Abstract: Adaptation of crops to climate change has to be addressed locally due to the variability of soil, climate and the specific socio-economic settings influencing farm management decisions. Adaptation of rainfed cropping systems in the Mediterranean is especially challenging due to the projected decline in precipitation in the coming decades, which will increase the risk of droughts. Methods that can help explore uncertainties in climate projections and crop modelling, such as impact response surfaces (IRSs) and ensemble modelling, can then be valuable for identifying effective adaptations. Here, an ensemble of 17 crop models was used to simulate a total of 54 adaptation options for rainfed winter wheat (Triticum aestivum) at Lleida (NE Spain). To support the ensemble building, an ex post quality check of model simulations based on several criteria was performed. Those criteria were based on the “According to Our Current Knowledge” (AOCK) concept, which has been formalized here. Adaptations were based on changes in cultivars and management regarding phenology, vernalization, sowing date and irrigation. The effects of adaptation options under changed precipitation (P), temperature (T.),[CO2] and soil type were analysed by constructing response surfaces, which we termed, in accordance with their specific purpose, adaptation response surfaces (ARSs). These were created to assess the effect of adaptations through a range of plausible P, T and [CO2] perturbations. The results indicated that impacts of altered climate were predominantly negative. No single adaptation was capable of overcoming the detrimental effect of the complex interactions imposed by the P, T and [CO2] perturbations except for supplementary irrigation (sI), which reduced the potential impacts under most of the perturbations. Yet, a combination of adaptations for dealing with climate change demonstrated that effective adaptation is possible at Lleida. Combinations based on a cultivar without vernalization requirements showed good and wide adaptation potential. Few combined adaptation options performed well under rainfed conditions. However, a single sI was sufficient to develop a high adaptation potential, including options mainly based on spring wheat, current cycle duration and early sowing date. Depending on local environment (e.g. soil type), many of these adaptations can maintain current yield levels under moderate changes in T and P, and some also under strong changes. We conclude that ARSs can offer a useful tool for supporting planning of field level adaptation under conditions of high uncertainty. (C) 2017 Elsevier Ltd. All rights reserved.

Abstract: As a part of the MACSUR task L2.4, we probabilistically calibrated the CARAIB dynamic vegetation model by Markov chain Monte Carlo (MCMC) simulation with the DREAMZS sampler. CARAIB is a mechanistic model that calculates the carbon assimilation of the vegetation as a function of the soil and climatic conditions, and can thus be used for simulating grassland production under cutting or grazing management. Bayesian model inversion was performed at 4 grassland sites across Europe: Oensingen, CH; Grillenburg, DE; Laqueuille, FR and Monte-Bodone, IT. Four daily measured variables from these sites: the Gross Primary Productivity (GPP), Net Ecosystem Exchange (NEE), Evapotranspiration (ET) and Soil Water Content (SWC) were used to sample 10 parameters related to rooting depth, stomatal conductance, specific leaf area, carbon-nitrogen ratio and water stresses. The maximized likelihood function therefore involved four objectives, whereas the applied Bayesian framework allowed for assessing the so called parameter posterior probability density function (pdf), which quantifies model parameter uncertainty caused by measurement and model errors. Sampling trials were performed using merged data from all sites (all-sites-sampling) and for each site (site-specific sampling) separately. The derived posterior parameter pdfs from the all-sites sampling and site-specific sampling runs showed differences in relation with the specificities of each site. Analysis of these distributions also revealed model sensitivity to parameters conditioned on the measured data, as well as parameter correlations.