Predicting sunflower grain yield using remote sensing data and models

International audiencePredicting grain yield (GY) and quality (oil, protein) a few weeks before harvest is of strategical interest for the cooperatives which collect, store and market grains, Several approaches have been used tbr that on various crops: (a) monitoring a network of representative fields using sensors and crop indicators; (b) running agroclimatic or process-based models at field level then upscaling the predictions at landscape level; (c) combining observations from rernote sensing (satellite) and models (statistical or dynamic) for improving GY prediction. Climate variability and \Vater deficit related to global warming both increase the uncertainty on model predictions. Correcting the predictions by assimilating vegetation and soil data into crop simulations is a way to reduce titis uncertainty. Among field crops, sunflower has been scarcely used in remote-sensing studies in spite of the plasticity of green area index (GAI) and its key role in yield and oil achievement In this study, we combined remote sensing and two types of models (statistical and dynamic) to predict sunflower yield for a range of cultivars and crop practices. In 2014 and 2015, 187 sunflower fields \Vere monitorcd in MidiPyrénées (SW France) at different periods of the growing season and GY was provided by the farmers. GAI was estimated by the inversion of a radiative transfer mode! from satellite images (Landsat 8 and Deimos in 2014; Landsat 8 and Spot 5 in 2015) at 6 (2014) and 11 (2015) dates throughout the growing season. From these estimations, GAimax and GAD (Green Area Duration) on 10 August were calculated and used in several predictive linear models of GY. In parallel, the crop mode! SUNFLO was run on farmer's fields using soil conditions, daily weather, and management data (incl. variety). Both inversion and assimilation techniques based on GAI from remote sensing were used to correct the poor representation of environmental conditions (initial N and water conditions, soil type) and reduce the error on GY prediction. As statistical models used robust relations between GY and GAD and/or GAimax• a better performance \Vas achieved with these models. However, assimilating GAI and correcting initial conditions improved the predictions of GY by SUNFLO a!though GY was still overestimated after correcting it by limîting factors (weeds, fungal diseases, low plant population). Predicting oil \.Yill be the next issue to address with this approach