Abstract:Timely and accurate crop monitoring and grain yield prediction before harvest of winter wheat are helpful for accurate farmland management and decision-making. Aiming to explore the potential of multitemporal vegetation indices (VIs) extracted from unmanned aerial vehicle (UAV) based multispectral images in the whole growth period of winter wheat and improve the grain yield prediction, a UAV platform carrying multispectral camera was employed to collect the high resolution images of the whole growth period of winter wheat under different water deficit states. Different kinds of multispectral VIs were used to characterize the growth characteristics of winter wheat and the correlations between VIs and winter wheat grain yield were analyzed. The multi-temporal VIs were collected to form the data set, which was used to train the machine learning algorithm. Three algorithms, including partial least squares regression (PLSR), support vector regression (SVR) and random forest regression (RFR) were used to predict the grain yield of winter wheat. The results showed that with the growth of winter wheat, the leaf area index (LAI) was changed basically as parabolic, indicating the useful of MTVI2 in remote sensing retrieval of LAI. Meanwhile, the correlation coefficient between multiple VIs and grain yield was continually increased to 0.7 at the end of the filling stage. The linear regression determination coefficient (R2) between VIs and grain yield also reached the maximum. Moreover, the accuracy of VIs forecasting grain yield was also continuously improved, because of the multi-temporal VIs reflecting the changing characteristics of winter wheat growth. The multi-temporal VIs at the flowering and early stage of filling had higher accuracy than the VIs at a single growth period. For instance, the R2 of PLSR was increased by about 0.021 and the R2 of SVR was increased by about 0.015 and the R2 of RFR was increased by about 0.051. For the multitemporal vegetation index at the end of filling stage, different models had high estimation accuracy. The highest R2 and RMSE of PLSR were 0.459 and 1822.746kg/hm2, the highest R2 and RMSE of SVR were 0.540 and 1676.520kg/hm2and the highest R2 and RMSE of RFR were 0.560 and 1633.896kg/hm2, respectively. So the RFR trained in this data set had the highest estimation accuracy and better stability. These findings demonstrated that the proposed approach can improve the prediction accuracy of grain yield as well as achieve an efficient monitoring of crop growth. Under water deficit conditions, longterm water deficit had a great impact on the growth of winter wheat at the filling stage, in turn leading to a decline of winter wheat grain yield. In comparison with normal quantity of irrigation water, the long-term water deficit caused a decrease in winter wheat production by about 1/2.
