Comparative Evaluation of Machine Learning and Numerical Weather Prediction for Precipitation Estimation in Complex Orographic Terrain

Abstract

Accurate precipitation estimation in mountainous regions is critical for flood early warning systems, water resource management, and hydroelectric operations, yet it remains a formidable challenge due to complex orographic processes that Numerical Weather Prediction (NWP) models frequently fail to resolve. While Machine Learning (ML) and Deep Learning (DL) offer promising alternatives, a systematic, side-by-side comparative benchmarking against physics-based NWP under consistent experimental protocols is notably absent from the literature. This study introduces the Terrain-Aware Multi-Model Comparative Framework (TAM-CF) to rigorously evaluate the predictive performance of diverse ML/DL algorithms—including Random Forest, XGBoost, Artificial Neural Networks, Convolutional Neural Networks (CNN), Convolutional Long Short-Term Memory (ConvLSTM), and Graph Neural Networks (GNNs)—against the Weather Research and Forecasting (WRF) model for daily precipitation estimation in complex orographic terrain. Results demonstrate that all ML models significantly outperform NWP, with the topology-aware GNN achieving a 38.6% reduction in Root Mean Square Error (RMSE) and improving the Kling-Gupta Efficiency from 0.52 to 0.78. Stratified analysis reveals that ML provides the greatest relative advantage under heavy and extreme precipitation regimes, substantially mitigating the systematic dry bias inherent in NWP. Crucially, the optimal integration strategy is terrain-dependent: hybrid NWP+ML post-processing excels in valleys and moderate slopes, whereas standalone GNNs are demonstrably superior in data-sparse high-altitude peaks where NWP errors are extreme. This research advances the hydrometeorological community beyond simplistic "AI versus Physics" dichotomies, offering a stratified, terrain-sensitive benchmarking protocol that informs strategic model selection for operational forecasting in mountainous catchments.