High-resolution land cover mapping from coarse labels via a noisy label learning-guided cross-scale framework

High-resolution remote sensing images (HRSIs) provide essential data support for land cover mapping, where deep learning has shown great promise. However, deep learning-based methods rely on abundant high-quality annotations, while low-resolution coarse labels are difficult to use directly in HRSIs training. In this paper, a novel noisy label learning-guided cross-scale framework (NL-CSF) is proposed, which is designed to achieve high-resolution land cover mapping from coarse labels. First, a spectral-based label mask filtering strategy is developed to preliminarily optimize coarse labels. Then, an adaptive noise evaluation scheme is introduced that assigns loss weights based on the noise differences between image and label patches in the training set. Finally, we design a cross-scale transfer Transformer (CSTT) model based on the vision Transformer (ViT) architecture, and the training process is guided by a noise-weighted loss function. Two cross-scale datasets are utilized to evaluate the performance of NL-CSF in multiple spatial scale differences (10 m to 3 m, 3 m to 0.5 m, and 10 m to 0.5 m). Experimental results demonstrate that NL-CSF improves overall accuracy (OA) by at least 7%, 6%, and 4% across the three cross-scale tasks in the first dataset, and by at least 2%, 9%, and 4% in the second dataset, respectively, compared with existing methods. Furthermore, the proposed framework is applied to cross-scale mapping across Jianye District of Nanjing (urban), Sheyang County of Yancheng (agricultural), and the Yellow River Delta of Dongying (wetland), leveraging a low-resolution land cover product and high-resolution PlanetScope images to generate more precise land cover maps. These results demonstrate the effectiveness of the proposed framework in mitigating the impact of noisy coarse labels and generating reliable high-resolution land cover maps.