Using Sentinel-1 data for soybean harvest detection in Vojvodina province, Serbia

Abstract

Information on crop harvest events has become valuable input for models related to food security and agricultural management and optimization. Precise large scale harvest detection depends on temporal resolution and satellite images availability. Synthetic Aperture Radar (SAR) data are more suitable than optical, since the images are not affected by clouds. This study compares two methods for harvest detection of soybean in Vojvodina province (Serbia), using the C-band of Sentinel-1. The first method represents a maximum difference of ascending VH polarization backscatter (σVH) between consecutive dates of observation. The second method uses a Radar Vegetation Index (RVI) threshold value of 0.39, optimized to minimize Mean Absolute Error (MAE). The training data consisted of 50 m point buffers’ mean value with ground-truth harvest dates (n=100) from the 2018 and 2019 growing seasons. The first method showed better performance with Pearson correlation coefficient r=0.85 and MAE=5 days, whereas the calculated metrics for the RVI threshold method were r=0.69 and MAE=8 days. Therefore, validation was performed only for the method of maximum VH backscatter difference where mean values of parcels with ground-truth harvest dates for 2020 had generated the validation dataset (n=67). Performance metrics (r=0.83 and MAE=3 days) confirmed the suitability for accurate harvest detection. Ultimately, a soybean harvest map was generated on a parcel level for Vojvodina province.