Quantitative inversion of water quality parameters in the Jiulong River estuary based on machine learning and UAV hyperspectral data

Background The combination of machine learning and unmanned aerial vehicle (UAV) hyperspectral remote sensing demonstrates great potential for dynamic and refined water quality monitoring. Most existing studies focus on stable water bodies such as lakes, rivers and reservoirs, while relevant research remains relatively scarce for estuaries with complex optical properties, affected by intense coupled disturbances from natural processes and human activities. Objective The study aims to verify the effectiveness and feasibility of the proposed technique for application in estuarine regions. Methods This study conducted continuous UAV hyperspectral measurements and simultaneous water sampling across multiple tidal phases in the Jiulong River estuary. On this basis, quantitative inversion models for chlorophyll-a and suspended sediment were developed and evaluated, integrating different feature selection methods Pearson correlation analysis, principal component analysis (PCA), successive projections algorithm (SPA) and machine learning algorithms support vector regression (SVR), random forest (RF), multilayer perceptron (MLP). Results The results indicated that the optimal inversion model for chlorophyll-a was Pearson-RF, with R² of 0.83, RMSE of 0.41 μg/L and RPD of 2.50. The optimal model for suspended sediment was Pearson-RF, with coefficient of determination (R²) of 0.90, root mean square error (RMSE) of 3.63 mg/L and residual predictive deviation (RPD) of 3.26. The PCA-MLP model also showed reliable predictive performance in the retrieval of suspended sediment (R²=0.89, RMSE=3.85 mg/L, RPD=3.07). Conclusion Through this study, the Pearson-RF model exhibits excellent feature extraction capability and robustness, along with favorable inversion accuracy and generalization performance. In practical applications, the Pearson-RF model demonstrates prominent advantages in the small-sample and weak-signal scenarios, rendering it a universal algorithm suitable for the inversion of chlorophyll-a and suspended sediment concentrations. Specifically, the PCA-MLP model can achieve reasonable prediction accuracy for suspended sediment with distinct spectral response characteristics. These findings can provide methodological reference and technical support for quantitative inversion, dynamic monitoring and refined management of water quality in estuarine regions.