Catherine Nakalembe

The problem is when you look deeply at any place, you start to see problems with a lot of the existing products.

They're not tailored to the underground contexts.

Sometimes it's simply that the data is just wrong.

A lot of the models are trained very well to predict for European or U.S.

agriculture.

In Europe, most of the farms have like a single crop.

They're really big and it's very easy to model.

In Kenya, in Uganda, in Rwanda, however, the fields are so tiny.

They have so many different crops in them, and farmers do things so differently.

It's like a tapestry.

With those images, fields are misrepresented, so there are places where there are no crops but are labeled as crops.

There are places where there should be crops and people that are completely missing.

When you go and try to assess something for the ministry and you use that as input, you're basically feeding them garbage.

You want to make sure that what you're feeding in is really good, and that actually requires work.

To train a model to understand that complexity, you need a lot of examples.

You have to go on the ground.

What we did is we used GoPros.

You wear a GoPro as you're driving on a motorcycle, or you can do it in the car.

And as you drive, we take pictures.

Basically, Google Street View, not for streets, but for crops, so the camera is actually facing towards the field.