Statistical analysis and modelling of crop yield and nitrogen use efficiency in China

Summary

Excess nitrogen (N) application not only decreases the economic efficiency of fertilizer application, but can also result in serious environmental problems. This thesis aimed to explore the effect of explanatory variables on crop yield and nitrogen use efficiency (NUE), and quantify the uncertainty sources of NUE model predictions. Provincial data about crop yield and NUE from 1978 to 2015 were used to establish stepwise multiple linear regression (SMLR) models. The explanatory variables included soil properties, crop types, economy, climate, topography and agricultural management. Results showed that NUE varies considerably in space and time in China. The partial factor productivity of nitrogen (PFPN) was larger in north- and south-east China than in other regions. The partial nutrient balance of nitrogen (PNBN) was lower in south China than in other parts of China. The national PFPN declined slightly from 32 kg kg-1 in 1978 to 27 kg kg-1 in 1995 and went up gradually to reach 38 kg kg-1 in 2015. The national PNBN decreased from 0.53 to 0.36 kg kg-1 from 1978 to 2003; thereafter it stabilized at around 0.40 kg kg-1 yr-1 between 2004 and 2015. Multiple linear regression explained 74% of the NUE variation. Crop types and various soil properties were identified as major influential factors in the PFPN model; while crop types, climate and soil properties accounted for most of the PNBN variation. The aggregated yield for all crops was higher in provinces in the eastern coast and south China than in the inland west and north China. The SMLR explained 95% of the aggregated yield spatio-temporal variation. Crop types, soil covariates, economic variables and agricultural management practices were the major explanatory variables in modelling the aggregated crop yield. I concluded that enhancing economic growth could be an adequate solution to meet the growing food demand for an increasing population with limited agricultural land resources, in combination with better management practices, crop composition, breeding and planting technologies. Analysis about NUE at a finer county scale in northeast China revealed different patterns and variable importances than the coarse provincial scale analysis. NUE decreased in most of the counties in northeast China during the study period and was highest in Heilongjiang province. The soil, crop and climatic covariates had higher relative importance in the SMLR model of NUE variation. The random forest (RF) model (model efficiency coefficients 0.84 and 0.89 for PFPN and PNBN, respectively) had a superior performance than the SMLR model (model efficiency coefficients 0.44 and 0.67 for PFPN and PNBN, respectively), which indicated a non-linear relation between explanatory variables and NUE. As expected, RF performed better than SMLR for modelling NUE variation, but the analysis also revealed that neither of the two models was perfect. Therefore, it is pivotal to quantify the uncertainty of the NUE predictions and determine contributions of uncertainty sources for the 31 provinces from 1978 to 2015. The uncertainty of the NUE prediction caused by measurement uncertainty in yield, N input and N removal was quantified using Monte Carlo simulation in three scenarios, while model uncertainty was assessed using quantile regression forest (QRF). The prediction uncertainty for both NUE indicators decreased over time. In 2015, PFPN had a higher 90% prediction interval ratio (PIR90) of input data in south and west China and a higher 90% Prediction Interval Width (PIW90) in south and east-costal China, while PNBN had a higher PIR90 in north China and a higher PIW90 in northeast China. The NUE prediction uncertainty propagated from QRF models has similar spatial patterns as the input data. NUE in most provinces had smaller input uncertainty than model uncertainty, except PNBN, which had larger model uncertainty than input uncertainty after 2010. Overall, the uncertainties in NUE predictions were substantial. A series of recommendations were made to improve the accuracy of NUE predictions. These may be applied by the government, in order to inform sustainable nitrogen management in food systems.