Mapping sunflower areas using high resolution Sentinel-2 images

Abstract

In most countries, growing sunflower is becoming less profitable due to high input costs. Farmers often rely on subsidies and lower insurance premiums incentivized by the government but the lack of timely and accurate statistical databases at global, regional and national levels about cropland performance and spatial distribution convolutes this task for the government. Characterizing performance on a field level can furthermore improve crop growth monitoring and provide information for farmers to identify yield gaps and improve management practices. In this study we used Sentinel-2 satellite imagery and ground truth data to provide highly accurate sunflower cropland mapping. For the three growing seasons of interest (2017-2019), crop maps were created using Random Forest (RF) classification algorithm and the time-series of multispectral satellite images. Images are acquired with a spatial resolution of 10 m for visible and near-infrared bands and 20 m spatial resolution for six additional bands in red-edge and SWIR (short wave infrared) spectral domains. In each season the ground truth data was manually collected by labeling parcels using a laptop connected to a high-precision GPS receiver across Vojvodina region in northern Serbia. It served as the training data for the RF model which performed pixel-based classification of land in Vojvodina according to the crop type (sunflower/other crops). The model was cross-validated and the overall accuracy proved to be higher than 85%. The subject of future work will be the improvement of the model using other advanced machine learning models and the model transfer from Serbia to other major sunflower regions with similar climate and soil characteristics.