A Hybrid Spatiotemporal Approach for Severe Rainfall Nowcasting Using Kriging and Recurrent Neural Networks

Abstract

Accurate nowcasting of localized extreme rainfall remains particularly challenging when observations are based on sparse and irregular rain-gauge networks. This article presents a framework for 1-h-ahead severe rainfall nowcasting that integrates spatial reconstruction, spatiotemporal deep learning, and extremes-oriented assessment while explicitly decoupling interpolation uncertainty from forecasting performance. Rain-gauge observations are reconstructed into regular precipitation fields through ordinary kriging (OK) and used to train forecasting models exclusively on severe rainfall events. Using a leave-one-event-out protocol, persistence, long short-term memory (LSTM), and ConvLSTM architectures are compared under continuous and threshold-based categorical metrics designed for extreme rainfall detection. Results show that ConvLSTM models reduce average prediction errors in several configurations, whereas persistence remains a strong baseline for detecting localized extreme cores. The proposed framework establishes a reproducible methodology for developing and fairly evaluating nowcasting systems over sparse observation networks, providing a foundation for future AI-based forecasting methods focused on high-impact rainfall events.