Spatial-frequency collaborative feature constraint based on interval type-2 fuzzy set and wavelet transform for high-resolution remote sensing image segmentation

• Published in: Information sciences Vol. 721; p. 122639
• Main Authors: Chong, Qianpeng, Ma, Rong, Xu, Jindong, Wei, Guangyi, Long, Yao, Zeng, Wenyi, Cheng, Dong
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.12.2025
• Subjects: Interval type-2 fuzzy set; Remote sensing; Semantic segmentation; Wavelet transformer; Interval type-2 fuzzy set; Semantic segmentation; Wavelet transformer; Remote sensing
• ISSN: 0020-0255
• DOI: 10.1016/j.ins.2025.122639

Semantic segmentation of high-resolution remote sensing images (HRSIs) is a fundamental but challenging task in the computer vision and image processing fields and is paramount for various practical applications. Although numerous efforts have been made, bottlenecks still exist in accurately segmenting each object. We analyse in depth its challenges from two main sources: 1) The feature representation of objects in the spatial domain fluctuates and is uncertain because of a cluttered distribution; however, this uncertainty still lacks effective modelling methods. 2) The inherent complexity of HRSIs allows them to contain more information; however, the frequency representation is rarely considered. To compensate for the above inadequacies, a novel spatial-frequency collaborative feature constraint network (S-FCFNet) is proposed to provide an effective feature constraint method for the spatial and frequency domains. Specifically, S-FCFNet contains two core components to respond to the aforementioned problems, i.e., an interval type-2 fuzzy feature constraint module (IT2F2CM) and a wavelet transform frequency feature constraint module (WTF2CM). IT2F2CM designs a high-order fuzzy method and uses its interval property to model uncertain feature representations in the spatial domain. The WTF2CM introduces the Haar wavelet transform to constrain the frequency feature by decomposing different frequency information. IT2F2CM and WTF2CM provide effective feature constraints for the spatial domain and frequency domain, respectively. Finally, S-FCFNet uses the CNN–transformer hybrid backbone to leverage global and local features simultaneously and align various semantic representations. Extensive evaluations are conducted on three remote sensing datasets, and the experimental results illustrate that S-FCFNet effectively mitigates the above problems and further achieves excellent performance. The extensive results qualitatively and quantitatively illustrate the superiority of S-FCFNet, which achieves the best mIoU values of 73.86%, 73.15% and 80.49% on the three benchmark remote sensing datasets. The code will be publicly available at https://github.com/goccchong/IT2FM.