Ground Truth – Why is it needed?

HSAT has built the world’s largest database of crops. This data helps understand food security and the impact of climate change.

This data is used by agriculture firms, insurance companies and traders.

But why?

Satellite data has the ability to “see” crops around the world, however the “view” is not very good.  This is because the satellite data used for crop predictions has 10 meter resolution, this means that its not possible to identify crops simply by looking at the image. This means that the smallest area seen is 10mx10m – or several times bigger than a car.  This is too low resolution to identify a crop from a picture.

As a comparison the satellite data seen in Google Maps is 0.2 meter resolution

The Challenge - Low Resolution

The Solution – Machine Learning

To resolve this issue machine learning is used to build models and identify the signal from different crops

Crop Signals

These signals are created using different spectral bands (e.g infra-red) examined across time

This data then results in probability maps of every field across an entire country. These maps are then used to predict the area

Probability of Crops

Machine Learning – Needs Ground Truth

These signals are all built with machine learning and machine learning requires “ground truth”

The ground truth is used to teach of “train” the system.  The satellite cannot recognise what sugar or wheat is, so we build a model using thousands of examples of sugar, wheat, rice, etc.  The machine then learns from this data and then generates the “machine learning” model that predicts the crops

Machine learning is very powerful but requires ground truth.