Explainable and Efficient Deep Learning for Rainfall Nowcasting

: Feature-Aware Evaluation and Temporal–Architectural Optimization

Abstract

ABSTRACT
Extreme rainfall events pose increasing threats to human safety, infrastructure, and the economy, especially under accelerating climate change. Short-term, high-resolution precipitation forecasting, known as rainfall nowcasting, is critical for flood preparedness, urban infrastructure resilience, and real-time disaster response. This thesis advances the field of deep learning-based rainfall nowcasting through three major contributions: (i) proposing a rainfall feature-aware evaluation framework, (ii) introducing an explainable framework for optimizing input–output configurations, and (iii) introducing an efficiency-oriented residual 3D convolutional framework that jointly learns spatial–temporal features while remaining deployable in data- and resource-constrained
settings. First, the thesis presents a novel feature-aware evaluation framework that overcomes the limitations of conventional benchmarking, which often masks the influence of specific rainfall features. Model skill is assessed for four features: event severity, motion velocity, rotational structure, and rainfall type, using clustered real-world events and controlled synthetic scenarios. In real-world cases, deep learning models (DLMs) generally outperform optical flow methods (OFMs), particularly for stratiform rainfall, but all models degrade for convective events. DLMs are affected by cumulative error and blurring over time, while OFMs cannot capture intensity changes. Synthetic tests isolating translational velocity show that DLM accuracy declines with increasing speed even without intensity change, confirming velocity as a key factor in their
performance. Second, the thesis addresses the lack of systematic optimisation in temporal input–output configurations. A novel unified, model-agnostic framework evaluates 24 combinations of historical input lengths and forecast horizons using a U-Net-based architecture. Optimal configurations vary with forecast length: three input frames yield the best results for single-frame outputs, while four inputs are optimal for longer forecasts. Explainable AI methods further quantify temporal redundancy and identify the most informative frames, supporting dynamically optimised configurations that improve accuracy and interpretability. Third, architectural trade-offs are examined by comparing standard 2D and 3D CNNs and exploring efficient 3D variants within the proposed residual 3D convolutional framework. While
2D CNNs capture spatial patterns, 3D CNNs model joint spatial–temporal dynamics and achieve better performance on complex sequences. Benchmarks of standard and efficient 3D U-Nets assess predictive skill, parameter count, memory use, and inference time. Several efficient designs match standard 3D CNN accuracy with reduced computational demands, enabling real-time and
resource-flexible deployment, especially in data-scarce conditions.
Overall, this thesis delivers a unified and interpretable methodology for designing and evaluating deep learning-based rainfall nowcasting systems, offering insights that support the development of robust, efficient, and operationally viable frameworks for evolving hydrometeorological risks.

Date of Award	9 Dec 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Dawei Han (Supervisor), Miguel A Rico-Ramirez (Supervisor) & Weiru Liu (Supervisor)