Efficient metric-resolution land cover mapping using open-access low resolution annotations with prototype learning and modified Segment Anything model

Abstract

Large-scale metric-resolution land cover mapping is crucial for a detailed understanding of large areas of the Earth surface, supporting fine-scale decision-making in sectors such as agriculture, forestry, and conservation. Despite advances in remote sensing technology, this type of mapping is often limited by a lack of high-quality manually labeled data. To address this issue, here we propose a novel framework called Segment Anything Model with Prototype Learning and Enhanced Refinement (SAMPLER), using open-access and lower-resolution land cover products (LCPs) as labels for metric-resolution land cover mapping. Specifically, SAMPLER is built upon Segment Anything Model (SAM), leveraging its advanced feature extraction capability for precise and stable segmentation across diverse scenarios. To bridge the resolution gap between LCPs and metric-resolution imagery, we designed the class prototype-based object-oriented decoder (CPO-Decoder) to accurately classify small-scale features and maintain spectral consistency, effectively managing complex intra- and inter-class variations by dynamically aligning object-level features with semantic prototypes. Additionally, a label refinement strategy iteratively updates noisy LCP labels by replacing low-confidence annotations with high-precision model predictions, thereby mitigating label ambiguity caused by coarse resolution and enabling the model to progressively adapt to fine-scale features while improving classification accuracy throughout the training process. This framework exhibits exceptional adaptability and performance across various sensor types and diverse geographic regions, achieving an overall accuracy (OA) of 85.3% and a frequency-weighted intersection over union (FWIoU) of 74.7% across four study areas, with an OA increase from 5.6% to 9.5% and an FWIoU enhancement from 5.2% to 11.8% relative to the original LCP labels. Comparative experiments demonstrate that the proposed SAMPLER outperforms other deep learning models by up to 6.9% in OA and 9.8% in FWIoU. Ablation experiments further prove the effectiveness of CPO-Decoder, prototype learning and label-iterative-refined strategy. The overall results highlight the potential of SAMPLER for efficient metric-resolution land cover mapping at a large scale without manually labeled data, providing a valuable tool for fine-scale applications from environmental conservation to urban planning.