Rethinking multilabel remote sensing image classification under dual-distribution instance-dependent noise

Zhang, Xinlei, Liu, Fan, Zhang, Chuanyi, Ji, Xiaoying, Wang, Wenhui, Zhou, Wei and Zhou, Jun 2026. Rethinking multilabel remote sensing image classification under dual-distribution instance-dependent noise. IEEE Transactions on Geoscience and Remote Sensing 10.1109/tgrs.2026.3740916

Full text not available from this repository.

Abstract

Multilabel scene classification (MLC) is a fundamental task in remote sensing (RS). However, its performance is often limited by noisy annotations, where subtractive noise occurs when inconspicuous targets are missed and additive noise arises when visually confusing categories are falsely added. To simulate such annotation errors, the recently proposed Instance- Dependent Multilabel Noise (IDMN) method introduces the first instance-dependent noise generation strategy for RS MLC, driven by the intuition that hard label instances are more susceptible to mislabeling. Specifically, it leverages a single probability distribution of present labels to derive thresholds for both subtractive and additive noise generation. This design determines additive noise by requiring absent-label occurrence probabilities to exceed an irrelevant present-label threshold, leaving its quantity difficult to control and overlooking potentially confusing absent labels. To address this limitation, we propose Dual-Distribution Instance-Dependent Multilabel Noise (DD-IDMN), which explicitly decouples the score distributions of present and absent labels for calibrated noise generation. Empirical analysis under DD-IDMN reveals that additive noise yields more severe performance degradation than subtractive noise. To combat this DD-IDMN challenge, we propose the Prompt-guided Activation Interference Rectification Network (PAIR-Net). Through cross-modal attention with learnable text prompts, it learns more discriminative RS feature representations. Furthermore, we introduce an activation-aware dynamic filtering module to identify unreliable present labels for suppressing the more destructive additive noise under DD-IDMN. Experiments on UCMerced, AID, and MLRSNet demonstrate that PAIR-Net consistently outperforms representative noise-robust multilabel methods under complex instance-dependent noise. Our code is available at https://github.com/BasilLeave/PAIR-Net.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics
Publisher:	Institute of Electrical and Electronics Engineers
ISSN:	0196-2892
Last Modified:	12 Oct 2026 13:30
URI:	https://orca.cardiff.ac.uk/id/eprint/190159

Actions (repository staff only)

Edit Item