Dual-Level Prototype Alignment via Cross-Attention for Few-Shot Remote Sensing Semantic Segmentation

Abstract

Analyzing remote sensing imagery relies heavily on accurate pixel-level classification to interpret complex geographical scenes. However, training robust deep learning models demands densely annotated datasets, which are expensive and time-consuming to create. Few-shot semantic segmentation (FSS) provides a practical alternative by adapting to new categories from limited labeled examples. Despite success in natural images, direct application to remote sensing faces unique challenges, such as massive scale variations, strong visual similarities among densely packed objects, and severe background interference. Existing FSS methods typically rely on masked average pooling for prototype extraction, which discards critical spatial geometries and fine-grained details. While some approaches attempt multi-scale or global-local modeling, they lack effective, decoupled alignment between support prototypes and query features, limiting their ability to simultaneously capture broad scene context and precise object boundaries in cluttered aerial views. To address these limitations, we propose DLPANet, a novel dual-level prototype alignment network centered on Prototype-Guided Spatial Attention (PGSA). Our framework constructs global and semantic prototypes from hierarchical backbone features, enabling simultaneous modeling of scene context and fine-grained details. Dual cross-attention modules then align these prototypes with decoupled global and semantic query feature maps, facilitating dynamic and precise feature enhancement that better distinguishes adjacent objects and suppresses background ambiguity. A learnable decoder integrates the enriched features, optimized via Dice loss to handle extreme class imbalance. Evaluated on the iSAID-5 benchmark, DLPANet achieves 36.12% and 41.05% mIoU in 1-shot and 5-shot settings, respectively, consistently outperforming state-of-the-art methods. These results demonstrate that our decoupled dual cross-attention mechanism provides superior prototype-query alignment compared to prior global-local strategies. https://github.com/mansoorfateh83-lab/dlpanet