Abstract:Dynamic monitoring of crop growth and accurate estimation of crop yield can provide effective support for agricultural operators’ field management and national food policy formulation. In order to improve the estimation accuracy of maize yield, a study was carried out in central plain of Hebei Province, including Baoding City, Shijiazhuang City, Cangzhou City, Hengshui City and Langfang City, from 2010 to 2018. The experiment was characterized by remotely sensed vegetation temperature condition index (VTCI) and Savitzky-Golay filtered leaf area index (LAI), which were closely related to maize growth and yield. Because the effects of water stress on maize yield at different growth stages were different, the weights of VTCI and LAI in the main growth stages (seedling-jointing, jointing-booting, booting-milking, milking-mature) of maize were determined by using the random forest regression method. The results showed that the weights based on the random forest regression were consistent with the actual growth of maize. Based on the determined weights, the weighted VTCI and LAI at the main growth stages of maize in each county (district) were calculated, and the univariate and bivariate estimation models of weighted VTCI and LAI with maize yield in 2010—2016 (except 2012) were constructed. The results showed that the accuracy of the bivariate estimation model (R2=0.303) was higher than that of the univariate estimation models, and the bivariate model reached a very significant level (P<0.001), indicating that maize yield was related to VTCI and LAI. In summary, the bivariate estimation model based on the random forest regression had the highest accuracy. The bivariate estimation model based on the random forest regression was used to estimate the maize yield in each county (district) of the study area in 2012. The results showed that the average relative error between estimated yield and actual yield of 53 counties (districts) was 985%, and that of 31 counties (districts) were below 10%, 7 counties (districts) were between 10% and 15%, 15 counties (districts) were more than 15% and the root mean square error was 824.77kg/hm2. In order to further verify the accuracy of the bivariate estimation model, a linear regression analysis model between actual yield and estimated yield of maize in 2012 was established. It could be seen that there was a significant positive correlation between estimated yield and actual yield (P<0.001) and R2 reached 0.540, further indicating that the accuracy of the bivariate estimation model based on random forest regression was high. The bivariate estimation model based on the random forest regression was used to estimate the yield of maize in the region from 2010 to 2018. The results showed that the spatial distribution of maize yield was the highest in the western region of the plain, the next was in the north and south regions, and the lowest was in the eastern region. The distribution in time was characterized by a tendency to decrease first in the fluctuations and then increase. This was consistent with the actual spatial and temporal distribution characteristics of maize yield. The research result can provide reference for maize growth monitoring and yield estimation.
