Developing soybean yield prediction model based on multicolor space and texture features of UAV images

• Skew distribution parameters provided more leaf color information. • Canopy multistage image feature parameters expanded data dimension for prediction model. • BPNN model constructed by multiple parameters of multistage performed accuracy of 88.2%. In order to intelligently and non-destructively estimate soybean yield, the fusion of multicolor space and texture feature parameters were considered to construct a yield prediction model. In this study, the yield prediction model was developed through different stands formed by a management experiment of different planting densities and nitrogen application strategies, and validated through a variety test of 28 soybean varieties. Images of the soybean canopy were collected during the key period for yield (the florescence, podding, and grain-filling stages) by unmanned aerial vehicle (UAV). The multicolor space and texture feature parameters of the RGB images of soybean canopy were extracted for these three periods, and soybean yield prediction models were constructed for the florescence, podding, grain-filling, and multiple growth stages based on different parameters combinations, by using methods of multiple linear regression (SMLR), random forest (RF), and back propagation neural networks (BPNN). The results showed that the color space and texture features of the soybean canopy images exhibited significant differences and different trends during the florescence, podding, and grain-filling stages. Models built with the combinations of texture features (TF) and Hue, Saturation and Value (HSV) parameters had little changes in accuracy, while those incorporating skewed parameters (SP) had better model accuracy. Among all the models, the model combining the SP, TF and HSV parameters demonstrated significantly greater accuracy. The accuracy of models based on individual reproductive stage was lower than those based on entire reproductive period. Thus, the best-performing model was a BPNN model using a combination of the SP, HSV and TF parameters of entire reproductive period as input factors, achieving an R 2 of 0.765, a prediction accuracy (PA) of the validation set of 88.2 %, and a root mean square error (RMSE) of 430.50 kg/ha. The accuracy of this model in predicting soybean yields across different varieties was PA = 80.4 %, and the RMSE = 514.28 kg/ha. This article provides an effective and low-cost method for accurately estimating soybean yield in the field. This method performs robustly in different varieties and agricultural practices, and has practical value.