Demonstrating Loamin algorithms for soil mapping at scale

Loamin have developed a geospatial machine learning approach to estimating soil properties, such as soil carbon and pH. These models are trained using over 20’000 geolocated ground truth soil samples. In addition, building on work developed by Hugh Sturrock while in academia, Loamin have developed novel algorithms to optimise further soil data collection which may be required to improve the precision of estimates and/or adhere to specific protocols/regulations.

The maps below show estimates of current soil carbon stock (t/ha) and pH at 0-20cm depth for 2 farms supplied by Agrii, as well as the recommended locations at which to collect further soil samples to improve the maps. These sample locations are selected by finding an optimal balance between sampling areas of the map with high uncertainty, while also ensuring a good spread of points across geographic space. While it is possible to optimise sampling for more than one soil indicator, for this exercise, given that lab tests for soil carbon are more expensive than for pH, the survey has been optimised for soil carbon stock mapping.