Abstract: Modelling N transformations within cropping systems is crucial for N management optimization in order to increase N use efficiency and reduce N losses. Such modelling remains challenging because of the complexity of N cycling in soil–plant systems. In the current study, the uncertainties of six widely used process-based models (PBMs), including APSIM, CROPSYST, DAISY, FASSET, HERMES and STICS, were tested in simulating different N managements (catch crops (CC) and different N fertilizer rates) in 12-year rotations in Western Europe. Winter wheat, sugar beet and pea were the main crops, and radish was the main CC in the tested systems. Our results showed that PBMs simulated yield, aboveground biomass, N export and N uptake well with low RMSE values, except for sugar beet, which was generally less well parameterized. Moreover, PBMs provided more accurate crop simulations (i.e. N export and N uptake) compared to simulations of soil (N mineralization and soil mineral N (SMN)) and environmental variables (N leaching). The use of multi-model ensemble mean or median of four PBMs significantly reduced the mean absolute percentage error (MAPE) between simulations and observations to less than 15% for yield, aboveground biomass, N export and N uptake. Multi-model ensemble also significantly reduced the MAPE for net N mineralization and annual N leaching to around 15%, while it was larger than 20% for SMN. Generally, PBMs well simulated the CC effects on N fluxes, i.e. increasing N mineralization and reducing N leaching in both short-term and long-term, and all PBMs correctly predicted the effects of the reduced N rate on all measured variables in the study. The uncertainties of multi-model ensemble for N mineralization, SMN and N leaching were larger, mainly because these variables are influenced by plant-soil interactions and subject to cumulative long-term effects in crop rotations, which makes them more difficult to simulate. Large differences existed between individual PBMs due to the differences in formalisms for describing N processes in soil–plant systems, the skills of modelers and the model calibration level. In addition, the model performance also depended on the simulated variables, for instance, HERMES and FASSET performed better for yield and crop biomass, APSIM, DAISY and STICS performed better for N export and N uptake, STICS provided best simulation for SMN and N leaching among the six individual PBMs in the study, but all PBMs met difficulties to well predict either average or variance of soil N mineralization. Our results showed that better calibration for soil N variables is needed to improve model predictions of N cycling in order to optimize N management in crop rotations.

Abstract: Statistical learning is a fairly new term referring to a set of supervised and unsupervised modelling and prediction techniques. It is based on traditional statistics but has been highly enhanced inspired by developments in machine learning and data mining. The main focus of statistical learning is to estimate the functions that quantify relations between several parameters and observed responses. These functions are further used for prediction, inference or a combination of both. For a particular case of quantitative responses, regularization techniques in regression are developed to overcome the weaknesses of ordinary least square (OLS) regression in prediction. These new shrinkage methods outperform OLS for prediction, especially in large datasets. In this study, a large dataset of field experiments on winter oilseed rape in Denmark for 22 years (1992 to 2013) was collected. Biweekly climatic data along with sowing date, harvest date, soil type and previous crop are considered as the explanatory variables. Yield of winter oilseed rape is considered as response variable. LASSO and Elastic Nets are the regularization techniques used to estimate the functions. Hold-one-out cross validation method for testing the prediction power reveals that these techniques are much useful in both prediction and inference. Since these techniques are included in recent versions of some software packages (e.g. R), they can be easily implemented by users at any level. The estimated function (model) is further used to predict the oilseed rape yield responses to climate change for several scenarios using representative weather data produced by a weather generator.

Abstract: Highlights • Regularization techniques for regression outperformed the classical regression techniques in predicting crop yields. • Different regression techniques with similar prediction accuracy showed different responses of major climatic variables to crop yield. • The regression models showed some responses of crop yield to climatic conditions that is mostly absent in process based crop models. Abstract Statistical regression models represent alternatives to process-based dynamic models for predicting the response of crop yields to variation in climatic conditions. Regression models can be used to quantify the effect of change in temperature and precipitation on yields. However, it is difficult to identify the most relevant input variables that should be included in regression models due to the high number of candidate variables and to their correlations. This paper compares several regression techniques for modeling response of winter oilseed rape yield to a high number of correlated input variables. Several statistical regression methods were fitted to a dataset including 689 observations of winter oilseed rape yield from replicated field experiments conducted in 239 sites in Denmark, covering nearly all regions of the country from 1992 to 2013. Regression methods were compared by cross-validation. The regression methods leading to the most accurate yield predictions were Lasso and Elastic Net, and the least accurate methods were ordinary least squares and stepwise regression. Partial least squares and ridge regression methods gave intermediate results. The estimated relative yield change for a +1°C temperature increase during flowering was estimated to range between 0 and +6 %, depending on choice of regression method. Precipitation was found to have an adverse effect on yield during autumn and winter. It was estimated that an increase in precipitation of +1 mm/day would result in a relative yield change ranging from 0 to −4 %. Soil type was also important for crop yields with lower yields on sandy soils compared to loamy soils. Later sowing was found to result in increased crop yield. The estimated effect of climate on yield was highly sensitive to the chosen regression method. Regression models showing similar performance led in some cases to different conclusions with respect to effect of temperature and precipitation. Hence, it is recommended to apply an ensemble of regression models, in order to account for the sensitivity of the data driven models for projecting crop yield under climate change.