Greetings friends readers, in this opportunity back here I bring you something that I found very interesting. Particularly because in my free time from time to time or when there is some important date, I resort to it, and that is to see the weather report of the day and week.
As I learned in general geography, a couple of years ago, the weather is predicted but not the climate, and although it is possible to determine the weather in a given period, it is usually complicated to get it right 100% (complicated eye, but not impossible). All this thanks to the changes that can occur, which affect the weather during a certain interval (assuming that the weather report is for one day, for example).
It should be noted that it was in 340 BC when the Greek philosopher Aristotle, began with that of "meteorology", in which he tried to make weather forecasts. According to Aristotle, meteorology was nothing more than "all the effects that can be called common to air and water, the forms and parts of the earth and their effects. And it certainly wasn't that far removed from reality.
Since much has happened since then and today, various tools are used to achieve the goal, such as ground stations, Doppler radars, multispectral imaging, etc.
Now, as we know from some time ago, AI has started to be used for everything, from systems for surveillance, autonomous cars, making calculations, etc. And well the company of the big G, decided to develop an AI to make weather reports, in a "different" way so to speak.
As we know we have lived and will continue to live in a larger scale with a very drastic climate change, which is why we have seen droughts and rains in areas that usually do not happen on certain dates for example.
So from Google contribute that a system of immediate forecast of high resolution, is an essentially necessary tool, to adapt to the climate change, particularly even more in extreme climates.
Here we have three boxes, in which the
is the prediction made by the HRRR, in the
is what really happened on that date and finally on the
the predictions made by the Google model.
The proposed model is an automatic learning one, which is designed to face this kind of events, emphasizing that these predictions are highly localized and that they are applied to an immediate future. The advantage of this is that at present there are computer models capable of making predictions, but in a certain way they are somewhat slow. The company of the big G, focuses on immediate forecasts and its AI can make use in part of the inference (that is computationally economic) to generate such forecasts of the following 6 hours in just about 5/10 minutes, including within this what would be the data collection.
It is also commented that rains are related in some way to clouds, but it is emphasized that they are not perfectly correlated. It is also explained that radar data are usually derived from land stations and this would be a limitation in the case of the oceans. This is without taking into account the geographical location of the stations, which affects their coverage.
Therefore, the Google option opts for an autonomous approach of physics based on data. The simple translation to all this is that the neural network will learn to approximate atmospheric physics only from training examples, not incorporating a priori knowledge of how the atmosphere really works... And given a sequence of radar images for the last hour, it predicts what the radar image will be in N hours, from now on, where N will vary in a range of 0 to 6 hours. Since the radar data are organized in images, we can pose this prediction as a computer vision problem, deducing the meteorological evolution from the sequence of input images.
The type of neural network used for all this is called U-Net, which is a convolutional neural network that was developed at the computer department of the University of Freiburg in Germany. It was originally set up for the segmentation of biomedical images.
Flowchart of the process using the neural network (U-Net)
The U-Net entry is an image that contains one channel per multispectral satellite image. So if, for example, 10 models of satellite images taken with the same number of different wavelengths are added in one hour, the model input is a resulting image of one hundred channels.
For this AI, Google used the historical data from the USA between 2017 and 2019 for the training of this network. These were divided into four-week periods, with the first three weeks focused on training and the remainder on evaluation. Once these results were obtained, they were compared with traditional forecasts from NOAA's HRRR (High Resolution Rapid Refresh), an optical flow algorithm and a persistence model.
It is commented that optical flow algorithms have a problem, since they are focused on tracking moving objects along an image sequence, so in the case of rainfall, it can be assumed that rainfall over large areas is constant during the time of the forecast. The persistence model, on the other hand, may seem very simple (since it consists of analyzing today's environmental conditions in order to anticipate tomorrow's), but it is commonly used for these practices given the difficulty that climate prediction represents.
The final results were placed in a precision and recovery graph. Those obtained by Google, only the blue line, on the other hand those of the HRRR and the persistence model, as well as the optical flow algorithm, do not have the capacity to exchange precision and recovery, their results will be expressed in points.
Therefore, it can be noted that the forecast obtained by Google is higher (than that of the three models mentioned above) under a certain average number of hours, but it should be noted that the HRRR model tends to be more effective when the forecast covers approximately 5/6 hours.
In the end, it can be stated that the Google model is of instantaneous predictions, while the HRRR has a computational latency of between 1 and 3 hours. Therefore, Google's model would be effective for very short-term predictions, while the HRRR is more effective for long-term periods.