AI Model Enhances West African Rainfall Prediction by Downscaling Climate Data

Researchers have developed an AI-driven approach to improve the resolution of daily precipitation data for West Africa. This study utilizes U-Net and GAN-based architectures to statistically downscale ERA5 atmospheric reanalysis fields, which typically have a coarse resolution, to the finer resolution of IMERG V07 satellite precipitation data. This enhancement is particularly crucial for West Africa, where rainfall is predominantly generated by convective systems that global models often fail to resolve accurately.

The project focuses on the monsoon seasons from June to October, spanning 2000 to 2024. Inputs to the AI model include various atmospheric parameters like geopotential, temperature, humidity, and wind at different pressure levels, alongside elevation and land fraction. The inclusion of ERA5 precipitation as an additional input further refines the model's performance in this complex meteorological region.

Initial deterministic U-Net models proved to be too dry and struggled to capture rainfall variance at regional scales. This limitation led the researchers to explore Generative Adversarial Networks (GANs), which showed more promise in accurately simulating precipitation patterns. The study also highlighted the critical role of mid-tropospheric moisture, specifically 500 hPa specific humidity, as a primary driver for deep convection, underscoring its importance in West African rainfall dynamics.

This advancement has significant implications for climate modeling and agricultural planning across West Africa. More accurate, higher-resolution rainfall predictions can aid in better water resource management, disaster preparedness for floods and droughts, and improved food security strategies in a region highly vulnerable to climate variability.