Few-shot semantic segmentation (FSS) for remote sensing faces major challenges due to substantial intra-class variation. Existing methods primarily rely on support features for guidance, often resulting in incomplete or inaccurate segmentation when the support and query images exhibit significant differences. To address this, we propose a Progressive Query-Driven Learning (PQDL) framework that fully exploits query-specific semantics through a two-stage process. In the first stage, the Query-Driven Anchor Generation (QDAG) module fuses complementary semantics from both support and query domains to produce a discriminative anchor prototype that guides the initial segmentation and suppresses irrelevant features. In the second stage, the Progressive Prototype Adaptation (PPA) module refines the segmentation by extracting a new query-specific prototype from high-confidence predictions and applying a self-correlation mask to enforce local semantic consistency. This transition from support-dependent to query-centric learning allows the model to adapt more effectively to diverse query instances. Extensive experiments on iSAID-5 \(^i\) and LoveDA-2 \(^i\) benchmarks demonstrate that PQDL significantly outperforms state-of-the-art methods, confirming its robustness and adaptability in complex remote sensing environments.