Why do forest-based models work for radar rainfall estimation? Insights from 10-minute and hourly QPE experiments

Accurate precipitation estimation is fundamental for hydrological forecasting and disaster risk management, with radar-based Quantitative Precipitation Estimation (QPE) providing high-resolution rainfall input for real-time applications. In recent years, machine learning approaches have been widely adopted to improve radar QPE, with both forest-based models (e.g., Random Forest, gradient-boosted trees) and deep learning architectures outperforming traditional reflectivity-rain rate (Z-R) relationships. Despite this progress, most studies emphasise performance comparisons, offering limited insight into how forest-based models exploit radar-derived information across different temporal scales.

In this study, we move beyond accuracy benchmarks to investigate the predictive behaviour of forest-based models for radar rainfall estimation. We conduct a systematic set of experiments in which the input feature space is progressively expanded to include three-dimensional reflectivity profiles, derived radar products (e.g. MaxDBZ, VIL and so on), dual-polarization variables (e.g. Kdp), and geographical information.Model performance and feature importance are analysed for QPE at both 10-min and 1-h timescales.

Our results reveal clear, scale-dependent patterns in model behaviour. At the hourly timescale, predictive performance is primarily governed by simplified radar intensity measures (i.e. MaxDBZ) combined with geographic information, suggesting a dependence on regional weather patterns. In contrast, at the 10-min timescale, performance is more strongly associated with three-dimensional and vertically integrated radar features, indicating a more localized and dynamic regime. These findings highlight that forest-based models adapt their effective use of radar information depending on temporal scale, motivating further diagnostic analyses of ensemble behaviour to better characterise how tree-based models balance local and aggregated information in radar QPE.