Detecting the world’s highest accuracy field boundaries to power precision agriculture

Using deep neural network models and super-resolution Sentinel-2 EO-Data for detecting the world’s highest accuracy field boundaries

DigiFarm is a Norwegian based ag-tech startup established in 2019. DigiFarm’s core vision is to detect the world’s most accurate field boundaries and seeded acres to power precision agriculture, leveraging the latest advancements in Artificial Intelligence technology. This is achieved through developing deep neural network models for automatically detecting field boundaries through super-resolving Sentinel-2 satellite imagery (to 1 meter resolution). DigiFarm has successfully validated the model on 15+ million hectares of fields achieving detection accuracies of above 96%, 12-15% higher than existing boundary data (Cadastral, LPIS in EU and CLUs in US). 

The pilot will include: developing and modeling a training deep neural network model for detection of entire-country sized regions including Germany, Austria, Belgium and the United Kingdom. 

The solution will allow agricultural organizations to optimize operations, making better data-driven decisions, reducing seasonal uncertainty, minimising production costs, increasing crop-revenue and enabling key developments in carbon capture, leading to reduction in Co2 footprints. DigiFarm delivers software as a service seamlessly integrated into clients (B2B/B2G) digital solutions (API). The pilot results will be implemented into corporate partners (KWS) commercial FMS-solutions after the pilot has been finalized. 


All precision farming services and in-field analytics start with accurate field boundaries and seeded acres. Unfortunately, existing field boundary data is ofoutdated and inaccurate, which affects the entire agricultural value chain from pre-production to grain-processing. Large scale boundary data is managed through national agencies (Cadastral): NIBIO in Norway dates back to 1990, Land Parcel Identification Systems for all 27 EU-regions and the Common Land Units in the US are 13 years old. These boundaries were created manually drawing field boundaries into digital GIS (Geographical Information System) solutions. Creating, updating and maintaining manually delineated field boundaries is time-consuming, resource intensive, slow and costly. As boundaries and seeded acres change every year, existing data becomes increasingly unreliable and inaccurate. For example, we discovered that 20% of all field boundaries in Norway are inaccurate, affecting €1 billion in annual subsidies. In the 27 EU-regions, the same issue affects over €45.5 billion subsidies annually!  Our research showed similar levels of inaccuracies in LPIS data and CLUs, approx. 12-15%, across 10 million hectares (comparing LPIS and our modeling). Farmers are also affected: 40% of all agricultural fields are overfertilized and farmers are losing 15-20% in yield potentials through inadequate input application.

Work Plan

Collect and process manual training data: Manually delineated boundaries (drawing boundaries for training data in QGIS-solution which will be used to train different models) for each of the selected areas, have to make sure the data is collected representing a variety of different areas with different fields and in-field artifacts

Commence training of multiple different deep neural network models with different hyperparameters in each region in order to assess which ones perform best with the highest accuracy. DigiFarm intends to use IoU (Intersection over Union) as an accuracy measurement, common for image segmentation and object detection algorithms. The goal in this activity is to achieve an Intersection over Union (with 1 being the highest, and 0 the lowest accuracy) of 0.965 or better. This is the threshold we have determined in collaboration with clients as an acceptable accuracy level (~12-15% higher accuracy than current Cadastral models).



Business Partner

Supporting project