Core Concepts
Ensemble learning improves probabilistic predictions in spatial interpolation.
Abstract
The study introduces nine quantile-based ensemble learners for predicting precipitation, focusing on merging remote sensing and gauge-measured data. Stacking methods with QR and QRNN outperformed other algorithms, demonstrating improvements ranging from 3.91% to 8.95%. A novel feature engineering strategy using distance-weighted satellite data was employed to reduce predictor variables while maintaining performance. The dataset comprised 91,623 samples, each containing 10 values, split into three sets for training and testing algorithms. Performance comparisons were based on quantile scores at multiple levels, showcasing the potential of ensemble learners in improving predictive uncertainty estimation.
Stats
Stacking with QR and QRNN yielded improvements ranging from 3.91% to 8.95%.
Dataset comprised 91,623 samples with 10 values each.
Novel feature engineering strategy used distance-weighted satellite data to reduce predictor variables.