Forecasting Sunflower Grain Yield Using Remote Sensing Data and Statistical Models

Forecasting crop production a few weeks before harvest is of strategical interest for the cooperatives which collect, store and market grains. The recent development of Sentinel satellites opened new avenues for yield forecasting at field and farm level, thanks to their operational spatial resolution and revisiting time. In this study, we combined remote sensing data (in-season green area index, GAI) and statistical modeling to forecast sun-flower yield at field level for a range of cultivars and crop practices over different small production areas and years in southwestern France. From 2014-2016, 359 sunflower fields were monitored throughout the growing season in the 'Haute-Garonne' and 'Gers' administrative departments (SW France). From the satellite GAI esti-mates, two variables were calculated: GAImax (maximum GAI, between F1 stage and F1 + 10 days) and GAD (Green Area Duration). Different statistical modeling procedures were tested namely a linear regression (LR), a second degree polynomial regression (PR), a random forest regressor (RF) and a Gaussian process (GP). In each case, the models were tested using either GAD, GAImax or both variables, and each model was trained using in a first time GAD and GAI(max) obtained with linear interpolation, and in a second time, the same variables computed using the double sigmoid interpolation. In a perspective of yield prediction, GAD was calculated from anthesis to maturity but also from anthesis to 10/07, 20/07, 30/07 and 10/08 using remote sensing data. Sunflower grain yield at maturity was predicted with 10 models differing by their forms and the agronomic variables involved. At individual level, GY was slightly better predicted by models including GAD + GAI(max) or GAD, while models based only on GAI(max) were the less accurate. This was consistent with the major importance of post-anthesis radiation interception and senescence dynamics in the development of grain yield in sunflower. Better pre-dictions were achieved in 2014, then in 2015 and finally 2016. However, at the grain catchment area level, PR models including GAD were the most accurate ones with absolute errors ranging from 0.53 to 4.68 q.ha(-1) as a function of years. Only the predictions obtained with 2014 data and over the 3 years were sufficiently accurate to be of operational value for a cooperative manager.