PaperlayerView exact revision on arXiv

Source-linked AI summary

CDMamba: Incorporating Local Clues into Mamba for Remote Sensing Image Binary Change Detection

Haotian Zhang, Keyan Chen, Chenyang Liu, Hao Chen, Zhengxia Zou, Zhenwei Shi

arXiv:2406.04207v2cs.CV

TL;DR

Remote-sensing change detection requires detailed local information, while existing approaches face computational or receptive-field limitations and Mamba-based methods often focus on global modeling. CDMamba combines Mamba and convolution through SRCM and uses AGLGF for guided bi-temporal fusion; it achieves the best F1 and IoU among compared Mamba-based methods and provides refined local detection.

  • Problem

    High-resolution remote-sensing change detection involves complex ground conditions, while Transformer-based processing can be computationally costly and existing Mamba-based methods often lack detailed local clues.

  • Method

    CDMamba combines global Mamba features with convolutional local clues through SRCM and dynamically guides bi-temporal feature interaction with AGLGF.

  • Results

    CDMamba achieves the best F1 score and IoU among the compared Mamba-based methods, although it is not optimal on some precision and recall measures.

  • Takeaways & Limitations

    By leveraging global and local modeling, CDMamba provides more refined local detection and stronger resistance to interference from irrelevant changes.

Abstract

from arXiv · show

Recently, the Mamba architecture based on state space models has demonstrated remarkable performance in a series of natural language processing tasks and has been rapidly applied to remote sensing change detection (CD) tasks. However, most methods enhance the global receptive field by directly modifying the scanning mode of Mamba, neglecting the crucial role that local information plays in dense prediction tasks (e.g., binary CD). In this article, we propose a model called CDMamba, which effectively combines global and local features for handling binary CD tasks. Specifically, the Scaled Residual ConvMamba (SRCM) block is proposed to utilize the ability of Mamba to extract global features and convolution to enhance the local details to alleviate the issue that current Mamba-based methods lack detailed clues and are difficult to achieve fine detection in dense prediction tasks. Furthermore, considering the characteristics of bi-temporal feature interaction required for CD, the Adaptive Global Local Guided Fusion (AGLGF) block is proposed to dynamically facilitate the bi-temporal interaction guided by other temporal global/local features. Our intuition is that more discriminative change features can be acquired with the guidance of other temporal features. Extensive experiments on five datasets demonstrate that our proposed CDMamba is comparable to the current methods (such as the F1/IoU scores are improved by 2.10%/3.00% and 2.44%/2.91% on LEVIR+CD and CLCD, respectively). Our code is open-sourced at https://github.com/zmoka-zht/CDMamba.

I. INTRODUCTION

Remote sensing change detection requires both broad contextual modeling and fine local detail, but existing approaches often emphasize one at the expense of the other. CDMamba addresses this gap by combining global/local feature extraction with guided bi-temporal fusion.

  • Remote sensing change detection supports urban planning, land-cover analysis, disaster assessment, ecosystem monitoring, and resource management.
  • Higher spatial resolution provides detailed texture and geometry but increases regional heterogeneity, limiting traditional empirically designed methods in complex scenes.
  • Transformers model global relationships but incur quadratic image-patch complexity, while Mamba offers data-dependent global modeling with linear complexity.
  • Most Mamba-based change-detection methods modify scanning to improve global receptive fields while neglecting local information crucial for accurate dense prediction.
  • CDMamba combines Mamba and convolution through SRCM to capture global and local clues, while AGLGF guides bi-temporal interaction using other-temporal global/local features.
  • Experiments on WHU-CD, LEVIR-CD, and LEVIR+CD report state-of-the-art results for CDMamba.

A. CNN-based CD Models

CNNs provide strong local feature extraction but struggle with long-range dependencies, whereas Transformers and Mamba address global modeling with different computational trade-offs. The paper positions Mamba as an efficient complement to convolution for change detection.

  • CNN-based CD Models: CNN-based change-detection models use fully convolutional or Siamese designs, dense connections, and deep supervision to learn bi-temporal or differential features.
  • CNN-based CD Models: CNNs’ local receptive fields make long-range dependency modeling difficult, especially when changed objects are sparse.
  • CNN-based CD Models: Transformer-based change-detection methods model long-range relationships but incur quadratic complexity with image-patch length, increasing computational costs for dense prediction.
  • CNN-based CD Models: State Space Models such as Mamba provide linear computational complexity for input sequence length while modeling long sequences and long-term dependencies.
  • CNN-based CD Models: Existing visual Mamba methods mainly modify scanning methods to enlarge global receptive fields, although local information remains crucial for accurate change detection.

