Importantly, in areas with extensive, high-quality weather radar precipitation datasets, including portions of the USA and Europe, machine learning (ML) methods have demonstrated the ability to outperform traditional optical flow-based solutions for certain verification metrics and use cases, while reducing compute costs and latency. Consequently, implementations of recently published academic and industry-developed ML models are now utilized in both government and commercial precipitation prediction. This work presents relevant comparisons between traditional and ML nowcasting methods tailored to end-user needs, identifies challenges in employing ML nowcasting across the Globe’s heterogeneous observation networks, provides examples of regional ML-driven nowcasts integrated into a global system, and demonstrates verification that supports both product development and end-user utilization.
Successes and lessons learned from implementing the current nowcast system are also shaping a refined scope for nowcasting applications at Tomorrow.io. In addition to projecting near real-time precipitation estimates to their future location at short time horizons, nowcasting can augment near-real time precipitation estimation by enabling the synchronization of precipitation retrievals from multiple instruments that have differing latency and skill. This topic is becoming increasingly relevant as Tomorrow.io's constellation of satellite precipitation radars and microwave sounders comes online in 2024.

