Wavelet-Based Bi-Dimensional Aggregation Network for SAR Image Change Detection

Jiangwei Xie, Feng Gao, Xiaowei Zhou, Junyu Dong

Abstract

Synthetic Aperture Radar (SAR) image change detection is a critical task in remote sensing, yet it remains challenging due to inherent speckle noise and the loss of fine-grained information during feature extraction. This paper introduces the Wavelet-based Bi-dimensional Aggregation Network (WBANet), a novel architecture designed to preserve high-frequency details. By replacing traditional lossy downsampling with Discrete Wavelet Transforms (DWT) in the attention mechanism and employing a bi-dimensional aggregation module, WBANet achieves state-of-the-art accuracy across diverse SAR datasets.

Introduction

SAR imaging is indispensable for Earth observation due to its all-weather and day-night capabilities. However, detecting changes between multi-temporal SAR images is complicated by speckle noise—a multiplicative noise that degrades image quality. Modern deep learning approaches, particularly those based on Transformers, often exacerbate information loss by using pooling operations to reduce computational complexity. These operations act as low-pass filters, discarding high-frequency textures essential for precise change detection.

Methodology

The proposed WBANet architecture addresses these limitations through two primary innovations: the Wavelet-based Self-attention Module (WSM) and the Bi-dimensional Aggregation Module (BAM).

Wavelet-based Self-attention Module (WSM)

In standard self-attention blocks, downsampling is typically achieved through pooling or strided convolutions, which permanently delete structural information. WBANet replaces these with the Discrete Wavelet Transform (DWT). DWT decomposes the input into four subbands: one low-frequency component (LL) and three high-frequency components (LH, HL, HH). Because DWT is mathematically invertible via the Inverse Discrete Wavelet Transform (IDWT), this process is essentially lossless.

Comparison of traditional self-attention and the proposed Wavelet-based Self-attention Module (WSM).

As shown in the figure above, the WSM utilizes DWT to reduce the spatial resolution of the "Key" and "Value" matrices while retaining the high-frequency details in the decomposed subbands. This allows the attention mechanism to capture long-range dependencies without sacrificing the fine textures necessary for accurate boundary detection.

Bi-dimensional Aggregation Module (BAM)

To further enhance feature representation, WBANet incorporates the BAM, which processes features through two parallel branches. The Channel Aggregation branch captures global context by modeling inter-channel relationships, while the Spatial Aggregation branch focuses on local spatial dependencies. These branches are merged using a broadcast mechanism, providing a more robust non-linear mapping than standard linear layers.

Experimental Results

The performance of WBANet was evaluated on three benchmark datasets: Chao Lake (Sentinel-1), Sulzberger (Envisat), and Yellow River (Radarsat-2). These datasets represent various challenges, including flooding, ice shelf breakup, and high speckle noise environments.

Quantitative Performance

WBANet consistently outperformed five state-of-the-art methods, including CWNN and CAMixer. The model achieved a Percentage of Correct Classification (PCC) of 98.33% on the Chao Lake dataset, 96.65% on Sulzberger, and 96.62% on the Yellow River dataset. Notably, the Kappa Coefficient (KC) showed improvements of up to 24% over baseline models on the Chao Lake dataset, indicating a high degree of reliability beyond mere chance agreement.

Qualitative Analysis

Visual inspection of the generated change maps confirms the quantitative findings. WBANet produces maps with significantly smoother boundaries and fewer false positives caused by speckle noise.

Qualitative comparison of change detection results across three datasets.

In the Yellow River dataset, which is particularly noisy, WBANet successfully identifies changed areas while ignoring the heavy speckle that leads to "salt-and-pepper" artifacts in other models. The researchers also noted a "Goldilocks" sweet spot for model depth, finding that performance peaked at five aggregation blocks before complexity led to diminishing returns.

Conclusion

WBANet represents a significant advancement in SAR change detection by integrating wavelet theory with modern attention mechanisms. By treating downsampling as a lossless decomposition rather than a reduction, the network preserves the high-frequency details critical for identifying subtle environmental changes. Future work will focus on automating hyperparameter selection to adapt the network's depth to different sensor types and terrains.

Summary updated

BibTeX
@article{xie2024wavelet,
  author = {Jiangwei Xie and Feng Gao and Xiaowei Zhou and Junyu Dong},
  title = {{Wavelet-Based Bi-Dimensional Aggregation Network for SAR Image Change Detection}},
  journal = {IEEE Geosci. Remote. Sens. Lett.},
  year = {2024},
  doi = {10.1109/lgrs.2024.3431532},
}
Figure