Catherine Nakalembe

So I could try to map it really well.

I could get all the information directly on the ground.

But I'd still do a pretty bad job because the model has learned about other things that have nothing to do with our reality.

So there's one side of it.

One side of it is in order to get useful information from satellite data, we build models.

You have to train the model using some existing example.

And then the model goes through all the satellite data and tries to find that thing that you were looking for.

And that's how we would create like a crop type map.

Today, most models do very well for farmland in the United States.

The fields are really big.

They're homogeneous, so there's like one crop over a really large area.

There's been a lot of investment in collecting data to have those examples.

So when I build a model and train to predict what is growing in Kansas, it's easy.

But for me to produce a map of where maize is in Kenya, a whole different story, because I don't have those examples yet.

It's much harder to collect the data because there's no default data collection for this purpose.

The other side of it is that the satellites that we have access to, where we have open data that allows you to scale for all of Kenya, the resolution doesn't fit the small Kenyan fields.

And so if I use the same amount of labels but trained on the same data set for Kenya and the U.S.,

It wouldn't work because these fields are so much smaller and much more complex.

So the products we have are usually not relevant at the farmer's scale.

They're relevant at a larger scale.