DB-CRNet: A novel weakly supervised framework for Cross-Resolution land cover classification

• A dual-branch cross-resolution network is proposed to enable high-resolution land cover mapping with coarse labels. • Developed a label-filtering dual-branch loss (LFDB-Loss) to mitigate coarse-label noise and class imbalance. • DB-CRNet achieves superior accuracy and clearer boundaries on Chesapeake and Guilin, showing cross-resolution generalization. High spatial resolution land cover products play a crucial role in urban planning and renewal by providing detailed spatial information for decision-making. However, generating these products through fully supervised classification approaches requires extensive pixel-level annotations, which are time-consuming and costly. Weakly supervised methods offer a practical alternative by using low spatial resolution land cover products as coarse labels. Nevertheless, the mismatch in resolution introduces significant label noise and blurred boundaries, limiting classification accuracy. To address these challenges, this study proposes a Dual-Branch Cross-Resolution Network (DB-CRNet) that uses only low spatial resolution labels and high spatial resolution images for land cover classification. DB-CRNet adopts a non-downsampling dual-branch framework, consisting of a Convolutional Neural Network (CNN) branch, a Wavelet Visual State Space Model (WVSSM) branch, and a Label-Filtering Dual-Branch Hybrid Loss (LFDB-Loss) module. This design enables effective cross-resolution learning while preserving fine spatial details. Experimental results show that DB-CRNet outperforms existing benchmark models (e.g., Paraformer, L2HNet), achieving an mIoU value of 72.46% on the public Chesapeake dataset and 63.48% on the self-built Guilin dataset. The results demonstrate that the proposed network exhibits strong generalization across different data sources and spatial resolutions. Overall, this study provides an efficient and economic method for high spatial resolution land cover mapping using low spatial resolution coarse labels, significantly reducing the labeling cost and ensuring the precision of classification results.