Combining Machine Learning and Spatio-Temporal Filtering to Map Crop Types of Germany for 7 Years

Knowledge of the distribution of crop types and their sequence over the years is essential not only for scientific applications but also for supporting informed planning for food security, climate adaptation and mitigation, agroecology and landscape diversity. For Germany, crop type information is collected as part of the subsidy management, but data access is restricted for certain years and Federal States. Remote sensing time series of sensors such as Sentinel-1 and Sentinel-2 allow national mapping of crop types at field scale. This has been demonstrated in previous literature which describe crop type mapping of Germany for one to three years. Here, we demonstrate a two-phase crop type classification methodology based on Sentinel-1 and Sentinel-2 data. The approach combines machine learning based classification with spatial and temporal analyses. The methodology was used to create a novel and most recent seven-year time series (2018-2024) for Germany at 10m spatial resolution separating 18 classes. The two-phase approach led to annual overall accuracies of 0.81 to 0.83 with average class-specific F1-scores ranging from around 0.56 to 0.99.