The Sen2-Agri algorithms

The open source Sen2-Agri system has been developed to ingest and processed automatically Sentinel-2 and Landsat 8 time series in a seamless way to derive the four basic products thanks to streamlined processes based on machine learning algorithms Inglada et al. (2015), Matton et al. (2015) and Valero et al. (2016) described the respective methods and the benchmarking results. In the course of research, methods were adjusted to Sentinel-2 data and fine-tuned to near real-time operational conditions.

An objective selection process for the underlying algorithms

After the consolidation of requirements collected from end-users and the development of the product specifications, one of the most important step of the project was the benchmarking exercise to choose the algorithms to be included in the system.

The benchmarking exercise represents an essential activity as it is the basis for the implementation of the system to be developed. In practical terms, it consists in selecting the best algorithms to match the products specifications and therefore fulfill to a maximum extent the user requirements.

The benchmarking was implemented as follows:

  • For each product, a minimum of 5 concurrent algorithms were selected basing upon literature review and a preliminary exploratory phase carried out to identify the most interesting ones out of all possibilities offered by the state of the art;
  • These 5 algorithms were run on a Test Data Set including both Earth observation and in-situ data and covering 12 JECAM sites spread over the world, which  represent more than 17 major crop types;
  • The products were validated and inter-compared using concrete and measurable criteria agreed before the assessment;
  • The best algorithm according to these predefined criteria was selected to guide the design of the future Sen2-Agri system.

The benchmarking was therefore carried out  in an objective and transparent way, ensuring the selection of methodologies suitable for a set of sites showing a wide variety in terms of crop systems or climatic conditions.

During the selection process, specific attention was also paid to methodologies that will take the most of the spatial, temporal and spectral properties of the Sentinel-2 sensors.