AI Monitoring - Pilot study
| Reference number | |
| Coordinator | Statens Jordbruksverk |
| Funding from Vinnova | SEK 500 000 |
| Project duration | December 2020 - August 2021 |
| Status | Completed |
| Venture | AI - Competence, ability and application |
| Call | Start your AI-journey for public organizations - autumn 2020 |
Important results from the project
The aim of the project was to increase the competence of the Swedish Board of Agriculture (SJV) in AI and to spread the competence internally in SJV and externally. The goal was to develop AI algorithms that in satellite images over a limited area automatically can detect activities on grasslands and pastures during the year. This has been well achieved by: - develop methodology that helps automate the detection of changes in grasslands. - knowledge exchange between RISE and SJV and sharing of AI code - information about the project has been disseminated nationally and internationally
Expected long term effects
Due to the limited size of the project, it was not possible to develop a fully functional AI method. However, the knowledge exchange between SJV and RISE during the project has generated important results for AI development of future data such as: - Reuseable NetCDF datasets - Model based on unsupervised machine learning developed - Jupyter notebooks for knowledge transfer created - Increased performance of the Space Data Lab to extract the imagery SJV requests - High quality test material developed and suggestions for AI models for continued work after the project.
Approach and implementation
A dataset has been compiled containing the location and use of 11,700 agricultural fields and Sentinel-2 satellite images of the same areas for year 2020. Unsupervised machine learning was used to automatically group the pastures based on their use. This by analyzing satellite data and generating time series based on NDVI (Normalized Difference Vegetation Index) calculated with Gaussian Process Regression. Cluster analysis (Dynamic Time Warping and Hierarchical clustering) showed a correlation between NDVI time series and use of pastures.