Prediction of Rice Yield Based on Multi-Source Data and Hybrid LSSVM Algorithms in China

Accurate prediction of rice yield is essential for national food security and the development of the national economy. Currently, owing to the influence of data sources and model parameters, it is difficult to obtain simple and highly accurate models for rice yield prediction. In this study, nine typical rice ecological observation stations in China were selected to build a rice yield prediction model integrating multi-source data based on the least squares support vector machine (LSSVM) model. To improve the accuracy of the rice yield prediction model, the genetic optimization algorithm (GA), particle swarm optimization algorithm (PSO), and grey wolf optimization algorithm (GWO) were selected to optimize the parameters of the least squares support vector machine model. The correlation significances of yield with different influencing factors followed the order: total solar radiation (Ra) > number of spikes (NS) > plant height (H) > average pressure (P) > maximum temperature (Tmax) > relative humidity (RH) > precipitation (Pre) > average surface temperature (Ts) > minimum temperature (Tmin) > sunshine hours (n) > accumulated temperature (Ta), and it was highly significant with meteorological data (P = 63.1%) and significant with phenotypic data (P = 36.9%). With an increasing number of influencing input factors, the model accuracy tended to increase and then decrease when prediction model was constructed. The results showed that in the input models with different variables, the prediction accuracy was the highest when the input was Ra, NS, H, P, Tmax, RH, Pre, Ts, Tmin, and n (R 2 = 0.712–0.841, RMSE = 1.139–1.458 ton/ha, MAE = 0.814–1.085 ton/ha, and NSE = 0.702–0.831). With the reintroduction of input variables, the accuracy of the rice prediction model could not be significantly improved. Compared with stand-alone LSSVM models, hybrid optimization algorithms can significantly improve the accuracy of the LSSVM model prediction results. The results of the GA, GWO, and PSO algorithms optimized for LSSVM showed that GWO-LSSVM had the highest accuracy with R 2 = 0.841, RMSE = 1.139 ton/ha, MAE = 0.814 ton/ha, and NSE = 0.831. The best accuracies of PSO and GA were R 2 = 0.782, RMSE = 1.233 ton/ha, MAE = 0.882 ton/ha, NSE = 0.781, R 2 = 0.818, RMSE = 1.169 ton/ha, MAE = 0.863 ton/ha, and NSE = 0.798. This study suggests that the optimization algorithm is important for optimizing the hyperparameter parameters of the LSSVM model and that the GWO˗LSSVM yield prediction model is recommended for predicting rice yields in China. Graphical abstract