Center for Strategic Assessment and forecasts

Autonomous non-profit organization

Home / Politics and Geopolitics / Causes and consequences of abnormal heat in summer 2010 / Other
As Yandex has created a global forecast of precipitation using radar and satellites
Material posted: Publication date: 18-10-2018
The team Yandex.The weather there is a tradition to talk about new technologies on habré. We have described how machine learning methods helped to create a more accurate weather forecast, but also on how the neural network and data from the radars to anticipate rainfall.

Today we will tell to readers of Habra about the new technology, with which we were able to achieve prediction of precipitation with a minute precision even where data from radar just yet. And helped us in this satellite imagery.


Images of Meteosat-8 from space (source: EUMETSAT)

About radars and nowcast


Residents of the Central part of Russia were in luck, because there is installed a meteorological radar Roshydromet – modern tools that allow to map precipitation in a radius of 250 kilometres from the point of installation of the radar. While the spatial resolution of this map is 2x2 kilometer per pixel, and the interval between two successive snapshots in just 10 minutes.

imageThe left shows the look of the weather radar. (source: Actaris)

What this means for the average person, interested in the weather? In the territories covered by the radar umbrella, it is possible to report sediments with a precision of a residential neighborhood. Such data are highly precise traditional weather forecast, because I have information about what happened just a few dozen minutes ago. Learn more about the advantages of radar written in our previous article. Now look at their flaws.

The main one is the poor scalability due to the huge cost of such measurement tools and the complexity of the design. Recall now that the radar covered only part of European Russia and Barabinsk and Vladivostok. In addition, radar observations suffer from the buildings around, for example, tall buildings may obstruct the view in entire sectors, which affects the quality of the precipitation fields by radar imagery. An example of how can look averaged over several months of observation on each of our radars indicated in the illustration below.


Averaged data for measurements of the radar for a few months

As can be seen, severely affecting the quality of radar installed in Sheremetyevo and Vladivostok, and in Mineral waters falls the whole sector.

We see this in the complaints from our users. Also, in the case of low precipitation, radar can not see all 250 kilometers due to the curvature of the Earth's surface, which affects the ability to determine precipitation closer to the limit of visibility. In addition, the radar is out of order, sometimes for a long time, which affects our users who are accustomed to the rainfall map and warnings about the rain. Because of this, for example, unexpected "explosions" in the precipitation field, as it was recently in Vologda. Which, of course, becomes the occasion for a flurry of all sorts of memes.

Satellites as a source of information


Not to be tied to the radar, we decided to literally make a space product, what hints the title picture. In addition to radar measurements, there are similar approaches to the measurement of precipitation based on satellite imagery. There is a special group of meteorological satellites (in orbit there are about 30 units): polar orbiting, covering the Ground shots like the thread is wound on a clew, and geostationary satellites are at an altitude of about 36,000 kilometers from the Earth's surface and are rotated synchronously with the rotation of the Earth over the equator. The peculiarity of the orbits of the satellites of the second type allows to permanently "hang" over the same point and get the same pictures as in the beginning of the article. Satellite grouping geostazionari allows observations to cover almost the entire Earth, using for this purpose the European satellite (Meteosat), the U.S. (GOES) and the Japanese (Himawari). The products based on them have a spatial resolution of 0.5 to 3 kilometers per pixel, but there is a problem. The satellites hanging above the equator, so our latitude come to the edge of the image, causing the data suffer from geometric distortion.

Forecast for satellite images


The idea of using satellite imagery for weather forecasting is not new – the information has been used in traditional global weather prediction models. In addition, from satellite imagery extracted the useful information from monitoring volcanic ash and forest fires to search of phytoplankton. Of course, satellite imagery is used for monitoring precipitation and short-term forecasting. For example, there are standard algorithms for the detection of areas with rainfall – SAFNWC, but they work well only for the case with convective rainfall. According to our strict metrics that we use for radar nowcast these algorithms, unfortunately, are the outsiders. We wanted to make the product the quality is comparable with the current solution for radar data, but also globally scalable. For this we took the strengths from each approach and used the magic of machine learning.

Meteosat as a source of information


After the first experiments on standard algorithms, we decided that the satellite nowcast in the Weather to be. But here arose the first problem: where to get satellite images? In the digital age, it would seem that there should not be problems with obtaining satellite information. On the Internet you can find everything... except what you need. With the satellites, the situation is as follows: in order to quickly obtain images from space, you need to install the receiving station. Standard kit includes satellite dish, DVB-S tuner and the computer where pictures are stored and processed. Yes, the technology is completely analogous to the satellite TV, only dish a little more of human growth, and the tuner a little more expensive.

Thus, in our new DC in Vladimir, we got our own receiving station for satellite information. Data it come with the European series of satellites Meteosat second generation. To cover the Western part of the territory of Russia, we have chosen Meteosat-8, which hangs over the Indian ocean to the longitude of 41.5°. Snapshots are taken every 15 minutes – for them is a complete scan of the visible area of the satellite, and then the scan begins anew. Because of this, the pictures behind by 15 minutes from real time. The shooting takes place in 12 channels: 11 channels in the visible and infrared ranges with a resolution of 3x3 km and 1 channel in the visible spectrum with a resolution of 1x1 km (an example of shooting in various channels shown at right, source: EUMETSAT). Full the has a resolution of 3712х3712 pixels, or approximately 14 megapixels as the camera of a modern smartphone.

