Noise-Adaptive Regularization for Robust Multi-Label Remote Sensing Image Classification

A new regularization method called NAR aims to make multi-label remote sensing image classification more robust by explicitly distinguishing between additive and subtractive label noise, a distinction previous work has largely overlooked [1]. The method, detailed in a paper by Tom Burgert and submitted to arXiv on 13 Jan 2026, addresses a growing problem in remote sensing [1]. As the volume of satellite and aerial imagery expands, annotation procedures increasingly rely on thematic products or crowdsourced procedures to reduce costs, which often introduces partially incorrect labels [2]. This label noise can appear in three forms: additive noise, where incorrect labels are added; subtractive noise, where correct labels are missing; or a combination of both, known as mixed noise [3]. Existing approaches typically treat all noisy annotations as valid supervised signals, lacking mechanisms that adapt learning behavior to different noise types [3]. NAR operates within a semi-supervised learning framework and employs a confidence-based label handling mechanism. Label entries with high confidence are retained, entries with moderate confidence are temporarily deactivated, and low-confidence entries are corrected by flipping them [4]. This selective attenuation of supervision is integrated with early-learning regularization, or ELR, to stabilize training and mitigate overfitting to corrupted labels [2]. In experiments across additive, subtractive, and mixed noise scenarios, NAR consistently improved robustness compared with existing methods, with the most pronounced gains under subtractive and mixed noise [4]. The work builds on a broader research effort to handle noisy labels in multi-label classification. A separate framework, Adaptive Gradient Calibration, or AdaGC, proposed a gradient calibration mechanism with a dual exponential moving average module for robust pseudo-label generation, also targeting noise in remote sensing imagery [5]. The NAR paper was revised on 4 Jun 2026, with the submission size growing from 1,600 KB to 1,611 KB [1].