Posted by Jason Hickey, Senior Software Engineer, Google Research
The weather can affect a person's daily routine in both normal and serious ways, and the precision of predictions can greatly influence how they deal with it. Weather forecasts can inform people if they need to take a different route to work, whether they have to reschedule the planned picnic for the weekend or even if they have to evacuate their homes due to an approaching storm. But making accurate weather forecasts can be particularly challenging for localized storms or events that develop at time scales per hour, such as thunderstorms.
In “ Machine Learning for Precipitation Nowcasting from Radar Images ” we present new research into the development of machine learning models for precipitation forecasts that address this challenge by making highly localized “physics-free” predictions that apply to the near future. An important advantage of machine learning is that the inference is computationally inexpensive, given an already trained model, allowing for forecasts that are almost immediate and in the original high resolution of the input data. This precipitation now, which focuses on 0-6 hour predictions, can generate predictions with a resolution of 1 km with a total latency of only 5-10 minutes, including delays in data collection, which outperform traditional models, even at these early stages of development.
Beyond traditional weather forecasts
Weather agencies around the world have extensive monitoring facilities. For example, Doppler radar measures precipitate real-time weather satellites, multispectral imaging, ground stations measure wind and precipitation directly, etc. The figure below, compared artificially colored composition radar imaging precipitation via continental US clouds exposed by geosynchronous satellites, illustrates the need for weather information with multiple sources. The existence of rain is related to, but not perfectly correlated with, the existence of clouds, so extracting precipitation from satellite images alone is a challenge.
Unfortunately, not all of these measurements are present everywhere in the world. For example, radar data is largely from ground stations and is generally not available through the oceans. Furthermore, coverage varies geographically, and some locations may have poor radar coverage even when they have good satellite coverage.
Yet there are so many observation data in so many different varieties that prediction systems have difficulty recording it all. In the US, remote sensing data collected by the National Oceanic and Atmospheric Administration (NOAA) now reaches 100 terabytes per day. NOAA uses this data to feed the huge weather forecast engines that run on super computers to provide 1- to 10-day global forecasts. These engines have been developed during the last half century and are based on numerical methods that directly simulate physical processes, including atmospheric dynamics and numerous effects such as thermal radiation, vegetation, lake and ocean effects and more.
The availability of calculation sources, however, limits the power of numerical weather forecast in various ways. For example, arithmetic requirements limit the spatial resolution to approximately 5 kilometers, which is not sufficient to solve weather patterns in urban areas and agricultural land. Numeric methods can also take several hours. If the 6 hours to calculate a prediction is that only allow 3-4 runs per day, resulting in predictions based on 6+ hours old data, some knowledge of some bad turn now limited .daarentegen with Nowcasting isname useful for direct decisions from traffic routing and logistics to evacuation planning.
As a typical example of the type of predictions that our system can generate, consider the radar-to-radar prediction problem: give a series of radar images from the past hour, predict what the radar image will be N hours from now, where N usually ranges from 0-6 hour. Because radar data is organized in images, we can put this prediction as a problem with computer vision, where the meteorological evolution is derived from the series of input images. On these short time scales, evolution is dominated by two physical processes: advection for cloud movement and convection for cloud formation, both of which are significantly influenced by local terrain and geography.
Above (from left to right): the first three panels show radar images from 60 minutes, 30 minutes and 0 minutes earlier, the point at which a forecast is desired. The panel on the right shows the radar image 60 minutes later, that is, the basic truth for a nowcasting forecast. Bottom left: for comparison, a vector field induced by applying an optical current (OR) algorithm to model advection on the data of the first three panels above. Optical flow is a computer vision method developed in the 1940s that is often used to predict weather evolution in the short term. Bottom right: an example OR prediction. Note that it follows the movement of the precipitation in the lower left corner well, but does not take into account the rotting force of the storm.
We use a data-driven approach to physics-free, meaning that the neural network will learn to approach the atmospheric physics only the training examples, not by prior knowledge of how the atmosphere really works. We treat weather forecast as an image-to-image translation problem and use the current state of affairs in image analysis: convolutional neural networks (CNNs).
CNNs are usually composed of a linear series of layers , where each layer is a set of operations that convert a certain input image into a new output image. Often a layer changes the number of channels and the overall resolution of the image that is given, in addition to convolving the image with a set of convolutional filters . These filters are small images themselves (for us these are usually only 3×3 or 5×5). Filters drive much of the power of CNNs and result in operations such as edge detection, identification of meaningful patterns, etc.
