What is it about?
This study presents a new open benchmark to help researchers improve short-term rainfall forecasts for Latin America. Reliable forecasts for the next few hours are essential for warning people about floods, landslides, and severe storms, but many parts of Latin America do not have enough ground-based weather radars or rain gauges. To address this problem, the study uses satellite rainfall data and artificial intelligence models to predict hourly rainfall up to six hours ahead. The benchmark provides organized datasets, standard testing procedures, and ready-to-use computer code so that different forecasting methods can be compared fairly. The study tests several deep learning models and compares them with existing forecasting approaches. The results show that AI models can learn useful rainfall patterns across the region, but predicting intense rainfall remains difficult. By making the data, methods, and baselines easier to use, this work supports more transparent and collaborative development of AI-based rainfall forecasting systems for Latin America.
Featured Image
Why is it important?
This work is important because Latin America is highly vulnerable to extreme rainfall, floods, and landslides, but many areas still lack dense radar and rain gauge networks for reliable short-term weather forecasting. By using satellite rainfall data and artificial intelligence, this study provides a practical way to evaluate and improve rainfall forecasts in regions where ground observations are limited. What makes this work unique is that it introduces an open and reproducible benchmark specifically designed for hourly precipitation nowcasting across Latin America. It brings together curated datasets, standardized evaluation procedures, baseline models, and scalable computing tools, allowing researchers and forecasting centers to compare methods fairly and build on shared results. This is timely because AI-based weather forecasting is advancing rapidly, but many existing systems are developed and tested mainly in regions with better observational infrastructure. The benchmark can help accelerate the development of more accurate, transparent, and regionally adapted forecasting systems. In the long term, this may support better early warnings, improved disaster preparedness, and stronger hydrometeorological services across Latin America.
Perspectives
This publication represents an important step toward making artificial intelligence for weather forecasting more useful in regions where observational data are still limited. Much of the recent progress in AI-based nowcasting has been driven by datasets and infrastructure from countries with dense radar networks, but Latin America has different challenges, including sparse ground observations, complex topography, tropical convection, and strong regional variability. This article was an opportunity to help build a more open and reproducible foundation for research in this context. I see AINPP-PB-LATAM not only as a benchmark, but also as an invitation for collaboration among researchers, operational forecasting centers, and institutions interested in improving short-term rainfall forecasts for Latin America. I hope this work encourages more studies that adapt AI models to regional needs, rather than simply transferring solutions developed elsewhere. I believe that open datasets, transparent evaluation protocols, and shared computational tools are essential if AI is to have a real impact on early warning systems and disaster risk reduction. My hope is that this publication helps accelerate that process and supports the development of more reliable precipitation nowcasting systems for communities that are highly exposed to extreme rainfall, floods, and landslides.
Adriano Almeida
National Institue for Space Research
Read the Original
This page is a summary of: A Regional Benchmark for Deep Learning-Based Hourly Precipitation Nowcasting in Latin America, IEEE Access, January 2026, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/access.2026.3670767.
You can read the full text:
Contributors
The following have contributed to this page
