AI models learn from past observations: where no measurements are taken, they make educated guesses. Down To Earth advocates funding instruments for poorer countries rather than providing them with free forecasts.
In a forum published on October 12, Down To Earth highlights a blind spot in the rise of weather by IA: these models do not learn physics, they reproduce the regularities of decades of observations. Where measurements are dense, they are excellent; Where they're missing, they're extrapolating. Kenyan researchers have found that GraphCast makes rain and temperature errors on Africa about twice as much as those on Europe, on a continent where 95% of agriculture depends on precipitation.
A very uneven map of instruments
Germany alone has more stations complying with international standards than all of Africa, where only one in ten surface stations are required to transmit data as required by global models. The article also recalls the August 26 glacial flood in Nepal, which killed more than 1,400 people: the last standing gauge never triggered an alert, designed to measure the height of the water and not its rate of rise. The budget of the Nepal Hydrological Service has been divided by more than two in three years.
The author proposes four measures: closing the funding of the United Nations SOFF mechanism ($400 million targeted by 2027, another quarter to be found); giving weather services a share of the revenue generated from commercial forecasts on their data; financing the maintenance of networks and not just their installation; and publishing the quality of forecasts by country income level each year.
Read the full article on the original source
Go to the source →