Zhang et al. (2025) A Dual-Branch Framework Integrating the Segment Anything Model and Semantic-Aware Network for High-Resolution Cropland Extraction

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-10-13
• Authors: Dujuan Zhang, Yi‐Ping Li, Yucai Shen, Hengliang Guo, Haitao Wei, Jian Cui, Gang Wu, Tian He, Lingling Wang, Xiangdong Liu, Shan Zhao
• DOI: 10.3390/rs17203424

Research Groups

[Information not provided in the paper text.]

Short Summary

This study proposes SAM-SANet, a novel dual-branch framework integrating the Segment Anything Model (SAM) with a semantically aware network, to overcome challenges in precise boundary localization and cross-domain adaptability for high-resolution cropland extraction, demonstrating superior performance on custom datasets.

Objective

• To address the limitations of the Segment Anything Model (SAM) in semantic expressiveness and cross-domain adaptability when applied to high-resolution cropland extraction, specifically focusing on enhancing precise boundary localization.

Study Configuration

• Spatial Scale: High-resolution remote sensing imagery for cropland extraction.
• Temporal Scale: Not specified.

Methodology and Data

• Models used: Segment Anything Model (SAM), SAM-SANet (proposed dual-branch framework), semantic segmentation network.
• Data sources: High-resolution remote sensing imagery. Three custom cropland datasets: GID-CD, JY-CD, and QX-CD.

Main Results

• SAM-SANet achieved mIoU scores of 87.58%, 91.17%, and 71.39% on the GID-CD, JY-CD, and QX-CD datasets, respectively.
• SAM-SANet achieved mF1 scores of 93.54%, 95.35%, and 82.21% on the GID-CD, JY-CD, and QX-CD datasets, respectively.
• Comparative experiments confirmed the superior performance of SAM-SANet in high-resolution cropland extraction compared to mainstream semantic segmentation models.

Contributions

• Proposal of SAM-SANet, a novel dual-branch framework integrating SAM with a semantically aware network, specifically designed for high-resolution cropland extraction.
• Introduction of a boundary-constrained SAM branch to enhance attention to boundary information and improve extraction performance.
• Development of a boundary-aware feature fusion module to aggregate multi-scale edge information for enhanced boundary representation.
• Implementation of a prompt generation and selection module within the SAM branch to generate boundary-relevant prompts for precise localization.
• Construction of three new cropland datasets (GID-CD, JY-CD, QX-CD) for evaluating high-resolution cropland extraction methods.

Funding

[Information not provided in the paper text.]

Citation

@article{Zhang2025DualBranch,
  author = {Zhang, Dujuan and Li, Yi‐Ping and Shen, Yucai and Guo, Hengliang and Wei, Haitao and Cui, Jian and Wu, Gang and He, Tian and Wang, Lingling and Liu, Xiangdong and Zhao, Shan},
  title = {A Dual-Branch Framework Integrating the Segment Anything Model and Semantic-Aware Network for High-Resolution Cropland Extraction},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs17203424},
  url = {https://doi.org/10.3390/rs17203424}
}