Vision transformer-based feature harmonization network for fine-resolution land cover mapping

Autonomous high-resolution (HR) land cover mapping is essential for analysing and understanding the Earth’s surface, particularly in addressing ecological challenges. However, this task is hindered by the lack of precise training datasets and the time and space complexity of existing algorithms. To address these challenges, this study introduces Vision Feature Harmonization learning (VHF-Para Learning), a Vision Transformer and CNN-based Feature Parallel learning framework designed to harmonize network knowledge acquisition through weak supervision. This deep learning approach leverages freely available global low-resolution (LR) land-cover products (GLC10) to guide fine-scale mapping. The GLC10 dataset supervises the network’s knowledge acquisition, while high-resolution Google Earth imagery is used to train the network. Traditional CNNs excel at preserving local ground information but often struggle to model global contexts effectively for multi-scale feature representation. VHF-Para incorporates Multi-layer Agent Graphical (MLA) modules for improved edge refinement and feature mask extraction using CNNs, along with a noise-label-assisted training (NLAT) module to refine low-resolution labels for weekly supervised semantic segmentation. Experiments on three large datasets show that VFH-Para surpasses existing methods in enhancing land-cover maps with GLC10 data, achieving an overall accuracy of 89.87% and a Kappa coefficient of 0.78.