.png){kind=icon}
.png){kind=icon}
Return
Adaptive frequency collaboration for remote sensing change detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108234.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning methods have recently begun exploiting frequency information to improve remote sensing change detection. However, they typically aggregate high- and low-frequency components for detection without explicitly distinguishing their respective roles, leading to a decline in performance. In particular, high-frequency components capture fine spatial details associated with object heterogeneity, potentially introducing spurious differences and interfering with accurate change detection. In contrast, low-frequency components maintain stable energy distributions and better preserve the global structure of land cover objects, thus benefiting the localization of actual changed objects. To overcome this issue, we propose an adaptive frequency collaboration network (AFCN) to construct change features from a frequency-domain perspective. To achieve frequency disentanglement, we design a position-specific low-pass filter that adaptively extracts the low-frequency component from the spatial feature. Inspired by the wavelet reconstruction principle, the high-frequency counterpart is obtained by subtracting the low-frequency part from the spatial feature. The low-frequency part is used to generate change features for locating changed objects. Meanwhile, the high-frequency part is employed to extract edge features that enhance spatial details through an auxiliary edge detection task. This auxiliary task contributes to more accurate and detail-preserving change detection. We evaluate AFCN on three benchmark datasets, including LEVIR-CD, PX-CLCD, and WHU-CD. Experimental results demonstrate that AFCN achieves state-of-the-art performance, with an intersection over union (IoU) of 85.30 %, 94.13 %, and 90.03 % on the three datasets, respectively.
Journal
IF:
6.3
Papers:
7.8K
Citations:
3.0W
View more organizations