A study on automatic annotation methods for watershed environmental elements based on semantic segmentation models

Satellite remote sensing data provide a crucial spatiotemporal foundation for advancing digital-twin water conservancy construction in China. Given the complex and variable nature of watersheds and their surrounding elements, using remote sensing imagery to intelligently interpret watershed elements...

Full description

Saved in:
Bibliographic Details
Main Authors: Peipei He, Taoxing Shen, Yafei Wang, Dantong Zhu, Qingfeng Hu, Hui Li, Weibo Yin, Yi Zhang, Ang Yang
Format: Article
Language:English
Published: Taylor & Francis Group 2025-12-01
Series:European Journal of Remote Sensing
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/22797254.2025.2473939
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Satellite remote sensing data provide a crucial spatiotemporal foundation for advancing digital-twin water conservancy construction in China. Given the complex and variable nature of watersheds and their surrounding elements, using remote sensing imagery to intelligently interpret watershed elements poses significant challenges. Deep neural network models offer a potential unified solution; however, their effectiveness hinges on the quality and diversity of the training data. In a data-driven analytical approach, this study explores the segmentation accuracy and efficiency of intelligent-interpretation neural networks from the perspectives of watershed monitoring element types and sample annotation methods. Hence, experiments are conducted in three modules: (1) Determination of watershed environmental element sample types. (2) Semi-automated annotated sample production method. This study utilizes the Random Forest classifier (RF) and Segment Anything Model (SAM) to create label sets, replacing open-source land cover datasets to achieve superior watershed-environment segmentation results. (3) Comparison of different neural network segmentation capabilities. In the study area, SegFormer combined with SAM labels achieved an overall accuracy (OA) of 96.71%. The OA of SegFormer combined with RF labels was 91.71%, representing improvements of 6.24% and 1.24%, respectively, over the Sentinel-2 Land Cover Explorer (OA of 90.47%).
ISSN:2279-7254