Study on Yield Estimation of Wheat Varieties Based on Multi-Source Data

Pre-production estimation of wheat production is related to the formulation of agricultural production plans, food security, national economy and macro-decision-making, and the application of drones can estimate wheat production in a non-destructive, fast, accurate, timely and efficient manner. The machine learning method is used to fully tap the potential of multi-source remote sensing data to estimate the grain yield of multiple wheat varieties and to clarify the effect of multi-source data fusion on improving the yield estimation accuracy of cultivars. It is significant for crop field management and ensuring a high and stable yield in wheat. In this study, field trials of winter wheat were carried out with 140 main wheat varieties in the Huanghuai wheat region as materials. The drone platform equipped with red green blue (RGB) and multispectral sensors were used to collect the canopy information of 140 winter wheat varieties during the grain filling period. Six machine learning algorithms were used, namely Ridge Regression (RR) support vector regression (SVR) Random Forest Regression (RFR) Gaussian Process (GP), k-Nearest Neighbor (k-NN) and Cubist, to build yield estimation models from single sensor data and multi-source data fusion. Coefficient of determination (R-2), root mean square error (RMSE) and relative root mean square error (RRMSE) were used to evaluate the estimation model. The results showed that the selected 10 visible vegetation indices and 13 multispectral covered indices were significantly correlated with the measured yield (p<0. 05), and the absolute value of the correlation coefficient from high to low was multispectral vegetation index (0. 54 similar to 0. 83), color index (0. 45 similar to 0. 61), texture feature (<0. 45), all six machine learning algorithms have the highest yield estimation and prediction accuracy when using multi-source data fusion. Multi-source data fusion yield estimation accuracy (average coefficient of determination R-2 = 0. 50 similar to 0. 71) > multi-spectral sensor yield estimation accuracy (R-2 = 0. 53 similar to 0. 69) > RGB sensor yield estimation accuracy (R-2 = 0. 35 similar to 0. 51). Compared with RGB data, the R-2 of multi-source data fusion increases by 0. 17 similar to 0. 23, and the mean root mean square error (RMSE) decreases by 0. 06 similar to 0. 09 t center dot hm(-2); compared with multi-spectral data, the R-2 increases by 0. 01 similar to 0. 06, and the RMSE decreases by 0. 01 similar to 0. 03 t center dot hm(-2). Compared with the other five algorithms, the multi-source data fusion model established by the Cubist algorithm has the highest yield estimation accuracy, with an R-2 of 0. 71 and an RMSE of 0. 29 t center dot hm(-2). It shows that compared with the yield estimation model of single sensor data, multi-source data fusion can effectively improve the yield estimation accuracy of winter wheat varieties, and the Cubist algorithm can better process multi-mode data to improve the yield prediction accuracy, providing theoretical guidance for predicting the yield of different wheat varieties.