Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CD-Lamba: Boosting Remote Sensing Change Detection via a Cross-Temporal Locally Adaptive State Space Model

Zhenkai Wu, Xiaowen Ma, Rongrong Lian, Kai Zheng, Mengting Ma, Wei Zhang, Siyang Song

TL;DR

CD-Lamba tackles RSCD by addressing locality loss inherent in prior SSM-based approaches that flatten images for sequential processing. It introduces a multi-scale Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module with a Locally Adaptive Scan (LASS), Cross-Temporal Scan (CTSS), and Window Shifting and Perception (WSP) to tightly couple local details with global context, aided by a Siamese backbone and a lightweight detector. The method achieves state-of-the-art F1 on four RSCD benchmarks with strong efficiency (28.74M parameters, 15.26G FLOPs) and shows clear qualitative improvements in boundary sharpness and reduction of false alarms.abl Ablation studies confirm the benefits of adaptive locality, cross-temporal fusion, and multi-scale integration, while the authors acknowledge remaining challenges in separating true locality from pseudo locality and propose future directions for autonomous window learning through state transfer and knowledge distillation.

Abstract

Mamba, with its advantages of global perception and linear complexity, has been widely applied to identify changes of the target regions within the remote sensing (RS) images captured under complex scenarios and varied conditions. However, existing remote sensing change detection (RSCD) approaches based on Mamba frequently struggle to effectively perceive the inherent locality of change regions as they direct flatten and scan RS images (i.e., the features of the same region of changes are not distributed continuously within the sequence but are mixed with features from other regions throughout the sequence). In this paper, we propose a novel locally adaptive SSM-based approach, termed CD-Lamba, which effectively enhances the locality of change detection while maintaining global perception. Specifically, our CD-Lamba includes a Locally Adaptive State-Space Scan (LASS) strategy for locality enhancement, a Cross-Temporal State-Space Scan (CTSS) strategy for bi-temporal feature fusion, and a Window Shifting and Perception (WSP) mechanism to enhance interactions across segmented windows. These strategies are integrated into a multi-scale Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module to effectively highlight changes and refine changes' representations feature generation. CD-Lamba significantly enhances local-global spatio-temporal interactions in bi-temporal images, offering improved performance in RSCD tasks. Extensive experimental results show that CD-Lamba achieves state-of-the-art performance on four benchmark datasets with a satisfactory efficiency-accuracy trade-off. Our code is publicly available at https://github.com/xwmaxwma/rschange.

CD-Lamba: Boosting Remote Sensing Change Detection via a Cross-Temporal Locally Adaptive State Space Model

TL;DR

CD-Lamba tackles RSCD by addressing locality loss inherent in prior SSM-based approaches that flatten images for sequential processing. It introduces a multi-scale Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module with a Locally Adaptive Scan (LASS), Cross-Temporal Scan (CTSS), and Window Shifting and Perception (WSP) to tightly couple local details with global context, aided by a Siamese backbone and a lightweight detector. The method achieves state-of-the-art F1 on four RSCD benchmarks with strong efficiency (28.74M parameters, 15.26G FLOPs) and shows clear qualitative improvements in boundary sharpness and reduction of false alarms.abl Ablation studies confirm the benefits of adaptive locality, cross-temporal fusion, and multi-scale integration, while the authors acknowledge remaining challenges in separating true locality from pseudo locality and propose future directions for autonomous window learning through state transfer and knowledge distillation.

Abstract

Mamba, with its advantages of global perception and linear complexity, has been widely applied to identify changes of the target regions within the remote sensing (RS) images captured under complex scenarios and varied conditions. However, existing remote sensing change detection (RSCD) approaches based on Mamba frequently struggle to effectively perceive the inherent locality of change regions as they direct flatten and scan RS images (i.e., the features of the same region of changes are not distributed continuously within the sequence but are mixed with features from other regions throughout the sequence). In this paper, we propose a novel locally adaptive SSM-based approach, termed CD-Lamba, which effectively enhances the locality of change detection while maintaining global perception. Specifically, our CD-Lamba includes a Locally Adaptive State-Space Scan (LASS) strategy for locality enhancement, a Cross-Temporal State-Space Scan (CTSS) strategy for bi-temporal feature fusion, and a Window Shifting and Perception (WSP) mechanism to enhance interactions across segmented windows. These strategies are integrated into a multi-scale Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module to effectively highlight changes and refine changes' representations feature generation. CD-Lamba significantly enhances local-global spatio-temporal interactions in bi-temporal images, offering improved performance in RSCD tasks. Extensive experimental results show that CD-Lamba achieves state-of-the-art performance on four benchmark datasets with a satisfactory efficiency-accuracy trade-off. Our code is publicly available at https://github.com/xwmaxwma/rschange.
Paper Structure (34 sections, 30 equations, 14 figures, 7 tables)

