Automated Crop Harvest Detection Algorithm Based on Synergistic Use of Optical and Radar Satellite Imagery.

This study presents an automated harvest detector based on optical Sentinel-2 and radar Sentinel-1 indices. The harvest detection model is obtained by training a sequential neural network on fields with known harvest dates in Belgium. This work shows that the combination of the ratio of the Sentinel-1 based VH (vertical-horizontal) and VV (vertical-vertical) polarization and the Sentinel-2 based FAPAR are suitable for harvest detection, while coherence has no added value for the studied crops. In general, the model validation demonstrates that harvest can be predicted within 5–6 days accurately. Furthermore, the model can handle both Sentinel-1 data from ascending and descending orbits for harvest prediction. Sometimes the model fails to predict harvest accurately, especially for crops where the harvest does not equal the removal of the above-ground biomass. Future research will focus on the validation of the detector on a wider geographical area and on a wider range of crops.