TL;DR
CDMamba is a new architecture for remote sensing change detection that combines the global modeling power of Mamba (State Space Models) with the local precision of convolutions. By integrating a novel Scaled Residual ConvMamba (SRCM) block and a cross-temporal Adaptive Global Local Guided Fusion (AGLGF) module, it achieves state-of-the-art accuracy on major benchmarks while maintaining the computational efficiency needed for high-resolution satellite imagery.
What Problem Does This Paper Address?
Remote sensing Change Detection (CD)—identifying differences between images of the same location taken at different times—is a "dense prediction" task that requires two conflicting capabilities. First, the model needs a global receptive field to understand the context of sparse changes across large landscapes. Second, it needs fine-grained local clues to accurately delineate the boundaries of buildings, roads, or land cover.
Traditional CNNs excel at local details but struggle with long-range dependencies. Vision Transformers (ViTs) solve the global modeling problem but suffer from quadratic computational complexity, making them prohibitively expensive for high-resolution dense prediction. Recently, the Mamba architecture (based on State Space Models) emerged as a linear-complexity alternative to Transformers. However, the authors argue that current Mamba-based methods focus too much on global scanning and neglect the local spatial information essential for precise change detection.
Methods
The authors propose CDMamba, an encoder-decoder network designed to balance global and local feature extraction.
1. Scaled Residual ConvMamba (SRCM) Block
Instead of relying solely on Mamba's 1D scanning, the SRCM block introduces a three-branch "ConvMamba" module. Two branches perform standard Mamba-style global modeling, while a third branch uses 2D convolutions to specifically capture local spatial features. These are fused via a scaled residual connection, ensuring the model doesn't lose track of fine edges while processing global context.
2. Adaptive Global Local Guided Fusion (AGLGF) Block
Change detection requires comparing two images ($T_1$ and $T_2$). Standard methods often just subtract or concatenate features. CDMamba introduces AGLGF, which uses one temporal feature to "guide" the other.
- Global-Guided Fusion (G-GF): Uses global context from one time-step to highlight potential change regions in the other.
- Local-Guided Fusion (L-GF): Uses local spatial clues for the same purpose. A dynamic gating mechanism then fuses these guided features, allowing the model to adaptively focus on the most discriminative changes.
Datasets
The model was evaluated on three representative building change detection benchmarks:
- WHU-CD: High-resolution (0.2m) aerial imagery from New Zealand.
- LEVIR-CD: A large-scale dataset of Google Earth images (0.5m resolution) focusing on building additions and removals.
- LEVIR+-CD: An extension of LEVIR-CD featuring more diverse building types and complex urban scenes.
Key Insights and Ideas
- Mamba needs a "Local Anchor": The researchers found that while Mamba is great at long-range sequences, it can be "blurry" at the pixel level. Adding a dedicated convolutional path within the Mamba block acts as a spatial anchor, significantly sharpening detection results.
- Guidance is better than Subtraction: Simply subtracting features from two time-steps can lead to noise. By using an attention-like "guidance" mechanism (AGLGF), the model learns to look for changes in one image based on the context provided by the other.
- Efficiency without Sacrificing Resolution: Unlike some Mamba models that downsample images heavily to save time, CDMamba processes original-sized images more effectively, leading to better small-object detection.
Key Results
CDMamba consistently outperformed CNN-based (e.g., SNUNet), Transformer-based (e.g., Changeformer, BIT), and recent Mamba-based (e.g., RS-Mamba) competitors.
- Accuracy: On the WHU-CD dataset, CDMamba achieved an F1 score of 93.76% and an IoU of 88.26%, representing a significant jump over previous Mamba-based SOTA methods.
- Visual Quality: Qualitative results show that CDMamba produces much cleaner building edges and fewer false positives in complex scenes compared to models like BIT or Changeformer.
- Efficiency: Despite its high accuracy, CDMamba remains lightweight, with only 11.90M parameters, compared to 41.02M for Changeformer and 51.95M for RS-Mamba.
Limitations and Future Work
While CDMamba is efficient in terms of parameters, the authors note that its training time per epoch is slightly higher than some competitors. This is because CDMamba avoids aggressive initial downsampling to preserve detail, which increases the computational load during the early stages of the network.
For future work, the authors suggest exploring self-supervised learning to pre-train the Mamba backbone, which could further improve performance in scenarios where labeled change detection data is scarce.
Summary updated