Reconstructing Daily Discharge in a Megadelta Using Machine Learning Techniques

In this study, six machine learning (ML) models, namely, random forest (RF), Gaussian process regression (GPR), support vector (SVR), decision tree (DT), least squares (LSSVM), and multivariate adaptive spline (MARS) were employed to reconstruct the missing daily-averaged discharge in a mega-delta from 1980 2015 using upstream-downstream multi-station data. The performance accuracy of each ML model assessed compared with stage-discharge rating curves (RCs) four statistical indicators, Taylor diagrams, violin plots, scatter time-series heatmaps. Model input selection was performed mutual information correlation coefficient methods after three data pre-processing steps: normalization, Fourier series fitting, first-order differencing. results showed that models are superior their RC counterparts, MARS RF most reliable algorithms, although achieves marginally better than RF. Compared RC, reduced root mean square error (RMSE) by 135% 141% absolute 194% 179%, respectively, year-round However, developed for climbing (wet season) recession (dry limbs separately worsened slightly Specifically, RMSE falling limb 856 1,040 m3/s, while obtained 768 789 respectively. DT is not recommended, GPR SVR provide acceptable results.