Tropical Cyclone Precipitation Nowcasting
Alawadhi, Omar Hussain Mohamednabi
Alawadhi, Omar Hussain Mohamednabi
Citations
Altmetric:
Supervisor
Department
Computer Vision
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
In current tropical cyclone precipitation nowcasting tasks, most existing approaches rely on numerical weather prediction (NWP) systems or deterministic deep learning frameworks; although these methods have achieved notable improvements, their forecasts often suffer from limitations in capturing highly localized and rapidly evolving rainfall structures, particularly under extreme weather conditions. These models are also sensitive to initial condition errors and may struggle to represent uncertainty effectively. On the other hand, recent advances in diffusion-based generative models have demonstrated strong capabilities in modeling complex data distributions and generating high-fidelity predictions. However, applying these models to precipitation nowcasting remains challenging due to computational complexity, difficulty in capturing fine-scale dynamics, and limited adaptation to tropical cyclone scenarios.
To address these limitations, I introduce a multimodal diffusion-based framework for tropical cyclone precipitation nowcasting that integrates observational and environmental information within a unified probabilistic forecasting model. Instead of directly predicting future rainfall intensities, the proposed approach models rainfall in the residual domain, enabling the framework to focus on temporal changes and better capture evolving recipitation patterns. The model leverages multiple data sources, including historical rainfall observations, gridded atmospheric variables, and scalar environmental descriptors, to provide rich contextual information for forecasting. The overall goal is to generate accurate and physically consistent short-term rainfall predictions while effectively capturing multiscale spatiotemporal dynamics.
Based on these insights, my contributions specifically include:
1. I reformulate tropical cyclone precipitation nowcasting as a residual-based diffusion problem, allowing the model to focus on temporal rainfall variations rather than absolute intensity values
2. I develop a multimodal conditioning framework that integrates rainfall observations, atmospheric variables, and scalar environmental features through a spatiotemporal diffusion backbone
3. I propose architectural enhancements, including gated linear attention for improved efficiency and wavelet-domain diffusion for enhanced multi-scale precipitation modeling
Through extensive experiments on a custom tropical cyclone precipitation dataset, I evaluate the effectiveness of the proposed approach using both continuous and event-based metrics, including MSE, MAE, CSI, HSS, and ETS. The results demonstrate that the proposed framework achieves competitive performance in both rainfall reconstruction and event detection across multiple thresholds, while effectively capturing the spatiotemporal evolution of precipitation systems. Overall, this work highlights the potential of diffusion-based generative models for weather forecasting and provides a flexible and effective framework for short-term precipitation prediction in complex and dynamic environments.
Citation
Alawadhi, Omar Hussain Mohamednabi, "Tropical Cyclone Precipitation Nowcasting," M.S. Thesis, Computer Vision, MBZUAI, 2026.
Source
Conference
Keywords
Nowcasting, Tropical Cyclone, Precipitation, Diffusion, Wavelet