遥感
计算机科学
分割
人工智能
遥感应用
图像分割
大气模式
空格(标点符号)
计算机视觉
国家(计算机科学)
地质学
合成孔径雷达
空间技术
雷达成像
钥匙(锁)
信号处理
作者
Ye Li,Shuli Cheng,Anyu Du,Zhixiong Yang,Chao Liu
标识
DOI:10.1109/tgrs.2026.3690600
摘要
In remote sensing semantic segmentation, man-made objects often exhibit strong directional structures that manifest multi-frequency cues ranging from low-frequency contextual layouts to high-frequency boundary details. However, most existing approaches either prioritize spatial-domain modeling or treat frequency cues only as auxiliary enhancements and fusion signals, leaving hierarchical frequency organization and inter-band dependencies underexplored. To address this, we propose FBSNet, the first frequency-band state space network for remote sensing semantic segmentation. The key insight of FBSNet is to shift dependency modeling from the spatial domain to an ordered frequency-band sequence, enabling explicit cross-band coordination from low-frequency context to high-frequency directional details. First, its core Frequency-Band State Space (FBSS) module learns a structured frequency-band decomposition and serializes sub-bands into a direction-aligned sequence, along which state space modeling captures cross-band dependencies and preserves orientation-consistent cues from context to boundary details. Second, we integrate a Hierarchical Spatial Mamba (HSM) block into the decoder to enforce cross-scale structural continuity, yielding a balanced frequency-spatial representation. Moreover, a Fourier-guided refinement strategy leverages frequency-domain cues to mitigate boundary information loss during upsampling. Extensive experiments on ISPRS Vaihingen, ISPRS Potsdam, and LoveDA consistently validate the effectiveness of FBSNet, which attains state-of-the-art mIoU (85.84% / 87.90% / 55.47%). The code is available at: https://github.com/liye1213/FBSNet.
科研通智能强力驱动
Strongly Powered by AbleSci AI