Mapping and Spatiotemporal Evolution Analysis of Salinity Water Mass Structure Characteristics at the Pearl River Estuary—A Framework Coupling Multi-Source Remote Sensing Information from GEE with Machine Learning Methods

DOI: 10.22161/ijeab.112.1

Abstract:
The distribution and variation of salinity in estuarine regions serve as a critical foundation for understanding the evolution of coastal ecosystems. However, obtaining long-term, continuous in-situ salinity data in estuarine areas with complex hydrodynamics and limited observational conditions poses significant challenges. To address this, this study focuses on the Pearl River Estuary (PRE), utilizing multi-source remote sensing data from 2015 to 2025 on the Google Earth Engine (GEE) platform to develop an unsupervised machine learning framework driven by multi-dimensional features. The framework aims to identify water mass structures closely associated with salinity distribution and their spatiotemporal evolution characteristics. The proposed method integrates multiple features, including Landsat 8-derived Total Suspended Solids (TSS), Colored Dissolved Organic Matter (CDOM), Normalized Difference Chlorophyll Index (NDCI), as well as Sentinel-1 backscatter coefficients σ⁰and Digital Elevation Model (DEM) data. It employs K-means clustering and XGBoost-SHAP analysis to classify surface water features and assess contributing factors. The results statistically delineate three typical water mass structures: freshwater-dominated, seawater-dominated, and their transitional mixtures, revealing significant interannual variability over a decade. The observed trends correlate with runoff conditions in the Pearl River Basin and regional hydroclimatic factors. Notably, during the extreme drought years of 2021–2022, the marked expansion of seawater-dominated water masses validated the model's sensitivity to structural adjustments in water masses. This study demonstrates that, in the absence of long-term in-situ data, the fusion of multi-source remote sensing and unsupervised machine learning methods can effectively identify the relative spatiotemporal evolution of salinity-related water mass structures in estuaries. It offers a feasible, low-cost approach for studying salinity dynamics in large estuaries and supporting regional water environment management.

Keywords:
The Pearl River Estuary (PRE); K-means Clustering; Water Mass Structure; XGBoost-SHAP; Google Earth Engine (GEE) ; Machine Learning (ML)

Article Info:
Received: 17 Jan 2026; Received in revised form: 19 Feb 2026; Accepted: 25 Feb 2026; Available online: 04 Mar 2026

Cite this Article: