Phenological piecewise modelling is more conducive than whole-season modelling to winter wheat yield estimation based on remote sensing data

Abstract

Most of the existing remote sensing-based yield estimation methods adopt the mean or cumulative value of meteorological factors within the whole growing season, which may ignore the impact of adverse meteorological conditions on the growth of winter wheat in a certain phenological period. In this study, we distinguished the developmental progression of winter wheat as three phenological periods. In each phenological period, the vegetation indices and meteorological factors were optimized. Then the accuracy and spatiotemporal transferability of the phenological piecewise modelling was compared with that of the whole-season modelling based on four regression methods (i.e. multiple linear regression, artificial neural network, support vector regression and random forest). The results showed that the optimal combinations of variables for the whole-season modelling and the phenological piecewise modelling were different. Compared with the whole-season models, the R2 for the phenological piecewise models improved by 1.4% to 7.6%, the root mean square error (RMSE) decreased by 1.1% to 8.2% among four regression methods . In addition, compared with the whole-season models, the spatiotemporal transferability for the phenological piecewise models was generally better. The accuracies after spatiotemporal transfer for the phenological piecewise models were still higher than that for the whole-season models.

Keywords

• Meteorological factors
• phenology
• remote sensing
• winter wheat
• yield estimation
• machine learning