This paper contains 34 sections, 30 equations, 14 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. State Space Model
  4. Remote Sensing Change Detection
  5. Preliminaries and Findings
  6. Preliminaries
  7. State Space Models (S4)
  8. Discretization
  9. Selective State Space Models (S6)
  10. Findings
  11. Methodology
  12. Overall Architecture
  13. Siamese Backbone
  14. Proposed State-Space Scan Strategies
  15. Locally Adaptive State-Space Scan Strategy
  16. ...and 19 more sections

Figures (14)

  • Figure 1: Comparisons of State-Space Scan (SS) strategies among VMamba liu2024vmamba, LocalMamba huang2024localmamba and CD-Lamba. (a) An example of a pair of geographically co-registered remote sensing images along with the groundtruth of the change regions. (b) The SS strategy in VMamba leads to locality loss, reducing the model's ability to capture local details. (c) The SS strategy in LocalMamba is constrained by fixed windows, limiting its adaptability. (d) The step-by-step process of our proposed Locally Adaptive State-Space Scan (LASS) strategy. Pink windows are selected based on the top-$k$ score map. The sequence is formed by first flattening all yellow pixels and then sequentially appending the flattened connected regions from the pink windows. In terms of reducing the loss of locality, for example, when comparing LocalMamba with VMamba, the gap between the two rows of features in Change 2 has been reduced from the original 6 pixels to just 2 pixels. Furthermore, when comparing our CD-Mamba with LocalMamba, this gap is reduced to 0 pixels.
  • Figure 2: Spectral analysis between SS2D in VMamba liu2024vmamba and our proposed LASS in CD-Lamba. Low-frequency global features are closer to the center, while high-frequency local features are farther from the center. The brighter the pixel, the greater the energy. Therefore, the figure demonstrates that LASS not only (a) enhances the locality recognized by SS2D but also (b) activates the locality that SS2D cannot recognize.
  • Figure 3: More spectral analysis of typical SSM-based RSCD methods (e.g., RSMamba rsmamba and ChangeMamba changemamba). The comparison focuses on replacing their selective scan strategies with either the SS2D approach in VMamba liu2024vmamba or our proposed LASS in CD-Lamba. Low-frequency global features are closer to the center, while high-frequency local features are farther from the center. The brighter the pixel, the greater the energy. Therefore, the figure demonstrates that LASS not only enhances the locality recognized by SS2D but also activates the locality that SS2D cannot recognize.
  • Figure 4: (a) The overall architecture of CD-Lambda consists of a Siamese backbone, (b) a Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module, and (c) a lightweight change detector (LCD). In the CT-LASS module, we propose a novel Locally Adaptive State-Space Scan (LASS) strategy, where the score window is generated by average pooling the differences of bi-temporal features. The CT-LASS module efficiently models the global-local spatio-temporal context, enabling the modulation of feature generation based on bi-temporal feature flows and selectively enhancing changes of interest. (d) Additionally, we introduce a Window Shifting and Perception (WSP) mechanism, performing CT-LASS five times at each scale with shifts of 1/8 length in four directions, ensuring sufficient interaction between bi-temporal features.
  • Figure 5: Illustration of Cross-Temporal State-Space Scan (CTSS) strategy with bi-temporal selective scan strategy. We first unfold the bi-temporal windows into one-dimensional sequences along each of the four directions. For the sequence of bi-temporal features obtained in each direction, we perform cross-scanning for bi-temporal features at the same location. Finally, the output features from the four directions are merged to construct the final feature windows.
  • ...and 9 more figures