Building a custom crop type detection for Estonia from scratch

Company description

eAgronom is an Estonian AgTech company providing a Carbon Program, Sustainable farming loans as well as a Farm Management solution. Although focused on Central and Eastern Europe, they provide their services all over Europe with teams in Estonia, Latvia, Poland and Spain. 

They provide support to thousands of farmers over more than 1M hectares in optimizing their activity to help them sequestrate as much carbon in the soil as possible. Part of this activity is to Measure, Report and Verify (MRV) field activity, and this is why they need SpaceSense.

‍

The challenge

A machine learning-based crop detection algorithm relies on historical training data to learn how to differentiate between crops. The location where this data comes from will significantly influence the performance of model. If the model is trained on Ukrainian fields, it will not be very effective in Portugal for example. And the off-the-shelf solutions offered to eAgronom reflected this issue, and where not effective for Estonia, the first zone that eAgronom wanted to monitor. This is why they decided to use their internal data to build a custom model.

There comes the second challenge: They do not have internal expertise on machine learning with satellite imagery, and limited resources they can allocate to it. This is why they selected SpaceSense to build their solution.

The solution

eAgronom used SpaceSense’s crop type “solution template”. These “solution templates” are pre-written scripts designed to train a specific models, in this case a model to detect crop types. The user only has to integrate their custom training data, and the template does the rest automatically, and generates a trained custom model. The user can then tweak several parameters to improve performance if it is necessary.

‍

The template has the following steps:

1. Extract the fields boundaries and the growing periods of the proprietary data you want to use for training - 3 days
2. Customize the “Solution template” for crop detection by incorporating the proprietary training data, and validating all training parameters - 2 day
3. Train and test the crop type detection model. There might be several iterative loops where you change some parameters to improve the accuracy of the model - 5 days
4. Run the model over the region of interest to get results as often as needed - 1 day

The result

Thanks to SpaceSense, eAgronom was able to build a crop type detection model with an accuracy superior to 84% for 12 different crops in less than a month. Without any previous machine learning experience.

Not only does it now gives them a competitive advantage in Estonia, but they also saved 80% of their cost compared to building it in-house, and they now have more IP that they can value.