A dissolved oxygen inversion method based on multi-source spectral-spatial feature fusion and recursive feature elimination.

Coastal ecosystem health is crucial for regional sustainability and biodiversity maintenance. As a key indicator of eutrophication and hypoxia risk, the accurate monitoring of dissolved oxygen (DO) concentrations is essential for coastal ecological security. Traditional monitoring methods are limited by narrow spatial coverage, low temporal resolution, and high costs, while remote sensing offers advantages in wide-area, periodic observation. However, existing remote sensing-based DO inversion models suffer from insufficient stability, feature redundancy, and limited ability to characterize spatial heterogeneity.To address these issues, this study takes the Bohai Sea, China, as the study area and constructs the MSSF-RFE (Multi-Source Spectral-Spatial Fusion with Recursive Feature Elimination) model framework, which integrates multi-spectral bands from Sentinel-2 and Landsat 8/9 with latitude and longitude spatial information to form a multi-source spectral-spatial feature set. Recursive Feature Elimination (RFE) is employed for feature optimization, and its performance is compared with Sequential Backward Selection (SBS) and Sequential Forward Selection (SFS). Subsequently, model construction is carried out using XGBoost, Random Forests (RF), Light Gradient Boosting Machine (LightGBM), and CatBoost. The results indicate that the MSSF-RFE approach combined with machine learning models can effectively select key features and significantly improve inversion accuracy, outperforming the SBS and SFS methods. Among the models, MSSF-RFE-CatBoost achieves the best performance (R2 = 0.8753, RMSE = 0.4032 mg/L). The MSSF-RFE framework proposed in this study provides an efficient, stable, and promising method for the inversion of Bohai Sea dissolved oxygen concentrations driven by the synergy of multi-source remote sensing and spatial information, holding significant importance for ecological environment monitoring and long-term dynamic assessment in this region.