Answers are grounded in this paper and cite their sources.
Deep-learning-based change detection (CD) methods have become an important means of synthetic aperture radar (SAR) images to identify changes. To generate the accurate change map, these methods typically require a high-quality training set. As a frequently adopted way to extract training samples, preclassification has a crucial effect on CD precision. However, preclassification images are often generated using intensity-based CD algorithms that rely on SAR magnitude images without considering phase information. In addition, the statistical characteristics of SAR images are seldom considered. When designing the artificial intelligence CD models, it is expected to account for speckle noise inherent in SAR images while detecting more small-scale changes to obtain a high-accuracy CD map. Thus, we introduce a new CD approach for single-polarization SAR images based on image standardization and the improved residual network (I-ResNet). First, a strategy of fusing the coherent and noncoherent intensity changes for preclassification image generation is introduced to retain large-scale and small-scale changes. The noncoherent change acquisition part of the strategy involves an image standardization algorithm, which is derived from the Gaussian speckle model and is especially effective for images with different statistical characteristics. Then, the I-ResNet model combining the dual-tree complex wavelet transform with residual learning is presented, which aims at taking advantage of the wavelet transform in reducing the influence of speckle noise and ResNet in easy training and preservation of information integrity. Finally, experiments with different SAR image pairs demonstrate that the proposed method produces better CD maps than other related methods.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: