A Regional Benchmark for Deep Learning-Based Hourly Precipitation Nowcasting in Latin America

Accurate short-term precipitation forecasting is critical for Latin America, but the region lacks a standardized framework to evaluate data-driven approaches due to the sparse coverage in the ground. This study introduces the Artificial Intelligence for Nowcasting Pilot Project (AINPP) Precipitation Benchmark (AINPP-PB-LATAM), providing curated datasets and a scalable, optimized training pipeline designed to accelerate deep learning development in high-performance computing environments. Beyond establishing a baseline using seven years of satellite-based data (2018–2024), the framework reduces engineering barriers, enabling researchers to focus on fine-tuning strategies to extract the full predictive capacity of models for regional specificities. As a demonstration use case, we trained and evaluated five deep learning architectures, AFNO, Inception-V4, ResNet-50, U-Net, and Xception, comparing them against Lagrangian Persistence and the operational AI Nowcast. The results reveal critical trade-offs: spectral methods like AFNO excel in continuous skill by capturing large-scale dependencies, while convolutional architectures offer robust categorical performance. However, pixel-wise optimization challenges persist, with systematic under-prediction of heavy rainfall. By providing open-access code and optimized baseline implementations for distributed computing, AINPP-PB-LATAM establishes a scalable foundation for collaborative research, facilitating the advancement of operational AI-based nowcasting and transferability assessments in data-scarce regions.