Exploiting Sentinel-1 data time-series for crop classification and harvest date detection

Light source independence and the advantage of being less affected by weather conditions than optical remote sensing, as well as the sensitivity to dielectric properties and targets structure, make Synthetic Aperture Radar (SAR), particularly time-series data, a relevant tool for crop processes monitoring. This study aims to benefit from all the amplitude and phase SAR data to perform both a crop classification and a harvest date detection algorithm, supported by the first one for corn and soybean fields. Study area was located in Buenos Aires province, Argentina. To achieve this goal, time-series of Interferometric Coherence (IC) and backscattering values in vertical transmit and vertical receive (sigma(0)(VV)), and vertical transmit and horizontal receive (sigma(0)(VH)) polarizations were generated from Single Look Complex images acquired from C-band SAR satellites Sentinel-1A and -1B. The crop classification was performed using a Random Forest classifier with an overall accuracy of 97%. For its training, both sigma(0)(VV) and sigma(0)(VH) time-series along the entire crops life cycle were used. Harvest detection algorithm was accomplished by analysing both the IC and sigma(0)(VH) time-series in an individual way for both crops. IC changes could be linked to plant structure characteristics along their life cycle (from seeding to harvesting), surface structure induced by harvest operations and post-harvest crops stubble. Based on the latter, individual criteria for corn and soybean were adopted. Crop depending on the determination of the harvest date detection was supported by the crop classification obtained. Harvest detection accuracy over 80 fields was superior to 93% for both crops. The proposed methodology for harvest detection is focused on the crops structural characteristics along its life cycle and the post-harvest stubble, which could lead to different IC behaviours.

Automated summary

The study utilizes SAR data for crop classification and harvest date detection with an overall accuracy of 97%. By analyzing IC and sigma(0)(VH) time-series individually for corn and soybean, harvest detection accuracy over 93% was achieved.