Work

Large-scale tidal wetland mapping based on point-sample-driven debiased and label noise learning: A case study in Asia

Year: 2025

Type: article

Abstract: Reliable tidal wetland maps are essential for coastal monitoring, as the extent and spatial distribution of these ecosystems reflect the complex dynamics of the land–sea interface, shaped by both natural processes and human activities. However, the inherent spatiotemporal variability and dynamic flooding patterns of tidal wetlands pose substantial challenges to conventional remote sensing classification approaches. In recent years, deep learning-based methods have shown considerable promise for tidal wetland mapping, achieving high accuracy when ample training data are available. However, for large-scale wetland mapping, the existing frameworks struggle to obtain enough tidal wetland samples, as most available data are point-based rather than pixel-wise, whereas deep learning models generally require dense annotated labels. In this paper, given the challenges posed by the complexity of tidal wetland formation and succession on sample quantity and quality, we propose a point-sample-driven tidal wetland mapping method based on debiased sampling and noise-resistant learning. Historical point samples are first converted into pixel-wise pseudo-labels to train a deep learning classifier. To alleviate sample selection bias, covariate statistical characteristics are explicitly modeled, ensuring that the generated pseudo-labels better reflect the true spatial distribution of tidal wetlands. To mitigate semantic errors in pseudo-labels, confusion probabilities between tidal wetland categories are incorporated as prior knowledge, enabling the identification and suppression of label noise. The effectiveness of the proposed method was tested across the tidal regions of Asia, covering a coastline of approximately 3,800,000 km2. Comparative experiments demonstrate that our framework consistently outperforms existing approaches, yielding substantial improvements in classification accuracy. The mapping results are publicly available at https://github.com/yingxinwu/AsiaTidalWetland.git.

Source: International Journal of Applied Earth Observation and Geoinformation

DOI: 10.1016/j.jag.2025.104913

Authors: Yingxin Wu, Xian Xu, Yinhe Liu, Sunan Shi, Ailong Ma, Yanfei Zhong

Institutions: Wuhan University, State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing, Changjiang Water Resources Commission, Changjiang River Scientific Research Institute

Language: English

Topic: Remote-Sensing Image Classification

Subfield: Media Technology

Field: Engineering

Domain: Physical Sciences

SDG: Clean water and sanitation

Open Access status: gold

Estimated APC paid: $3,550