2026· IEEE Transactions on Geoscience and Remote Sensing· Vol 64, pp. 5531018-5531018· 0 citations· 55 references
TL;DR
A dynamic prototype interaction calibration module, which adaptively distinguishes typical samples from hard samples via the prediction confidence of query samples, and a synergistic global domain alignment strategy that jointly aligns macroscopic spatial distributions and spectral baselines while preserving class-discriminative structures is developed.
Abstract
Coastal wetland hyperspectral image classification (HSIC) is practically constrained by scarce labeled samples. Cross-domain few-shot learning (CDFSL) provides a promising solution by enabling knowledge transfer under limited supervision. However, coastal wetlands exhibit high spectral similarity and spatial heterogeneity. Existing CDFSL methods struggle to disentangle such complex geomorphology, leading to insufficient feature discriminability. Furthermore, severe prototype estimation bias and domain shift persist as inherent challenges. To address these issues, an adaptive calibration strategy for CDFSL is proposed. Specifically, a collaborative perception feature extractor (CPFE) with a spectral–spatial adaptive differencing module (SSAD) and a dual-stream gated fusion module (DGFM) is constructed to extract the fine-grained signatures and complex spatial structures. Building upon these robust representations, we design a dynamic prototype interaction calibration (DPIC) module, which adaptively distinguishes typical samples from hard samples via the prediction confidence of query samples, thereby constructing and fusing multiple auxiliary prototypes to enable class prototypes to better approximate the latent data distribution. Finally, we develop a synergistic global domain alignment (SGDA) strategy that jointly aligns macroscopic spatial distributions and spectral baselines while preserving class-discriminative structures. Extensive experiments on three self-constructed hyperspectral image (HSI) datasets and one public dataset demonstrate the superiority of the proposed CDFSL method for HSIC via adaptive calibration (ACFSL).
Few-shot hyperspectral image classification (FS-HSIC) remains highly challenging because the limited number of labeled samples not only makes traditional supervised learning models prone to overfitting but also restricts their ability to learn generalizable spatial–spectral representations. Moreover, existing methods a...
A collaborative feature and classifier adaptation network (CF-CAN) that jointly enhances representation robustness and decision reliability under open-set cross-scene conditions and improves both cross-scene generalization and open-set discrimination is proposed.
Wu-Li Wang, Yun Liu, Jian-Bu Wang et al.· IEEE Transactions on Geoscie...· 0 citations
The adaptive basis spectral network (ABSNet), a lightweight domain generalization (DG) framework that learns adaptive low-rank spectral bases for constructing sample-specific low-rank adapters that achieves accuracy comparable to state-of-the-art (SOTA) methods with only a small fraction of the parameters of the compar...
Jian Qiu, Meng-Qi Chai, Shu-Ming Xiong et al.· IEEE Transactions on Geoscie...· 0 citations
Cross-scene hyperspectral image classification (HSIC) under unsupervised domain adaptation (UDA) remains challenging due to substantial distribution discrepancies caused by variations in acquisition conditions. In this setting, labeled samples are only available in the source domain, while the target domain lacks super...
Hao-Nan Chang, Li-Na Yang, Guan-Yu Qin et al.· IEEE Transactions on Geoscie...· 0 citations
Vision-language multimodal learning has exhibited remarkable advantages in few-shot hyperspectral image (HSI) classification, where prompt learning effectively enhances feature-extraction accuracy and representation quality by guiding the model to focus on critical information. However, static prompts lack flexibility,...
Yu-Hang Li, Jin-Rong He, Xiang-Qing Zhang et al.· IEEE Transactions on Geoscie...· 1 citation
Most existing semi-supervised hyperspectral classification methods are built upon the closed-set assumption, which assumes that unlabeled samples originate exclusively from the known class space. However, this assumption is frequently invalid in real-world remote sensing scenarios. In addition, hyperspectral images (HS...
Hua-Ping Zhou, Shuai Xu, Man Chen et al.· IEEE Transactions on Geoscie...· 3 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.