The is divided into 8 parts (stripes on the latitude), which sometimes affects the quality of data – the loss of one part may render useless all the.

The detection of precipitation


Since our first experiments showed that the quality of the product when using a traditional approach of suffering, we decided to use what has brought us success in the case of conventional newcustom. Came to the aid of the neural network. As input parameters we used information from the 11 channels of satellite imagery, and were trained on radar images, consolidated in a single field on a grid of 2x2 km. We used traditional approaches, which solve such tasks in computer vision. Until recently, two competing architectures based on ResNet-like (authored by irina-rud) and U-Net-like (authored illusionww) models.

ResNet is used in problems of image classification and can be very deep, the increase in the number of layers gives a stable increase in quality. However, this architecture has drawbacks in use – we have to apply the trained model at each point of our geographical grid. Alternatively, was selected as U-net – architecture of the convolutional neural network, which is typically used in problems of image segmentation. Originally developed for biomedical purposes for fast working with large images. With such architecture it is faster to test our hypotheses, we may also apply the trained model is not point-by-point, which significantly affects the speed of processing of satellite images. Below is a comparison on the obtained metrics for the two architectures. As we have managed to bring U-net model in quality to ResNet, but U-net makes it possible to process satellite images that we used as a production solution.


This chart shows the F1 measure is a standard metric in classification problems, which shows how our satellite precipitation differ from the radar. Under ideal match she should be unity. As the graph shows, the quality of detection of precipitation depends on the time of day, as the in the visible range is an important source of information.

The task is complicated by the fact that it is necessary not only to highlight the cloud in the picture, but also to determine whether the will of the rain. In experiments, we found that information from the 11 channels is not enough. To make a quality product, it is important to consider parameters such as the angle of the sun above the horizon, the height of the terrain, the data on the next clock from meteorological models, such as the moisture content of clouds, humidity in the atmosphere at various levels, etc.
In the result, the trained model allows to allocate precipitation with great accuracy. So the figure shows the overlay of precipitation by satellites and radar field. Here the color purple – satellites outside of the radars, crimson – the intersection of radar and satellites, the blue detect the satellites, but not detect radar and red – what they find is a radar. In the figure, a slight systematic shift to the North, due to the fact that we will detektoriem the rain from below the clouds, as do radars and from above, from space and angle. This problem will be fixed in the next release, which will affect the growth of accuracy.

Satellite nowcast


As the technology is fairly new, we decided not to abandon the radar, and leave them in places where our users already are accustomed to using newcustom. Here's the problem: how to display products, differing in method of measurement, in a single interface. We decided on a bold experiment – to show the radars and satellites in a common, familiar user-interface radar nowcast, expanding the area using satellites. This was a huge work, because we need, first, to coordinate the satellites and radars in time, and, second, to properly glue them to the limit of visibility of the radar.

To lead a 15-minute satellite images to 10-minute intervals, already familiar to our users, we use Optical Flow to generate the intermediate frames between consecutive satellite images. Optical flow or optic flow is the technology that is used in computer vision to determine the shifts between images. Using the two snapshots, we can build a field of displacement vectors of the image at each point is such which allows to obtain the following from the previous snapshot. With vectors of migration, we can obtain intermediate frames and to bring them to a single scale at a time with 10 minute intervals. This algorithm (through the efforts of bonext and ruguevara ) is used and for prediction 2 hours ahead with 10-minute resolution for satellite imagery, and radar are being counted the old-fashioned algorithm based on neural networks described in our article about radar nowcast. In the next update we plan to switch completely to the transfer of the entire precipitation field using neural network architecture.

When gluing dissimilar materials, of course, sometimes artifacts can appear, for example, as shown on the left. A prominent standard for radar in the limit of visibility they did not see rain, and in the triangle between them gets information from the satellite successfully detects rain. Thinking about a solution to the problem of bonding together two fields of data of different nature, we are reminded of the objective of inpainting. Nvidia in a recent article, Image Inpainting for Irregular Holes Using Partial Convolutions showshow neural networks can restore image details in the irregular masks. On Yet Another Conference described how Dmitry Ulyanov using inpainting restored mural. This is the same approach we plan to use in our case, and already have successful practices, which will soon go into production that will allow to properly take into account heterogeneous information from different sources about the fact of rain.

What's next?


Right now 100% of our users running nowcast built as radar and satellite measurements (due to space support imalion and working teams of backend and front). I hope our users who have been waiting nowcast in the city, they began to use it and to receive timely information on impending rainfall. At the moment, the area covered, the limited visible area of the satellite on the North and East (just East of the Ob river). In the South we have restricted the area to the lower part of Cyprus, and West to Switzerland. Now you can watch for approaching fronts on the approach to your town and observe the beautiful weather cranky. And the difference in the coating visible to the naked eye.



Thus, we have covered much of Russia, CIS and some tourist destinations. Of course, we have not forgotten about the Eastern part of our country – now we are working with the Japanese Himawari satellite, which hangs over Australia and will soon to please newcustom our Eastern borders.

And then – the global precipitation map for the whole world free from childhood diseases, with increased accuracy of detection of precipitation, a single algorithm transfer and the proper bonding of all data on precipitation.

Source: https://habr.com/company/yandex/blog/425517/

Tags: Russia , climate


RELATED MATERIALS: Politics and Geopolitics
Возрастное ограничение