In spite of jokes on weather forecast we have to acknowledge the tremendous progress that have been made over the last 20 years, thanks to a pervasive network of sensors (both on the ground and up in the sky) and the increased processing crunching capabilities, leveraging on machine learning and AI in general.
However, the forecast becomes less and less reliable as we narrow the area and/or extend the time horizon. 10 days forecast is more an indication than a real forecast. Besides, the actual weather in a very specific area remains difficult to predict.
In July, 2021, my daughter wedding was to take place in a rural areas on the hills in Southern Piedmont (it did!) and in the preceding week we were anxiously monitoring the weather forecast. It started well with a sunshine for the whole week (that made us very happy) but as we approached the date the forecast turned sour with rain for five days in a row starting 2 days before the event.
As we approached the wedding day the forecast indicated scattered showers in the afternoon. Our concern was on where those “scattered” showers would hit. Will the location be spared or not?
We actually found out on the spot. There were plenty of ugly black clouds up to the very moment we started and then (probably because my wife made an angry face to the sky) they moved away without a single drop falling on the bride and the groom. Later we found out that in a place just 5 km away there has been a downpour with significant flooding.
All of this to point out that weather forecast has still ways to go before we can trust it!
Of course a wedding day may not be sufficient motivation ot invest money in bettering the weather forecast but agriculture surely is. Being able to predict accurate weather -both time and location- can be very important to farmers, saving money and protecting their crops. An accurate determination of rain can save water (irrigation), something that could be very important in dry areas, knowing if temperature will reach freezing point during the night over the coming two weeks can lead to different decision of the use of fertiliser (if temperature gets too low, the fertiliser may kill the sprouts).
This is why it is so important the study presented by Microsoft researchers on the use of artificial intelligence to localise weather predictions. The team presented DeepMC, a software based on machine learning and AI to evaluate predictions from different sources (accessed via public APIs) with respect to local micro-climate detected via sensors in the field (see photo).
The study has shown that provided a good training of the software on the micro climate takes place the predictions can become much more accurate. In general, the need for intensive local training is a problem since it is difficult (costly) to apply the system in new locations. To ease the re-use the team is experimenting with GAN (Generative Adversarial Networks) to automate the re-training.
This kind of systems are likely to find extensive application as more and more IoTs are being deployed in agriculture. They can generate the data required by DeepMC to localise global weather predictions.