FedRS-Bench: A Realistic Federated Learning Dataset and Benchmark for Remote Sensing

Remote sensing (RS) data are naturally distributed across regions, institutions, and sensing platforms, making centralized model training difficult due to privacy, ownership, and data-sharing constraints. Federated learning (FL) provides a promising solution by enabling collaborative model training without exposing raw data. However, existing FL studies in RS are typically evaluated on ad hoc client partitions of a single dataset, which fail to reflect the source diversity, client imbalance, and domain heterogeneity of real-world deployments. In this paper, we present \textbf{FedRS}, a realistic federated learning dataset for remote sensing. FedRS is constructed from eight public RS datasets and contains 135 clients under a source-consistent construction principle, where data from the same client come from the same source. This design preserves realistic federated properties, including skewed label distributions, highly imbalanced client data volumes, and substantial cross-client domain shifts. Based on FedRS, we build \textbf{FedRS-Bench}, a unified benchmark that evaluates 10 representative FL algorithms under multiple heterogeneity settings, two global test protocols, and partial client participation. Extensive experiments show that federated training consistently outperforms isolated local learning, but also reveal that no existing FL method is uniformly robust under realistic RS heterogeneity, that optimization stability remains a major challenge, and that stronger privacy protection incurs a clear utility cost. We hope FedRS-Bench will provide a standardized and scalable testbed for realistic FL in remote sensing and facilitate fair comparison across future work. The source code and dataset are available at https://fedrs-bench.github.io/.