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

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.

• The paper presents CD-Lamba, a novel remote sensing change detection method that integrates cross-temporal fusion with locally adaptive scanning to capture both regional details and global context.
• It introduces key strategies—LASS, CTSS, and a window shifting mechanism—that improve feature extraction, yielding state-of-the-art F1-score gains across multiple benchmarks.
• The approach refines traditional state space models, offering practical gains for applications such as urban planning and ecological conservation through more accurate spatio-temporal analysis.

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

The paper presents a novel approach to address inherent limitations in remote sensing change detection (RSCD) methods leveraging state space models (SSMs). While existing RSCD methods based on SSMs, such as Mamba, effectively facilitate global awareness with linear computational complexity, they fail to adequately capture the locality within change regions. In this context, the proposed CD-Lamba introduces significant advancements by integrating a Cross-Temporal Locally Adaptive State Space Model that simultaneously enhances locality and maintains global perception.

Contributions and Methodology

The primary contribution of this work is the development of CD-Lamba, which introduces several key strategies aimed at refining feature extraction in bi-temporal remote sensing images:

1. Locally Adaptive State-Space Scan (LASS) Strategy: Unlike previous SSM-based RSCD methods which flatten images for sequential scanning, the LASS strategy maintains spatial coherence by dynamically selecting and scanning significant local windows. This ensures that the model captures both the local details and global context necessary for effective RSCD.
2. Cross-Temporal State-Space Scan (CTSS) Strategy: The CTSS strategy enhances feature fusion across bi-temporal inputs by utilizing pixel-wise cross-scanning, thereby improving the alignment and interaction of temporal information in the feature space.
3. Window Shifting and Perception (WSP) Mechanism: Inspired by Swin Transformer’s shifted window concept, the WSP mechanism ensures seamless connectivity between scanned segments. This approach ameliorates discontinuity issues at feature map boundaries and maximizes spatial interactions.

The integration of these strategies into the multi-scale Cross-Temporal Locally Adaptive State-Space Scan (CT-LASS) module facilitates superior modulation of feature generation in bi-temporal inputs, significantly elevating CD-Lamba's RSCD capabilities.

Numerical Results and Performance

CD-Lamba was rigorously evaluated across four challenging RSCD benchmark datasets: WHU-CD, SYSU-CD, DSIFN-CD, and CLCD. The experimental results demonstrate that CD-Lamba achieves state-of-the-art performance, presenting improvements in the F1-score by 2.43%, 3.28%, 5.72%, and 8.06% on the respective datasets, compared to prior SSM-based methods like ChangeMamba. This robust improvement underscores both the accuracy and computational efficiency of CD-Lamba, thus verifying the efficacy of the proposed strategies in enhancing local-global spatio-temporal interactions for RSCD tasks.

Implications and Future Directions

The implications of this research extend to both practical and theoretical aspects of remote sensing and computer vision. Practically, CD-Lamba can facilitate more accurate monitoring and analysis of geographical and environmental changes, crucial for applications like urban development assessment and ecological conservation. Theoretically, CD-Lamba sets the stage for further integration of adaptive learning strategies in SSMs, potentially embedding self-regulating mechanisms for autonomous window learning and context-aware feature extraction.

In conclusion, CD-Lamba marks a substantial advancement in RSCD, offering a sophisticated blend of local enhancement and global awareness that could inspire future explorations in adaptive state space modeling within AI and remote sensing disciplines.