Abstract:
Accurate estimation of crops yield is of great significance in agricultural production and has a strong guiding significance for agricultural managers. It is necessary to use an effective technical means to estimate the yield of field crops quickly and accurately. Taking winter wheat in Xiaotangshan National Precision Agricultural Research Demonstration Base as the research object, this study compared the performance of unmanned aerial vehicle (UAV) digital image and hyperspectral data in winter wheat yield estimation. The field surveys and campaigns were conducted in three typical winter wheat growth stages including flagging, flowering and filling stages. The digital images and hyperspectral data were respectively acquired by digital camera and Cubert UHD 185 Firefly imaging spectrometer, which were mounted on a UAV platform. The wheat yield data were collected during harvest. Firstly, the typical digital image indices and hyperspectral parameters were extracted from UAV digital image and hyperspectral data, respectively. Then the correlation analyses between wheat measured yield and digital image indices and hyperspectral parameters were carried out. Nine digital image indices and hyperspectral parameters with high correlation were selected for each growth stages, respectively. The selected digital image indices and hyperspectral parameters were used as modeling factors and the yield were estimated by multiple linear regression (MLR) and random forest (RF), and the models constructed by the two remote sensing data were compared to optimize the remote sensing data and model. The results showed that the digital image indices and hyperspectral parameters had significant correlation with the wheat measured yield. Among them, the correlation of the best index of different growth stages was the reflectance of the red and the best hyperspectral parameter of the three growth stages were transformed chlorophyll absorption reflectance index optimized soil adjusted vegetation index (TCARI/OSAVI), simple ratio vegetation (SR), and TCARI/OSAVI, respectively. Through the digital image indices, analyzing the effect of the modeling set, the accuracy of the MLR model was significantly better than the RF model in different growth stages, and the estimation accuracy of the two models was the highest during the filling stage and the lowest during the flagging stage. The best R2 of the MLR model was 0.71 (RMSE = 730.66 kg/hm2, NRMSE = 12.79%), and the best R2 of the RF model was 0.57 (RMSE = 894.16 kg/hm2, NRMSE = 15.65%), indicating that the advantages of the MLR model were more obvious. MLR and RF model verification effect and modeling effect remain the same. The performance of MLR and RF models had gradually increased to the filling stage to achieve the best. NRMSE reached 13.56% and 17.22%, respectively. The yield effect was estimated based on the spectral index. For MLR and RF models, the accuracy of model modeling was gradually improved, and the fitting effect was getting better and better. Among them, the best R2 of the MLR model was 0.77 and the NRMSE was 10.32%; the best R2 of the MLR model was 0.61, NRMSE was 14.79%, the estimation accuracy of MLR model was better than RF model in different growth stages. As the growth stage progresses, the verification R2 gradually increased, and RMSE and NRMSE gradually decreased. This result was consistent with the effect of the modeling set, indicating that the validation effect was relatively stable. So using UAV hyperspectral remote sensing data, the estimation model of winter wheat yield established by the MLR method can quickly and easily predict the yield of crops, and can effectively monitor the growth of crops and the performance of yield estimation models in different growth stages.