Comparison of yield prediction models and estimation of the relative importance of main agronomic traits affecting rice yield formation in saline-sodic paddy fields

Abstract

Saline-sodic soils are widely distributed in the western Songnen Plain of Northeast China, and soil salinization has increasingly become an obstacle to local economic development and ecological security. The practice has proved that rice (Oryza sativa L.) cultivation is one of the most effective methods for the biological improvement and utilization of such land. However, the key limiting factors of rice yield are unclear because of lacking an effective method to quantitatively assess the relative importance of rice agronomic traits to yield formation, resulting in low rice yield in saline-sodic soils. In this study, the random forest model (RF) and the structural equation model (SEM) were conducted to quantitatively assess the relative importance of the rice agronomic traits to yield and the interrelationships and influence modes of rice yield components to yield, respectively. We also compared the prediction performance of different models using two-year data based on important agronomic traits and predicted yield-increasing potential in saline-sodic paddy fields. The results showed that the aboveground biomass and the straw weight were the most important explanatory variable for rice yield per hill, with the explained variation of 32.4 % and 19.4 %, respectively. The plant numbers per m2 and plant height were the most important explanatory variable for rice yield per m2, with the explained variation of 32.5 % and 20.6 %, respectively. Rice yield was directly affected by panicle numbers and panicle weight (decided by the spikelet number per panicle), and the yield-increasing potential of panicle numbers was greater than panicle weight. The random forest model (RF) performed much better than the traditional model (TR) and the multiple linear regression model (MLR) based on agronomic traits to predict yield in saline-sodic paddy fields, which showed a higher R2 and lower error indices (MAE and RMSE). The RF also predicted that rice yield can increase by 9.4 % if plant numbers per m2 increase by 18.4 % (when seedling number per hill reached the level of no salt-affected paddy fields) in saline-sodic paddy fields, and tillering capacity per hill plays a vital role in this process. Therefore, RF may be the preferred method for predicting rice yield based on agronomic traits in saline-sodic paddy fields. Rice cultivation in saline-sodic soils should pay attention to the regulation of dry matter accumulation and tillering capacity of rice during vegetative growth for ensuring sufficient panicle number in reproductive growth period to achieve the purpose of increasing yield.