Abstract: The estimation of grain yield by remote sensing is an important component of agricultural remote sensing. Timely and accurate early prediction of grain yield is of great significance for relevant national sectors to make grain marketing polices, to conduct macroeconomic regulation of food security, and to make grain trade decisions. Under different moisture treatments in arid regions of the Loess Plateau, we used ASD FieldSpec-3 spectrometer to determine the spectral reflectance of winter wheat during key growth stages. A total of 29 vegetation indices were calculated. The vegetation indices highly correlated with grain yield were screened out, and the winter wheat yield estimation models were constructed based on either the single vegetation index or the combination of multiple vegetation indices. The results showed that the most reliable and effective index for yield prediction was the spectral reflectance data collected during the booting and heading stages. The highest accuracy of yield estimation model constructed with single vegetation index was determined based on heading stage data (DVI-3) (R²=0.59, RMSE=977.60 kg·hm^-2). The models constructed with the combination of multiple vegetation indices were better than those with a single vegetation index, and generated greater accuracy of yield prediction, with the model built with heading stage data-VIs showing the best prediction result (R²=0.69, RMSE=889.55 kg·hm^-2). Our results can provide scientific basis for yield estimation of winter wheat by hyperspectral remote sensing in arid regions of the Loess Plateau.

Key words: winter wheat, drought stress, canopy spectrum, vegetation index, yield estimation.