The application of science and technology to forecast atmospheric conditions for a specific location and time is known as weather forecasting. People have tried to guess the weather informally for centuries and methodically since the nineteenth century. Presently, traditional physics-based techniques, powered by the world’s largest supercomputers, are used to forecast weather. But high computational needs constrain such methods, and they are also susceptible to approximations of the physical laws on which they are based.
Deep learning can offer a new approach to computing forecasts. Deep learning has been used to solve a variety of crucial problems, including cancer prevention and increasing accessibility already. Therefore, the use of deep learning models to anticipate weather can be helpful to humans on a daily basis. Deep learning models learn to predict weather patterns directly from observable data rather than applying explicit physical laws and can calculate predictions quicker than physics-based techniques. These methods have the potential to boost the frequency, scope, and accuracy of projected forecasts.
Deep learning algorithms have shown great promise in weather forecasting for nowcasting, predicting weather up to 2-6 hours ahead. Previous research has concentrated on using direct neural network models for weather data, extending neural forecasts from 0 to 8 hours using the MetNet architecture, generating radar data continuations for up to 90 minutes ahead, and interpreting the weather information learned by these neural networks. Deep learning, on the other hand, has the potential to improve longer-term forecasts.
Google AI has released the Meteorological Neural Network 2 (MetNet-2) for 12-hour precipitation forecasting. MetNet-2 outperforms its predecessor, MetNet, by a significant margin. MetNet-2 surpasses the state-of-the-art HREF ensemble model for weather forecasts up to 12 hours ahead of when compared to physics-based models.
MetNet-2 and other neural weather models relate measurements of the Earth to the chance of weather occurrences like rain over a city in the afternoon, wind gusts of 20 knots, or a sunny day ahead. By integrating a system’s inputs and outputs directly, end-to-end deep learning has the ability to both streamline and improve quality. MetNet-2 was designed with this in mind, and it tries to reduce both the complexity and the overall number of stages required to create a forecast.
Radar and satellite photos, which are also used in MetNet, are among the inputs of MetNet-2. MetNet-2 employs the pre-processed beginning state used in physical models as a proxy for this additional meteorological information to capture a more comprehensive snapshot of the atmosphere with information such as temperature, humidity, and wind direction – crucial for longer predictions of up to 12 hours.
Capturing enough spatial information in the input photos is one of the primary issues that MetNet-2 must overcome in order to create 12-hour forecasts. The team included 64 km of background in every direction for each additional forecast hour at the input. This leads to a 20482 km2 input context, which is four times larger than MetNet’s. MetNet-2 updates MetNet’s attentional layers with computationally more efficient convolutional layers due to the size of the input context.
The forecasts were evaluated using well-established metrics like the Continuous Ranked Probability Score, which measures the amount of a model’s forecasts’ probabilistic error in comparison to the ground truth observations. Despite not using any physics-based calculations, MetNet-2 is able to surpass HREF for both low and high levels of precipitation up to 12 hours in the future.
Because MetNet-2 does not rely on hand-crafted physical equations, its effectiveness begs the question: What kinds of physical weather relationships does it learn from the data during training?
The most striking conclusion is that MetNet-2 appears to mimic the physics described by Quasi-Geostrophic Theory, which is used to approximate large-scale weather occurrences. At the scale of a normal high- or low-pressure system (i.e., the synoptic-scale), MetNet-2 was able to detect changes in atmospheric forces that lead to advantageous precipitation conditions, which is a critical component of the theory.
MetNet-2 is a step toward enabling a new weather forecasting modeling paradigm that does not rely on hand-coding the physics of weather occurrences. Many obstacles remain on the way to fully realizing this goal, including incorporating more raw data about the atmosphere directly.
Paper: https://arxiv.org/pdf/2111.07470.pdf
Reference: https://ai.googleblog.com/2021/11/metnet-2-deep-learning-for-12-hour.html
Suggested
Credit: Source link
Comments are closed.