Adaptive Gradient Calibration for Single-Positive Multi-Label Learning in Remote Sensing Image Scene Classification

TL;DR

This work tackles scalable RS image scene classification under single-positive multi-label learning (SPML) by introducing AdaGC, a framework that calibrates gradients to mitigate false negatives, generates robust pseudo-labels with a dual-EMA scheme, and adaptively activates gradient calibration via learning-dynamics analysis. The approach integrates Mixup and provides a theoretical basis for early-learning detection, ensuring GC is engaged at an optimal training stage. Extensive experiments on reBEN and AID-multilabel under Random and Dominant SPML noise demonstrate state-of-the-art performance and robust generalization, with ablations confirming the contributions of the adaptive trigger, dual-EMA, and Mixup. AdaGC offers a practical, noise-robust solution for RS SPML, highlighting its potential for scalable, high-quality multi-label RS classification.

Abstract

Multi-label classification (MLC) offers a more comprehensive semantic understanding of Remote Sensing (RS) imagery compared to traditional single-label classification (SLC). However, obtaining complete annotations for MLC is particularly challenging due to the complexity and high cost of the labeling process. As a practical alternative, single-positive multi-label learning (SPML) has emerged, where each image is annotated with only one relevant label, and the model is expected to recover the full set of labels. While scalable, SPML introduces significant supervision ambiguity, demanding specialized solutions for model training. Although various SPML methods have been proposed in the computer vision domain, research in the RS context remains limited. To bridge this gap, we propose Adaptive Gradient Calibration (AdaGC), a novel and generalizable SPML framework tailored to RS imagery. AdaGC adopts a gradient calibration (GC) mechanism with a dual exponential moving average (EMA) module for robust pseudo-label generation. We introduce a theoretically grounded, training-dynamics-based indicator to adaptively trigger GC, which ensures GC's effectiveness by preventing it from being affected by model underfitting or overfitting to label noise. Extensive experiments on two benchmark RS datasets under two distinct label noise types demonstrate that AdaGC achieves state-of-the-art (SOTA) performance while maintaining strong robustness across diverse settings. The codes and data will be released at https://github.com/rslab-unitrento/AdaGC.

Figures (7)

• Figure 1: Single- and multi-label annotation examples from (a) AID-multilabel hua_relation_2020 and (b) refined BigEarthNet clasen_reben_2025_arxivonly datasets, where the corresponding CLC mask is presented for reference. Compared to the single-class labels, the multi-label annotations can more comprehensively describe the scene.
• Figure 2: Flowchart of the proposed Adaptive Gradient Calibration (AdaGC) method for single-positive multi-label learning in remote sensing image classification. Pseudo-labels $\mathbf{t}$ are generated by combining the teacher model’s predictions $\mathbf{p}^T$ and the student model’s predictions $\tilde{\mathbf{p}}^S$ according to \ref{['eq:pseudo-label']}. EMA is also applied to the student model’s predictions during the warm-up stage, which is omitted for simplification.
• Figure 3: True and noisy validation mAP trends of the teacher model during training on refined BigEarthNet clasen_reben_2025_arxivonly: (a) random SPML, (b) dominant SPML. The trends align with Propositions 1 and 2, and confirm the validity of the stationarity theorem.
• Figure 4: Histograms of the number of GT labels per image in the training sets.
• Figure 5: Validation accuracies (mAP and mF1) over training epochs obtained with predefined warm-up lengths (10, 20, 30) and our proposed early learning detection strategy (AdaGC).