SGMFNet: A Semantic-Guided Multifrequency PolSAR Data Fusion Framework Based on Scattering Mechanisms

Owing to the physical diversity across radar bands, multifrequency polarimetric synthetic aperture radar (PolSAR) data exhibit strong complementarity, making them valuable for enhancing remote sensing representations. Although multifrequency SAR fusion has garnered increasing attention, existing efforts predominantly focus on downstream tasks such as classification, whereas upstream data-level integration remains largely underexplored. Such integration is crucial for preserving band-specific physical properties and leveraging interfrequency complementarities, thereby enabling accurate scattering characterization and providing robust, interpretable inputs for subsequent tasks. To address this gap, we propose a semantic-guided multifrequency PolSAR fusion network (SGMFNet), which leverages physical scattering mechanisms to guide deep feature learning in a two-stage design. First, a semantic guidance map derived from polarimetric scattering mechanisms is embedded as a region-adaptive prior, encoding interpretation reliability and object-specific scattering sensitivity to guide fusion in a physically constrained manner. Second, SGMFNet employs channel-wise independent modeling to rigorously preserve the inherent decoupling among polarization channels. By combining multiscale feature extraction (MSFE) with a cross-frequency channel fusion module, the network effectively captures complementary spatial and scattering information across frequencies, enriching the fused representation. Extensive experiments conducted on two publicly available PolSAR datasets, covering various sensor types, frequencies, and platform configurations, systematically evaluate the performance of SGMFNet in multiple tasks. The results demonstrate that the fused data outperform single-band data in preserving scattering mechanisms and fine textural details while significantly improving the accuracy and robustness of downstream applications.