Spatial–Spectral Prototype Calibration Network for Few-Shot Multispectral Object Detection in Remote Sensing

Few-shot multispectral object detection remains a formidable challenge in remote sensing, constrained by the scarcity of annotated data across heterogeneous modalities and environmental conditions. Existing transformer-based detectors, while powerful, often exhibit overfitting in low-sample regimes and fail to preserve cross-spectral consistency between visible and infrared channels. To address these limitations, this article presents few-shot spatial–spectral prototype calibration network (Few-SSPC-Net), a spatial–spectral prototype calibration network designed for efficient and adaptive few-shot multispectral object detection. Unlike conventional transformer-driven pipelines, this framework employs a transformer-free dual-branch convolutional architecture—one branch emphasizing spatial semantics and the other spectral correlations—bridged by a Cross-Scale Interaction Module for fine-grained feature alignment across modalities. Central to this framework is the proposed Spatial–Spectral Prototype Calibration Module, which dynamically refines class prototypes through spectral correlation-guided calibration between support and query samples. This mechanism mitigates prototype drift and enhances generalization by enforcing spectral angular consistency within the embedding space. The entire architecture is trained under an episodic meta-learning paradigm, optimizing a joint objective of classification, localization, and spectral calibration regularization. Extensive experiments on benchmark datasets demonstrate that Few-SSPC-Net achieves consistent gains over state-of-the-art few-shot detectors, with up to +4.7% mAP improvement under five-shot settings, while maintaining competitive inference efficiency. The results affirm the positioning of Few-SSPC-Net as a robust framework for multispectral object detection in complex, data-limited remote sensing scenarios.