B. Overview

CDMamba processes bi-temporal images through hierarchical SRCM encoders and decoders, with AGLGF enabling guided interaction and adaptive fusion of multi-scale features. Differential features are decoded back to the original image resolution for change prediction.

  • B. Overview: CDMamba contains SRCM encoder and decoder blocks together with an Adaptive Global Local Guided Fusion block.
  • B. Overview: Two bi-temporal images are processed by convolutional streams to produce shallow feature maps, then cascaded SRCM encoder blocks extract features at multiple scales.
  • B. Overview: AGLGF applies global/local guided fusion and adaptive gating to multi-scale deep features for bi-temporal interaction and semantic-context learning.
  • B. Overview: Multi-scale differential features are obtained by absolute subtraction after bi-temporal fusion.
  • B. Overview: The decoder uses SRCM, convolution, upsampling, and adjacent-scale fusion to restore feature maps to the original image size before linear projection.

C. Scaled Residual ConvMamba Block

The SRCM block combines Mamba-based global modeling with convolutional local extraction, then uses a scaled residual connection and projection to refine the fused representation. Its ConvMamba module realizes this combination through complementary branches.

  • C. Scaled Residual ConvMamba Block: SRCM combines convolutional local features with Mamba global features to improve accurate dense prediction.
  • C. Scaled Residual ConvMamba Block: SRCM applies LayerNorm and ConvMamba to an input feature, fuses the result with the input through a scaled residual connection, then normalizes and linearly transforms it.
  • C. Scaled Residual ConvMamba Block: ConvMamba splits the input channels into two branches that use linear layers, Conv1d, SSM, and LayerNorm before Hadamard-product fusion for global features.
  • C. Scaled Residual ConvMamba Block: A third ConvMamba branch reshapes the input, applies Conv2d and SiLU, and flattens the result to extract local features.
  • C. Scaled Residual ConvMamba Block: The local branch output is added to the global features and linearly mapped to produce a representation integrating both information types.

D. Adaptive Global Local Guided Fusion Block

AGLGF dynamically combines global- and local-guided fusion to improve bi-temporal interaction and produce more discriminative change features.

  • AGLGF combines global-guided and local-guided features to provide more discriminative bi-temporal change features.
  • Global-guided feature fusion: The global-guided fusion module processes F1 through two branches and uses F2 to provide globally scanned guidance.Its branches use linear mappings, SiLU, Conv1d, SSM, LayerNorm, and gated processing.
  • Local-guided feature fusion: The local-guided fusion module extracts local features from F2 with Conv2d and fuses them with F1 through gating.
  • Dynamic fusion: A dynamic gate weights the global- and local-guided outputs after channel compression and concatenation, producing the fused feature FGL.
  • Difference-feature generation: Swapping F1 and F2 provides the reciprocal guided feature, and absolute subtraction generates difference features.

A. Data description

The experiments use three remote-sensing change-detection datasets with distinct image sources, scales, temporal spans, and patch-based partitions.

  • Experiments evaluate CDMamba on WHU-CD, LEVIR-CD, and LEVIR+-CD datasets.
  • WHU-CD: WHU-CD contains a 32507×15354 image pair from New Zealand at 0.2m/pixel, acquired in April 2012 and April 2016.The images cover 20.5 square kilometers and are partitioned into 256 × 256 patches.
  • LEVIR-CD: LEVIR-CD contains 637 Google Earth image pairs at 0.5m/pixel, with a 5-to-14-year time span focused on building changes.The study partitions the images into non-overlapping 256 × 256 patches.
  • LEVIR+-CD: LEVIR+-CD extends LEVIR-CD with 985 image pairs covering varied building types, including residences, garages, and warehouses.The images are partitioned into 256 × 256 patches for training and testing.
  • Encoder configuration: The encoder uses four stages with feature resolutions matching the input, 1/2, 1/4, and 1/8 of the original size.The stage channel numbers are {16, 32, 64, 128}, and the layer counts are {1, 2, 2, 4}.

2) Training details:

Training uses Adam optimization with a combined cross-entropy and Dice loss, while performance is assessed using pixel-level classification and overlap metrics.

