Crop Rotation Modeling for Deep Learning-Based Parcel Classification from Satellite Time Series

The Common Agricultural Policy (CAP) is responsible for allocating agricultural subsidies in the European Union, which nears 50 billion euros each year [36]. Consequently, monitoring the crop types for subsidy allocation represents a significant challenge for payment agencies, which have encouraged the development of automated crop classification tools based on machine learning [24]. In particular, The Sentinels for Common Agricultural Policy (Sen4CAP) project [19] aims to provide EU member states with algorithmic solutions and best practice studies on crop monitoring based on satellite data from the Sentinel constellation [10]. Despite the inherent difficulty of differentiating between the complex growth patterns of plants, this task is made possible by the nearly limitless access to data and annotations. Indeed, Sentinel-2 offers multi-spectral observations at a high revisit time of five days on average, which are particularly appropriate for characterizing the complex spectral and temporal characteristics of crop phenology.