NRMLC: A Noise-Robust Multilabel Remote Sensing Scene Classification Framework Based on Labelwise Attention and Transformer-Based Classification

Abstract

Multilabel remote sensing scene classification (ML-RSSC) aims to assign multiple semantic labels to complex high-resolution scenes. Although exhaustive expert annotation enables accurate supervision, its high cost limits large-scale applications. Low-cost labels from crowdsourcing, existing geographic products, or weak annotation pipelines provide a scalable alternative, but they often contain missing and false-positive labels. Such noisy supervision, together with object-scale variation and severe label imbalance, makes high-precision ML-RSSC challenging. To address this issue, we propose NRMLC, a noise-robust framework for learning from low-cost noisy annotations. NRMLC first constructs multiscale scene representations through a multiscale feature extraction module to preserve objects at different spatial scales. A labelwise enhanced attention module then converts these representations into label-specific attention maps and classwise feature embeddings, alleviating the bias caused by sparse positive labels. The resulting embeddings are processed by a transformer-based multilabel classification module to model label dependencies and generate dependency-aware prediction probabilities. During optimization, a multilabel self-adaptive training strategy dynamically updates soft targets to reduce the influence of missing and false-positive annotations. Experiments on MLRSNet and UC Merced under additive, subtractive, and mixed label noise demonstrate the effectiveness of NRMLC. On MLRSNet, NRMLC reduces Hamming loss by up to 43.3% and improves mean F1 by more than 10% under challenging noisy-label settings. Ablation and probability analyses further confirm that these components jointly enhance robustness under noisy supervision.