A particularly effective type of CNN is the U-Net . U-nets have a series of layers that are arranged in a single coding phase , where layers iteratively lower the resolution of the images passing through, and then a decoding phase in which the low-dimensional representations of the image created by the coding phase are extended back to higher resolutions. The following image shows all layers in our specific U-net.
(A) The general structure of our U-NET. Blue boxes correspond to basic CNN layers. Pink boxes correspond to down-sample layers. Green boxes correspond to up-sample layers. Solid lines indicate input connections between layers. Dotted lines indicate long skip connections that cross the coding and decoding phases of the U-NET. Dotted lines indicate short skip connections for individual layers. (B) The operations within our base layer. (C) The operations within our down-sample layers. (D) The operations within our up-sample layers.
The input for the U-Net is an image that contains one channel for each multispectral satellite image in the series of observations from the last hour. For example, if 10 satellite images were collected in the last hour and each of those multispectral images was taken at 10 different wavelengths, the image input for our model would be a 100-channel image. For radar-to-radar predictions, the input is a series of 30 radar observations during the past hour, with a distance of 2 minutes, and the output contains the forecast for N hours from now. For our first work in the US, we trained a network of historical observations on the continental US from 2017 to 2019. The data is split into four-week periods, with the first three weeks of each period used for training and the fourth week is used for evaluation.
We compare our results with three commonly used models. First, the high resolution Rapid Refresh numeric prediction with (HRRR) from NOAA. HRRR actually contains predictions for many different weather quantities. We compared our results with their total accumulated precipitation forecast of 1 hour , because that was their precipitation forecast of the highest quality. Secondly, an optical flow (OR) algorithm that attempts to track moving objects through a series of images. This latter approach is often applied to weather forecasting, although it makes the assumption that the total amount of rain is constant over large areas during the forecast time – an assumption that is clearly violated. Third, the so-called persistence model , is the trivial model in which it is assumed that every location will rain at the same speed in the future as it does now , that is, the precipitation pattern does not change. That may seem too simplistic a model to compare with, but it is common given the difficulty of weather forecast.
A visualization of predictions over the course of about a day. Left: the HRRR prediction of 1 hour at the top of every hour, the limit for how often HRRR gives predictions. Middle: the basic truth, that is, what we are trying to predict. Right: the predictions of our model. Our predictions are every 2 minutes (shown here every 15 minutes) at approximately 10 times the spatial resolution made by HRRR. Note that we record the general movement and the general form of the storm.
We use precision and recall (PR) graphs to compare the models. Because we have direct access to our own classifier, we offer a full PR curve (shown as the blue line in the image below). However, because we do not have direct access to the HRRR model, and because neither the persistence model nor OR has the ability to alternate and recall precision, those models are only represented by individual points. As can be seen, the quality of our neural network prediction surpasses all three of these models (because the blue line is above all the results of the other model). However, it is important to note that the HRRR model begins to outperform our current results when the forecast horizon reaches about 5 to 6 hours.
Precision and recall curves (PR) that compare our results (solid blue line) with: optical current (OR), the persistence model and the HRRR prediction of 1 hour. Because we do not have direct access to their classifications, we cannot provide a full PR curve for their results. Left: Forecasts for light rain. Right: Predictions for moderate rain.
One of the benefits of the ML method is that predictions are effectively instantaneous, meaning that our predictions are based on new data, while HRRR is hampered by latency of 1-3 hours. This leads to better predictions for computer vision methods for very short-term forecasts. The numerical model used in HRRR, on the other hand, can make better long-term predictions, partly because it uses a fully physical 3D model – cloud formation is more difficult to observe from 2D images, and it is more difficult for ML methods to learn convective processes . It is possible that the combination of these two systems, our ML model for rapid forecasting and HRRR for long-term predictions, generally produces better results, an idea that is central to our future work. We also look at applying ML directly to 3D observations. Regardless of the prediction, it is an important tool for real-time planning, facilitating decisions and improving lives.
thanks to Carla Bromberg, Shreya Agrawal, Cenk Gazen, John Burge, Luke Barrington, Aaron Bell, Anand Babu, Stephan Hoyer, Lak Lakshmanan, Brian Williams, Casper Sønderby, Nal Kalchbrenner, Avital Oliver, Tim Salimans, Mostafa Dehghani, Jonathan Heek, Lasse Espeholt, Sella Nevo, Avinatan Hassidim.
source: https://ai.googleblog.com/Tags: AI, google, Kunstmatige intelligentie, neerslag