Accurate and spatially detailed information on agricultural land use is critical for tracking developments in the agricultural sector and advancing toward more sustainable practices. In Switzerland’s national greenhouse gas (GHG) inventory, emissions and removals from cropland and grassland soils are modelled annually using the RothC carbon model. However, these simulations currently rely on aggregated land use data. Moving forward, there is a strong need for spatially explicit input variables, including yearly maps of main crops.
To address this, Agroscope has been developing a deep-learning-based system that classifies major crop types using satellite data from the Sentinel-2 mission (c.f. DeepField, Crop mAIpper). This system, developed from a prototype by Turkoglu et al. (2021), is now being transitioned into operational use, producing consistent, nationwide crop maps from 2017 onward.
Additionally, historical time series reaching back to at least 1990 are required. Since Sentinel-2 data only goes back to 2017, this project investigates the feasibility of extending crop classification further into the past using data from the Landsat satellite series, which has been monitoring Earth since the 1980s at 30-meter resolution.
Aims
- Adapt and evaluate Sentinel-2-based crop classification methods for use with Landsat data
- Assess the ability to distinguish among 19 major annual crops and multiple types of managed grassland relevant to the GHG inventory
- Analyze the impact of differences in sensor resolution, cloud cover, and data availability on classification accuracy
- Evaluate the consistency of the combined Landsat-Sentinel-2 time series with IPCC methodological guidelines
- Create spatially explicit crop and grassland maps from 1990 to today
Team
Partners
Andreas Schellenberger, Federal Office for the Environment (FOEN)
Funding
We are thankful to Federal Office for the Environment (FOEN) for funding this project.
