Near real-time yield forecasting of winter wheat using Sentinel-2 imagery at the early stages

Winter wheat is one of the main crops in Canada. Near real-time forecasting of within-field variability of yield in winter wheat at the early stages is essential for precision farming. However, the crop yield modelling based on high spatial resolution satellite data is generally affected by the lack of continuous satellite observations, resulting in reducing the generalization ability of the models and increasing the difficulty of near real-time crop yield forecasting at the early stages. In this study, the correlations between Sentinel-2 data (vegetation indices and reflectance) and yield data collected by combine harvester were investigated and a generalized multivariate linear regression (MLR) model was built and tested with data acquired in different years. In addition, three simple unsupervised domain adaptation (DA) methods were adopted for improving the generalization ability of yield prediction. The winter wheat yield prediction using multiple vegetation indices showed higher accuracy than using single vegetation index. The optimum stage for winter wheat yield forecasting varied with different fields when using vegetation indices, while it was consistent when using multispectral reflectance and the optimum stage for winter wheat yield prediction was at the end of flowering stage. This study demonstrated that the simple mean matching (MM) performed better than other DA methods and it was found that "DA then MLR at the optimum stage " performed better than "MLR directly at the early stages " for winter wheat yield forecasting at the early stages. The results indicated that the DA had a great potential in near real-time crop yield forecasting at the early stages.