FAD-Net: Frequency-Domain Amplitude-Phase Decoupling Network for Optical-Elevation Remote Sensing Segmentation

Multi-modal remote sensing semantic segmentation leverages multi-source information for fine-grained scene analysis. However, due to the inherent limitations of spatial-domain representations, achieving effective feature interaction and reconstruction remains challenging. In this paper, we propose the Frequency-Domain Amplitude-Phase Decoupling Network (FAD-Net). To resolve the fusion bottleneck caused by heterogeneous feature misalignment in the spatial domain, we design the Multi-Modal Mona (MM-Mona) module. It establishes an interaction mechanism based on decoupled attributes recombination, physically decoupling textural and structural components to achieve complementarity between heterogeneous modalities. To address mutual interference between deep and shallow features in spatial decoding strategies, we propose the Amplitude-Phase Recombination Pyramid (APRP) Decoder. A frequency-domain recombination mechanism is designed to reconstruct shallow details without losing deep semantic priors, guided by phase-geometric constraints. This effectively resolves the contradiction between preserving semantic consistency and recovering spatial details. Experimental results on the ISPRS Vaihingen and Potsdam datasets confirm that FAD-Net achieves state-of-the-art performance. The source code will be made available at https://github.com/Apsaras2005/FAD-Net.