GSMamba: GSD-Aware Autoregressive Pretraining for Mamba-Based Remote Sensing Foundation Models

Remote sensing (RS) foundation models (RSFMs) largely rely on self-supervised pretraining to improve downstream performance. The recent combination of Mamba and autoregressive pretraining offers an efficient direction for RS representation learning, yet existing methods process images at fixed pixel units and overlook the physical-semantic misalignment caused by varying ground sampling distance (GSD). To address this issue, we propose GSMamba, a GSD-aware autoregressive pretraining framework for Mamba-based RSFMs. At the data level, GSD-aware spatial resampling aligns inputs to an explicitly specified target GSD. At the feature level, Scale-Conditioned Channel Modulation (SCCM) and GSD-conditioned encoding (GSD CE) calibrate resampling-induced feature shifts and inject GSD information. During inference, GSMamba supports on-demand target GSD specification for controllable accuracy-computation trade-offs. Experiments show that GSMamba achieves consistent gains over RoMA and stronger adaptation under downstream GSD shifts.