Key-Region Prototype Contrastive Learning for Few-Shot PAN–MS Land-Cover Classification

Abstract

Few-shot panchromatic (PAN)–multispectral (MS) remote sensing image classification is challenging due to limited labeled samples and substantial cross-modal differences. PAN images provide rich spatial and structural details, whereas MS images contain discriminative spectral information. To address background interference and sample selection bias, we develop TK-Net to improve task construction and cross-modal representation learning under scarce supervision. TK-Net consists of two key components. First, a task-adaptive sampling (TAS) module replaces conventional random episodic sampling with a feature-driven strategy. It selects dispersed support samples and hard query samples near confusing decision boundaries, thereby improving prototype representativeness and boundary discrimination. Second, a key region mining and alignment (KMA) module identifies cross-modally consistent foreground regions through superpixel-guided region modeling and performs region-selective prototype contrastive alignment, thereby suppressing irrelevant background responses and improving cross- modal semantic consistency. Experiments on three dual-source remote sensing benchmarks demonstrate the effectiveness of TK-Net. Our code is available at: https://github.com/Xidian-AIGroup190726/TKnet