[Paper Review] A review of radar-based nowcasting of precipitation and applicable machine learning techniques

Rachel Prudden, Samantha V. Adams+5 more|arXiv (Cornell University)|May 11, 2020

Meteorological Phenomena and SimulationsEarth and Planetary Sciences73 references63 citations

TL;DR

This paper surveys radar-based nowcasting methods for very short-term precipitation predictions, reviewing traditional advection/extrapolation approaches and emerging machine learning techniques, and discusses ways to integrate ML with atmospheric science insights.

ABSTRACT

A 'nowcast' is a type of weather forecast which makes predictions in the very short term, typically less than two hours - a period in which traditional numerical weather prediction can be limited. This type of weather prediction has important applications for commercial aviation; public and outdoor events; and the construction industry, power utilities, and ground transportation services that conduct much of their work outdoors. Importantly, one of the key needs for nowcasting systems is in the provision of accurate warnings of adverse weather events, such as heavy rain and flooding, for the protection of life and property in such situations. Typical nowcasting approaches are based on simple extrapolation models applied to observations, primarily rainfall radar. In this paper we review existing techniques to radar-based nowcasting from environmental sciences, as well as the statistical approaches that are applicable from the field of machine learning. Nowcasting continues to be an important component of operational systems and we believe new advances are possible with new partnerships between the environmental science and machine learning communities.

Motivation & Objective

• Explain the historical and current state of radar-based nowcasting using extrapolation and advection.
• Summarize probabilistic and stochastic methods that quantify uncertainty in nowcasting.
• Discuss machine learning approaches and their potential to relax Lagrangian persistence assumptions.
• Identify integration points between atmospheric science and ML to improve convective nowcasting.
• Highlight open challenges and future research pathways in radar-based precipitation nowcasting.

Proposed method

• Review traditional Eulerian/Lagrangian persistence and advection-based nowcasting models (optical flow, semi-Lagrangian schemes, cell-based methods).
• Present variational and hierarchical approaches for estimating advection fields and enforcing spatial consistency.
• Summarize probabilistic and stochastic extensions (neighborhood methods, scale decomposition) to account for uncertainty.
• Discuss strategies for predicting convective development via diagnostics and hybrid advection–NWP frameworks.
• Survey machine learning architectures for dense spatiotemporal prediction, including spatiotemporal convolutions and flow/deformation-based models.
• Address challenges in applying ML to nowcasting, such as interpretability and avoiding blurred/extreme-miss predictions.

Experimental results

Research questions

• RQ1 What are the main radar-based nowcasting approaches used historically and today?
• RQ2 How can probabilistic and stochastic methods quantify and manage uncertainty in nowcasting predictions?
• RQ3 What machine learning techniques show promise for dense spatiotemporal precipitation prediction in nowcasting?
• RQ4 How can ML be integrated with physical diagnostics to improve convective initiation and evolution predictions?
• RQ5 What are the key challenges (interpretability, regime change, multi-scale dynamics) limiting ML in radar-based nowcasting?

Key findings

• Advective, extrapolation-based nowcasting offers high resolution forecasts but struggles with rapid evolution and convective initiation.
• Probabilistic and stochastic extensions address uncertainty by modeling advection errors and scale-dependent evolution.
• Convective development can be improved by combining advection with diagnostics such as CAPE/CIN, boundary layer convergence, and environmental analyses.
• ML approaches (dense spatiotemporal prediction, flows, and deformation models) show potential but face issues with stability, extreme events, and interpretability.
• Hybrid frameworks that fuse radar extrapolation with NWP analyses or environmental diagnostics yield the most promising improvements.
• The field advocates closer collaboration between atmospheric science and ML to exploit diverse data sources and domain knowledge.