局语义理解能力。尽管当前主流方法采用基于Transformer的架构进行全局依赖建模,但其标准自注意力机制对地理空间连续性的建模存在固有偏差,难以充分表达复杂的变化语义。为了解决这一问题,基于状态空间模型设计了一种视觉状态空间模块(vision state space module, VSSM),结构如图4所示。该模块通过结构化状态转移机制,增强模型对跨区域长程依赖的建模能力,进一步提升全局感知能力。
(3)METEOR(metric for evaluation of translation with explicit ORdering)[36]通过计算生成描述与参考描述之间的最长公共子序列(longest common subsequence, LCS)长度,来衡量它们在序列结构上的相似性。该指标通过强化连续匹配片段优化文本连贯性,并利用调和平均机制实现精确率与召回率的动态平衡,从而更全面地反映生成内容与参考内容的对齐程度。
ZhangTao, WangGuanghui, ZhengLijuan, et al. Remote sensing monitoring and spatiotemporal change characteristics of urbanconstruction land expansion in China’s urban agglomerations over the past 30 years[J]. Journal of Remote Sensing, 2024, 28(11): 2838-2849.
YuanYixin, WenQingke, XuJinyong, et al. Remote sensing monitoring of dynamic changes in mangrove forests in the Guangdong-Hong Kong-Macao Greater Bay Area from 1990 to 2020[J]. Journal of Remote Sensing, 2023, 27(6): 1496-1510.
YanHongbo, XuPengjun, LuXianjian. Response of vegetation to short-term drought in the Karst region based on GF6 remote sensing imagery[J]. Science Technology and Engineering, 2024, 24(27): 11576-11585.
LiuBo, NiuQuanfu, WangGang, et al. Analysis of forest cover changes and driving forces in the Loess Plateau (Gansu region) based onmultisensor remote sensing images[J]. Science Technology and Engineering, 2025, 25(1): 54-66.
[9]
YangY, LiuT, PuY, et al. Remote sensing image change captioning using multi-attentive networkwith diffusion model[J]. Remote Sensing, 2024, 16(21): 4083.
GaoChongqin, GuYanchao, ShenFuqiang, et al. Remote sensing image change detection based on improved deep convolutional gene-ration adversarial network[J]. Science Technology and Engineering, 2025, 25(23): 9911-9918.
SongJingkuan, ZengPengpeng, GuJiayang, et al. End-to-end image description generation based on visual region aggregation and bidirectional collaboration[J]. Journal of Software, 2023, 34(5): 2152-2169.
JinJiali, YuLu. Continual image captioning with dynamic Token-used fusion feature[J]. Computer Engineering and Applications, 2025, 61(4): 176-191.
[18]
HoxhaG, ChouafS, MelganiF, et al. Change captioning: a new paradigm for multitemporal remote sensing image analysis[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-14.
[19]
LiuC, ZhaoR, ChenH, et al. Remote sensing image change captioning with dual-branch transformers: a new method and a large scale dataset[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-20.
[20]
ChangS, GhamisiP. Changes to captions: an attentive network for remote sensing changecaptioning[J]. IEEE Transactions on Image Processing, 2023, 32: 6047-6060.
[21]
LiuC, YangJ, QiZ, et al. Progressive scale-aware network for remote sensing image change captioning[C]// IEEE International Geoscience and Remote Sensing Symposium (IGARSS). Pasadena: IEEE, 2023: 6668-6671.
[22]
YangC, LiZ, JiaoH, et al. Enhancing perception of key changes in remote sensing image change captioning[J]. arxiv Preprint arxiv: 2409.12612, 2024.
[23]
YuX, LiY, MaJ, et al. Diffusion-rscc: diffusion probabilistic model for change captioning in remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 1-13.
[24]
SunD, BaoY, LiuJ, et al. A lightweight sparse focus transfor-mer for remote sensing image change captioning[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024, 17: 18727-18738.
[25]
ZhouQ, GaoJ, YuanY, et al. Single-stream extractor network with contrastive pre-training for remote sensing change captioning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 1-14.
[26]
SmithJ T H, WarringtonA, LindermanS W. Simplified state space layers for sequence modeling[J]. arXiv Preprint arXiv: 2208.04933, 2022.
