Cloud-Top Temperature in Deep-Learning-Based Precipitation Nowcasting

Abstract

Short-term precipitation nowcasting is important for operational decision-making and the timely response to high-impact rainfall. Radar observations alone provide limited information on cloud development and convective evolution, which can restrict the prediction of newly developing precipitation. This study investigates the use of cloud-top temperature (CTT) in deep-learning-based precipitation nowcasting, focusing on CTT–radar fusion strategies and rain-regime-conditioned uncertainty in NowcastNet over the Netherlands.

The relationship between CTT and rainfall intensity was first examined to determine whether distinct precipitation regimes could be identified. Based on their temporal relationship, precipitation events were grouped into two clusters representing different CTT–rainfall characteristics. The identified relationship provided the basis for investigating the incorporation of CTT into NowcastNet. Three levels of CTT conditioning were considered: input-level fusion using concatenated and stacked radar–CTT inputs, half fusion introducing a separate CTT encoder into the evolution network, and full fusion additionally conditioning the generative network through multi-scale feature modulation. Model performance was evaluated using aggregated verification metrics and event-based spatial analyses. Finally, the two CTT–rainfall clusters were used to define a regime-dependent affine post-processing method for adjusting the latent noise employed in ensemble generation.

Our results show that incorporating CTT generally enables NowcastNet to detect a larger proportion of observed precipitation than the radar-only baseline. However, the effect depends strongly on the fusion strategy: the input-level fusion models perform poorly, the half-fusion model achieves the best overall CSI, and the full-fusion model produces the highest POD but also more false alarms. Event-based spatial analyses further show that the additional precipitation produced by the CTT-conditioned models cannot be consistently associated with either weak or intense rainfall, but varies with rainfall structure, forecast lead time, and the individual event. For the probabilistic forecasts, the two CTT–rainfall clusters respond differently to the cluster-dependent affine adjustment. The adjustment reduces CRPS mainly for low-intensity rainfall and selected events and lead times, whereas at higher intensities and in several events it increases CRPS. These findings indicate that CTT can improve precipitation detection and support rainfall-regime-based uncertainty characterization, but its effectiveness depends on the fusion architecture, rainfall regime, and event characteristics.