  • Training setup: CDMamba is trained for 300 epochs with Adam, an initial learning rate of 1e-4, and mini-batches of 6.The optimizer coefficients are β1=0.9 and β2=0.999.
  • Training setup: The loss combines cross-entropy loss and Dice loss.
  • Loss function: The loss equation uses λ1 and λ2 as loss coefficients, yi as the ith-pixel ground truth, ˆyi as its predicted probability, and N as the pixel count.
  • Evaluation metrics: Evaluation uses overall accuracy, precision, recall, F1 score, and intersection over union.F1 and IoU are noted as better reflecting model generalization ability.
  • Evaluation metrics: OA measures the proportion of correctly predicted pixels, while precision and recall measure positive-prediction correctness and positive-instance coverage.

C. Performance comparison

CDMamba is evaluated against CNN-, Transformer-, and Mamba-based methods using quantitative and qualitative comparisons across three change-detection test sets. It achieves strong overall performance, with detailed visual detections of edges and small changes.

  • Quantitative comparison: CDMamba is compared with CNN-, Transformer-, and Mamba-based methods under the same training conditions.The comparison covers WHU-CD, LEVIR-CD, and LEVIR+-CD test sets.
  • Quantitative comparison: CDMamba achieves superior overall performance across the WHU-CD, LEVIR-CD, and LEVIR+-CD test sets.Table I reports comparative results in percentage form, with the best three results highlighted by color.
  • Quantitative comparison: 0.97%/1.21%, 0.98%/0.59%, and 2.10%/2.24% are the F1-score improvements over Mamba-based methods on WHU-CD, LEVIR-CD, and LEVIR+-CD, respectively.The paired values correspond to the reported improvements on the three datasets.
  • Qualitative comparison: On WHU-CD visualizations, CDMamba produces more detailed change edges and detects small change areas that other methods miss in complex scenes.The visualizations use white for true positives, black for true negatives, red for false positives, and green for false negatives.
  • Qualitative comparison: On LEVIR-CD, CDMamba detects both large irregular buildings and small-scale building changes more effectively than CNN-, Transformer-, and Mamba-based methods.The authors associate improved sensitivity to small changes with integrated local feature extraction.
  • Qualitative comparison: On LEVIR+-CD, CDMamba achieves the best visual results across large- and small-change scenarios, with near-ground-truth detection in one small-change example.The cited example is Fig. 5(d), where competing methods almost completely miss the changes.

3) Model efficiency:

CDMamba improves change-detection performance through global-local feature integration and adaptive fusion, while remaining more efficient than a Transformer baseline. Ablations indicate that SRCM and early AGLGF placement are particularly important, whereas activation and dimensionality choices affect performance.

  • Model efficiency: CDMamba outperforms Changeformer in both parameter count and one-epoch training time, while remaining lightweight relative to RS-Mamba and ChangeMamba.Its training time is slightly higher than those Mamba-based methods.
  • Effects of Different Components in CDMamba: CDMamba outperforms the baseline by 6.45% in F1 and 10.78% in IoU when its key modules are added.The ablation results attribute the improvement to integrating global and local features and adaptively fusing differential features.
  • Effects of Different Components in CDMamba: SRCM produces clearer structures and edges than the baseline in complex scenes and scenes containing many added buildings.The qualitative comparison supports the role of global-local information integration in dense prediction.
  • Effects of different stages of AGLGF: AGLGF performs best when added at the second stage; adding it at deeper stages reduces performance because deep semantic features provide less detailed guidance.Shallower-stage fusion also focuses attention more strongly on changed regions in the differential feature visualizations.
  • Gate activation: Non-saturating gate activations generally outperform saturating functions, leading the model to select ReLU for its final gate activation.The paper attributes this preference to better preservation of detailed input features.

4) Effects of Different Dimensions:

The dimensionality study finds that larger L-GF convolutional feature dimensions improve performance, with diminishing gains at the largest tested expansion. CDMamba combines this design with global-local extraction and adaptive bi-temporal fusion for change detection.

  • Effects of Different Dimensions: Expanding L-GF convolutional feature dimensions progressively improves performance, although gains slow from 1.5×d-model to 2×d-model.The final design uses 2×d-model because it achieves the best performance.
  • Loss-function coefficients: The loss coefficients λ1 and λ2 achieve the best performance when relatively balanced, leading to the selected setting λ1=λ2=0.5.λ1 weights cross-entropy loss and λ2 weights dice loss.
  • Conclusion: CDMamba combines Mamba-based global feature extraction with convolution-based local-clue extraction to address insufficient detail in dense change detection.Its AGLGF block guides bi-temporal feature interaction using global and local features.
Loading 2406.04207v2…