Abstract:In order to fully exploit the time-series information and trend information of time-series remotely sensed parameters and further improve the yield estimation accuracy of winter wheat, vegetation temperature condition index (VTCI), leaf area index (LAI) and fraction of photosynthetically active radiation (FPAR), which were closely related to the growth and development of winter wheat, were selected as remotely sensed parameters, and a neural network was constructed based on variational mode decomposition (VMD) and gated recurrent unit (GRU). The VMD algorithm was applied to decompose each remotely sensed parameter series into multiple sets of intrinsic mode function (IMF) components, and the IMF components that were highly correlated with the original remotely sensed parameter series were selected for feature reconstruction, and the reconstructed features were used as the input of the GRU network to develop a combined model for yield estimation of winter wheat. The results showed that the VMD-GRU model for yield estimation had a coefficient of determination of 0.63, root mean squared error of 448.80kg/hm2, and mean relative error of 8.14%, with a highly significant correlation level (P<0.01), and its accuracy was better than that of the single model for yield estimation, indicating that the combined model for yield estimation can extract multi-scale and multilevel features of non-stationary time series and fully explore the internal linkage between remotely sensed parameters in each growth stage of winter wheat to obtain accurate yield estimation results and improve interpretability of model for yield estimation.
