Building effective models for predicting the weather has been an important area of focus for researchers and tech companies. Knowing precipitation predictions beforehand can have a critical role for industries that rely on rainfall and the general public. To take a step forward in this direction through deep learning algorithms, Google AI has released the Meteorological Neural Network 2 (MetNet-2) for 12-hour precipitation forecasting.
The tech giant says that deep learning methods provide a different perspective for forecasting by learning to predict directly from observed data. The computations are faster than physics-based techniques. MetNet-2’s predecessor, MetNet, released last year, provided eight-hour forecasting. Recently, DeepMind has also developed a deep-learning tool called DGMR (Deep Generative Models of Rain) for forecasting rain up to two hours ahead of time.
In the paper titled, “Skillful Twelve Hour Precipitation Forecasts Using Large Context Neural Networks“, Google focuses on longer forecasts with spatial resolution kept at 1 km and a time resolution of 2 minutes. It says that MetNet-2 outperforms the HREF ensemble model (physics-based) for weather forecasts up to 12 hours ahead.
Image: Google
The research paper says that MetNet-2 uses input observations from a 2048 km ×2048 km region and uses novel neural network architectural elements to deduce the larger context. MetNet-2 is trained over a 7000 km × 2500 km region of the Continental United States (CONUS).
Challenges
The Google AI team says that a big challenge to solve was capturing a sufficient amount of spatial context in the input images. To solve this, it used model parallelism. Here, the model is distributed across 128 cores of a Cloud TPU v3-128. MetNet-2 replaces the attentional layers of MetNet with computationally more efficient convolutional layers. MetNet-2 uses dilated receptive fields, whose size doubles layer after layer, to connect points in the input far apart from one another.
Architecture
The paper “Skillful Twelve Hour Precipitation Forecasts Using Large Context Neural Networks” puts forward how the model is built to forecast precipitation effectively.
The report points out that MetNet-2 and NWP models collect empirical observations to bring out the initial state of the atmosphere to form the basis for their forecasts. The sources for these observations are sensors located in weather stations, satellites, and aeroplanes. The researchers use two types of precipitation measures- instantaneous precipitation and hourly cumulative precipitation. They are both provided by the Multi-Radar Multi-System (MRMS).
Ground Truth
The MetNet-2 learns to predict both the instantaneous measure and the hourly cumulated measure at 2 minute and 60-minute intervals, respectively, the research says. It adds that the precipitation measures range from a rate of 0 mm/hr to 102.4 mm/hr.
Postprocess and Hybrid
The team considers other training modes for MetNet-2 that make use of outcomes of NWPs. The Postprocess takes NWP’s forecast as input and learns to map NWP’s forecast as closely as possible to the ground truth. The Hybrid learns to bring out information from available inputs, like from NWP, radar and assimilation inputs used in the default MetNet-2.
How does it perform?
Image: Google
The research says that Numerical Weather Prediction (NWP) model High Resolution Rapid Refresh (HRRR) is already operational at CONUS and makes hourly forecasts at a 3 km × 3 km resolution. Different evaluation methods have pointed out that MetNet-2 outperforms the NWP model at predicting precipitation up to 12 hours of lead time without physics-based atmospheric forecasts. A High Resolution Ensemble Forecast (HREF) is also present in the CONUS region at a 3 km × 3 km resolution. It outperforms that as well up to 12 hours for low and high levels of precipitation.
