's radar satellites use machine learning to punch above their weight | Tech Crunch

Spread the love

Those of us lucky enough to have a window seat can predict the weather just by looking outside, but for the less privileged, weather forecasting and analysis is getting better and better. has released results from its first two radar satellites, which, thanks to machine learning, can compete with big, old-school forecasting technology on Earth and in orbit.

The company has been planning this mission since it was called Climacell back in 2021, and the results being released today (and soon to be formally presented at a climate science conference) show that their high-tech approach is working.

Weather forecasting is complicated for many reasons, but the interaction between high-powered but legacy hardware (such as radar networks and older satellites) and modern software is huge. That infrastructure is powerful and valuable, but it takes a lot of work on the compute side to improve their output – and at some point you start getting diminishing returns.

It's not just “will it rain this afternoon” but more complex and important predictions like which direction a tropical storm will move or how much rain has fallen in a given area due to a storm or drought. Such insights will become increasingly important as the climate changes.

Space is an obvious place to invest, but the weather infrastructure is huge and heavy. NASA's Global Precipitation Measurement Satellite, the gold standard for the field, launched in 2014, uses Ka (26-40 GHz) and Ku (12-18 GHz) band radars and weighs 3,850 kilograms.'s plan is to create a new space-based radar infrastructure with a modern twist. Its satellites are small (only 85 kilograms) and use Ka-band exclusively. The two satellites, Tomorrow R1 and R2, were launched in April and June last year and are now starting to show their quality after a long shake-out and testing.

In a series of experiments the company plans to publish in a journal later this year, Tomorrow claims that with just one radar band and a fraction of the mass, their satellites can produce results on par with NASA's GPM and ground-based systems. In a variety of tasks, the R1 and R2 satellites were able to make predictions and observations as accurate or even better and more accurate than the GPM, and their results were also consistent with ground radar data.

Data from R1 and R2 satellites are examples.

They achieve this by using a machine learning model, acting as two tools in one, as described by Chief Weather Officer Arun Chawla. It is trained on data from GPM's two radars, but by learning the relationship between the observation and the difference between the two radar signals, it can make a similar prediction using just one band. Their blog post reads:

The algorithm is trained with these dual-frequency-derived precipitation profiles but only using Ka-band observations as input. However, the complex relationship between reflectivity profile shape and precipitation is “learned” by the algorithm, and the full precipitation profile is retrieved even in cases where Ka-band reflectivity is completely degraded by heavy precipitation.

It would be a big win for if these results pan out and generalize to other climate models. But the idea is not to replace the US infrastructure – GPM and the ground radar network have been here for a long time and are invaluable assets. The real problem is that they cannot be easily duplicated to cover the rest of the world.

The company's hope is to have a network of satellites that can provide this level of detailed forecasting and analysis around the world. Their 8 planned production satellites will be larger – about 300 kg – and more capable.

“We are working to provide real-time precipitation data anywhere in the world, which we believe will be a game changer in the field of weather forecasting,” said Chawla. “In that regard we are working on accuracy, global availability and latency (measured as the time between the signal captured by the satellite and the data available to enter the products).”

They are also making the inevitable data play with more detailed orbital radar images to train their own and other systems. For that to work, they'll need a lot more data, though — and they plan to speed up their collection with more satellite launches this year.

Source link

Leave a Comment