DeepMind Introduces AI-Based ‘Nowcasting’ System: A State-of-the-art Model That Predicts Rain Within The Next 1-2 hours

Weather plays an important role in our everyday lives. Among other weather conditions, rain influences our day-to-day decisions significantly. Weather forecasting has always been important to our communities and countries throughout history. Machine learning has found applications in almost every field, and weather forecasting is no longer an exception.

DeepMind’s recent study presents a cutting-edge model that forecasts rain (and other precipitation phenomena) within the next 1-2 hours. The researchers focus on value for decision-makers in collaboration with the Met Office, opening up new avenues for rain forecasting and pointing to the opportunities for AI in supporting the response to the challenges of decision-making in an ever-changing environment.

Short-term Weather Predictions

Numerical weather prediction (NWP) systems are used to make today’s weather predictions. NWPs provide critical planet-scale predictions several days ahead of time by solving physical equations. However, they have difficulty generating high-resolution forecasts with lead times of less than two hours. DeepMind presents Nowcasting that bridges the gap in performance during this critical time period.

Many industries, including water management, agriculture, aviation, emergency preparation, and outdoor events, could benefit from Nowcasting. High-resolution radar data, which monitors the amount of precipitation at ground level, is now available at a high-frequency thanks to advances in weather sensing (e.g., every 5 mins at 1 km resolution). With the availability of high-quality data, machine learning contributes to the development of Nowcasting.


Using Generative Models

Their proposed model focuses on making predictions about rain up to 2 hours ahead of the present time, capturing rainfall’s amount, timing, and location. They produce precise and realistic forecasts of future radar based on past radar using a technique known as generative modeling, which involves creating radar movies. This method reliably captures large-scale events while also generating many alternative rain scenarios (known as ensemble predictions), allowing us to investigate rainfall uncertainty. They used radar data from both the UK and the US.

This image shows a challenging event in April 2019 over the UK (Target is the observed radar). Our generative approach (DGMR) captures the circulation, intensity and structure better than an advection approach (PySTEPS), and does not blur like deterministic deep learning methods (UNet).

In April 2019, there was a challenging event in the United Kingdom. The generative technique (DGMR) was able to capture circulation, intensity, and structure better than an advection approach (PySTEPS). In addition, it predicted rainfall and motion in the northeast with more accuracy. Unlike deterministic deep learning systems, DGMR generates precise predictions (UNet).

The team was particularly interested in the models’ capacity to forecast medium to heavy rain events, which significantly impact people and the economy. They found statistically significant improvements in these regimes when compared to existing methods. 

Furthermore, they conducted a cognitive task assessment with more than 50 expert meteorologists from the Met Office, the United Kingdom’s national meteorological service. The experts rated this new approach as their first choice in 89 percent of cases compared to widely-used nowcasting methods, demonstrating its ability to provide insight to real-world decision-makers.

The team expresses that more work is needed to improve the accuracy of long-term predictions and accuracy on rare and intense events. They intend to provide more evaluation methods and plan to specialize these approaches for specific real-world applications.