Support Vector Machines with pre-and post-processing techniques for forecasting ice onset dates on the Yellow River

Abstract

The ice onset date is closely related to environmental conditions, which have fundamental impacts on river ecosystems and local communities. However, forecasting of this ecological indicator is seldom reported. In this study, several Support Vector Machine (SVM) models coupled with pre- and post-processing techniques were developed and evaluated to enhance prediction accuracy of ice onset dates at the Toudaoguai Station on the Yellow River, China. Initially, a five-factor SVM model was constructed using historical data of ice onset dates and associated factors. Subsequently, Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD) were employed to decompose the ice onset date series into multiple intrinsic mode functions (IMFs). These SVM models were then developed for each IMF and residual, and their predictions were aggregated. Furthermore, we explored SVM models with reduced input factors and integrated them using the Simple Averaging Method (SAM) to improve the overall performance. Given the limited dataset, the Leave-One-Out Cross-Validation (LOOCV) method was employed for rigorous model comparison. The results indicate that while the five-factor SVM model demonstrated strong predictive capability, the integration of pre-processing techniques (EMD and VMD) did not significantly enhance performance. However, the post-processing method using SAM with reduced input factors achieved the best results, highlighting the effectiveness of ensemble learning in this context. The practical usage of the proposed forecasting method was also discussed. This study provides novel tools for ice onset date forecasting, offering valuable insights for ice regime management and flood prevention in the Yellow River.