Spectral–Spatial Collaborative Calibration for Multi-Source Domain-Adaptive Semantic Segmentation of Remote Sensing Images

Multi-source unsupervised domain adaptation (MS-UDA) can reduce the annotation burden of remote sensing semantic segmentation, but heterogeneous source spectra and ground sampling distances can destabilize target pseudo-labels. This article presents a spectral–spatial collaborative calibration network (SSCCNet) for MS-UDA with RGB and infrared–red–green (IRRG) imagery. First, spectral adaptive normalization (SANM) performs category-conditioned statistical matching from the near-infrared channel of a labeled nonreference source to the red-channel statistics of an RGB reference source, while weak blending preserves source-specific contrast. Second, spatial granularity alignment filtering (SAF) provides resolution-aware contextual compensation for the lower-resolution Vaihingen branch through gated fusion of identity and dilated-convolution responses. Third, dual-threshold filtering (DTF) constructs detached pseudo-labels from the current network, combining posterior-confidence screening with a restricted set of near-infrared intensity checks and class-aware weights. Experiments are conducted on two adaptation tasks constructed from ISPRS Potsdam RGB, Potsdam IRRG, and Vaihingen IRRG data. SSCCNet obtains mIoU/OA scores of 40.53%/64.00% and 57.82%/79.36% on the two tasks, respectively. In the cumulative ablation, performance increases at each stage as SANM, SAF, and DTF are introduced in sequence.