YueRentian, NiuMeng, et al. Conflict detection and resolution strategy for low-altitude multiple UAVs based on flight conflict network[J]. Science Technology and Engineering, 2025, 25(22): 9631-9639.
[29]
GuA, GoelK, RéC. Efficiently modeling long sequences with structured state spaces[J]. arXiv Preprint arXiv: 2111.00396, 2021.
[30]
GuA, DaoT. Mamba: linear-time sequence modeling with selective state spaces[J]. arXiv Preprint arXiv: 2312.00752, 2023.
[31]
DongW, ZhuH, LinS, et al. Fusion-mamba for cross-modality object detection[J]. arXiv Preprint arXiv: 2404.09146, 2024.
[32]
LiuY, TianY, ZhaoY, et al. Vmamba: visual state space model[J]. Advances in Neural Information Processing Systems, 2024, 37: 103031- 103063.
YangRui, LiuYu. Remote sensing image building change detection method based on feature interaction and differential alignment[J]. Journal of Geo-information Science, 2025, 27(11): 2731-2740.
LiFangting, ZhangGuo, ZhouFangdong, et al. Building change detection based on dual-stream parallel omnidirectional scan Mamba network[J]. Geomatics and Information Science of Wuhan University, 2024, 49(9): 1574-1585.
ZhangYu, ZhuangHuifu, ZhangXiang, et al. MSF-UNet: an unsupervised change detection method for SAR images via spatial-frequency dual-domain feature fusion[J]. Journal of Geo-information Science, 2025, 27(9): 2213-2229.
[45]
MaY, JiJ, SunX, et al. Image captioning via dynamic path customization[J]. IEEE Transactions on Neural Networks and Learning Systems, 2025, 36(4): 6203-6217.
WuGuangli, GuoZhenzhou, LiLeiting. Combining bottom-up and top-down attention for image captioning[J]. Science Techno-logy and Engineering, 2022, 22(32): 14313-14320.
[50]
LuX, WangB, ZhengX, et al. Exploring models and data for remote sensing image caption generation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 56(4): 2183-2195.
[51]
PapineniK, RoukosS, WardT, et al. Bleu: a method for automatic evaluation of machine translation[C]// Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. Philadelphia: Association for Computational Linguistics, 2002: 311-318.
[52]
LinC Y. Rouge: a package for automatic evaluation of summaries[C]// Text Summarization Branches Out. Barcelona: Association for Computational Linguistics, 2004: 74-81.
[53]
BanerjeeS, LavieA. METEOR: an automatic metric for MT eva-luation with improved correlation withhuman judgments[C]// Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. Ann Arbor: Association for Computational Linguistics, 2005: 65-72.
[54]
VedantamR, Lawrence ZitnickC, ParikhD. CIDEr-D: consensus-based image description evaluation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Los Alamitos: IEEE Computer Society, 2015: 4566-4575.
[55]
KingmaD P. Adam: a method for stochastic optimization[J]. arXiv Preprint arXiv: 1412.6980, 2014.
[56]
ParkD H, DarrellT, RohrbachA. Robust change captioning[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Los Alamitos: IEEE Computer Society, 2019: 4624-4633.
[57]
QiuY, YamamotoS, NakashimaK, et al. Describing and localizing multiple changes with transformers[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Los Alamitos: IEEE Computer Society, 2021: 1951-1960.
[58]
XiaoC, LiM, ZhangZ, et al. Spatial-mamba: effective visual state space models via structure-aware state fusion[J]. arXiv Preprint arXiv: 2410.15091, 2024.
[59]
DaiT, WangJ, GuoH, et al. FreqFormer: frequency-aware transformer for lightweight image super-resolution[C]// Proceedings of the 33rd International Joint Conference on Artificial Intelligence. California: IJCAI Organization, 2024: 731-739.
[60]
WuW, ZhengC, YangZ, et al. Frequency guidance matters: skeletal action recognition by frequency-aware mixed transformer[C]// Proceedings of the 32nd ACM International Conference on Multimedia. New York: Association for Computing Machinery, 2024: 4660-4669.