Foundation Model Adaptation for Precipitation Nowcasting

Abstract

Accurate prediction of extreme precipitation is crucial for effective adaptation to the impact of climate change on precipitation intensity. Within the nowcasting time range of 0 to 3 hours, radar observations hold the most potential for informing predictions. Here we present our efforts to adapt recently published state-of-the-art machine learning nowcasting models to the Canadian context. Most machine learning tools capture only general structure and miss important details for longer lead times, often resulting in blurry images that fail to represent higher precipitation rates accurately. To address these issues, Zhang et al. (Nature 2023) presented the NowcastNet model for extreme precipitation that captures advective features at a scale of 20 km, and convective features at a 1–2km scale. The NowcastNet model only requires radar data as input and produces state-of-the-art results up to 3 hours ahead. In this work, we present a precipitation nowcasting system that combines the pre-trained Prithvi-WxC foundation model with the NowcastNet model. The pre-trained model processes the physical-evolution scheme whereas NowcastNet generative structure produces fine-grain details. The new structure is trained using a WGAN-GP mechanism with two time-scale update rule (TTUR) that makes training more effective and faster. To improve even further training time, we apply transfer learning from the original model to different components of the new model. New components can produce ensemble members faster with lower memory footprint. The model generates multiple ensemble members of precipitation maps up to 3 hours ahead with 10-minute time step. We first trained the modified NowcastNet model using ECCC radar composite data (at 2-km resolution) from 2022 and validated it against data from 2023. We then fed and fine-tuned the foundation model with hourly outputs from ECCC's Global Deterministic Prediction System (GDPS) at 15 km resolution to replace the evolution network of NowcastNet. We evaluated our models' performance for both the deterministic and probabilistic versions using metrics such as the CSI, CRPS, rank histograms, FSS and the Fréchet Inception Distance (FID), comparing with classical advection-based nowcasting and other machine learning approaches.