Abstract: Within the context of nitrogen (N) management, since 1950, with the rapid intensification of agriculture, farmers have often applied much larger fertiliser quantities than what was required to reach the yield potential. However, to prevent pollution of surface and groundwater induced by nitrates, The European Community launched The European Nitrates Directive 91/6/76/EEC. In 2002, in Wallonia (Belgium), the Nitrates Directive has been transposed under the Sustainable Nitrogen Management in Agriculture Program (PGDA), with the aim of maintaining productivity and revenue for the country’s farmers, while reducing the environmental impact of excessive N application. A feasible approach for addressing climatic uncertainty lies in the use of crop models such as the one commonly known as STICS (simulateur multidisciplinaire pour les cultures standard). These models allow the impact on crops of the interaction between cropping systems and climatic records to be assessed. Comprehensive historical climatic records are rare, however, and therefore the yield distribution values obtained using such an approach can be discontinuous. In order to obtain better and more detailed yield distribution information, the use of a high number of stochastically generated climate time series was proposed, relying on the LARS-Weather Generator. The study focused on the interactions between varying N practices and climatic conditions. Historically and currently, Belgian farmers apply 180 kg N ha(-1), split into three equal fractions applied at the tillering, stem elongation and flag-leaf stages. This study analysed the effectiveness of this treatment in detail, comparing it to similar practices where only the N rates applied at the flag-leaf stage were modified. Three types of farmer decision-making were analysed. The first related to the choice of N strategy for maximising yield, the second to obtaining the highest net revenue, and the third to reduce the environmental impact of potential N leaching, which carries the likelihood of taxation if inappropriate N rates are applied. The results showed reduced discontinuity in the yield distribution values thus obtained. In general, the modulation of N levels to accord with current farmer practices showed considerable asymmetry. In other words, these practices maximised the probability of achieving yields that were at least superior to the mean of the distribution values, thus reducing risk for the farmers. The practice based on applying the highest amounts (60-60-100 kg N ha(-1)) produced the best yield distribution results. When simple economical criteria were computed, the 60-60-80 kg N ha(-1) protocol was found to be optimal for 80-90% of the time. There were no statistical differences, however, between this practice and Belgian farmers’ current practice. When the taxation linked to a high level of potentially leachable N remaining in the soil after harvest was considered, this methodology clearly showed that, in 3 years out of 4,30 kg N ha(-1) could systematically be saved in comparison with the usual practice.

Abstract: Crop models describe the growth and development of a crop interacting with its surrounding agro-environmental conditions (soil, climate and the close conditions of the plant). However, the implementation of such models remains difficult because of the high number of explanatory variables and parameters. It often happens that important discrepancies appear between measured and simulated values. This article aims to highlight the different sources of uncertainty related to the use of crop models, as well as the actual methods that allow a compensation for or, at least, a consideration of these sources of error during analysis of the model results. This article presents a literature review, which firstly synthesises the general mathematical structure of crop models. The main criteria for evaluating crop models are then described. Finally, several methods used for improving models are given. Parameter estimation methods, including frequentist and Bayesian approaches, are presented and data assimilation methods are reviewed.