Intercomparison of models for simulating timothy yield in northern countries Panu Korhonen1, Taru Palosuo1, Mats Höglind2, Tomas Persson2, Guillaume Jégo3, Marcel Van Oijen5, Anne-Maj Gustavsson4, Perttu Virkajärvi1, Gilles Bélanger3 1 Natural Resources Institute Finland (Luke), Finland 2 Norwegian Institute of Bioeconomy Research (NIBIO), Norway 3 Agriculture and Agri-Food Canada (AAFC), Canada 4 Swedish University of Agricultural Sciences (SLU), Sweden 5 The Centre for Ecology & Hydrology (CEH), UK © Natural Resources Institute Finland Background • Forage-based livestock and dairy production are the economic backbone of agriculture in many northern countries. • In northern Europe and eastern Canada, forage grasses for silage are commonly grown for 2-4 years or longer in rotations with cereal crops and harvested 2-3 times per year. • In those regions, timothy (Phleum pratense L.) is one of the most widely grown forage grass species. • Models that simulate the growth and nutritive value have been developed for timothy, but the performance of different models has not been compared so far. 2 Timothy (Phleum pratense L.) © Natural Resources Institute Finland Research questions • How can current timothy models predict timothy yields of the first and second cut in northern areas of Europe and Canada where timothy is widely grown? • Are the models able to predict the timothy yield response to climatic factors and changes in management (e.g. changes in cutting times or N application rates)? • How do models perform with cultivar-specific vs. non-cultivar specific (generic) calibrations? • What is the magnitude of uncertainty associated to the yield predictions by different models? 3 © Natural Resources Institute Finland Model comparison setup 4 9.4.2015 • Three models: – BASGRA (The BASic GRAssland model, based on LINGRA) – CATIMO (CAnadian TImothy MOdel) – STICS (Simulateur mulTIdisciplinaire pour les Cultures Standard) • 7 study sites Country and site name Treatments (calibration+test) Canada 1. Fredericton 6 (4+2) (different N levels) 2. Lacombe 2 (2+0) 3. Quebec 9 (6+3) (different N levels) Finland 4. Maaninka 2 (2+0) 5. Rovaniemi 6 (4+2) (different N levels) Norway 6. Saerheim 6 (4+2) (early and late cut) Sweden 7. Umeå 2 (2+0) Altogether ~1500 observations of dry-matter yield (also for leaf and stem fractions), crop height, leaf area index and specific leaf area. 3 1 4 2 5 6 7 © Natural Resources Institute Finland Calibrations • Model users were free to use preferred calibration method – BASGRA and CATIMO applied Bayesian calibration – STICS was calibrated using the integrated optimization tool (simplex algorithm) • Data from 24 treatments were used for calibration and the remaining 9 treatments were used to assess model performance • Two different calibrations – Cultivar-specific calibration – Generic calibration applying data from all sites and cultivars 5 Study site Cultivar Years Fredericton, Canada Champ 1991-1993 Lacombe, Canada Climax 2004-2005 Quebec, Canada Champ 1999-2001 Maaninka, Finland Tammisto II 2006-2007 Rovaniemi, Finland Iki 1999-2001 Særheim, Norway Grinstad 2000-2002 Umeå, Sweden Jonatan 1995-1996 © Natural Resources Institute Finland Simulated and observed time course of dry-matter accumulation and leaf area index 6 Example: Særheim, Norway, year 2000 Dry matter yield Leaf area index © Natural Resources Institute Finland Model performance for the 1st and 2nd cuts 7 Simulated and observed maximum yields of the 1st and 2nd cut of each treatment using cultivar-specific calibration © Natural Resources Institute Finland Cultivar-specific vs. generic calibration 8 FIN FIN NOR CAN CAN SWE CAN 25% 50% 75% RMSE quartiles Arrows depict treatments used to assess model performance (not included in calibration). © Natural Resources Institute Finland Yield responses to N levels 9 Fredericton, year 1993, Cultivar-specific calibration © Natural Resources Institute Finland Uncertainty related to model predictions 10 FIN FIN NOR CAN CAN SWE CAN © Natural Resources Institute Finland Discussion • All models generally managed to estimate the DM yields satisfactorily and none of them worked clearly better than the others at all sites. • Cultivar-specific calibration provided better simulation accuracy than the generic calibration. Calibration effect on simulated yields varied among sites and treatments. • Models differed in their ability to simulate a response to nitrogen fertilization. • Uncertainties in simulated yield estimates in models are still quite wide and they are related to deficiencies in models process descriptions, uncertainties in model parameters and input data. 11 © Natural Resources Institute Finland Next steps • MACSUR2 LiveM task 1.2 - grassland quality modelling – Model survey of how current grass growth models simulate the nutritive value of forage grasses is currently going on • Related workshop to be held in connection with EGF 2016 in Trondheim (Norway) in September – Contact panu.korhonen@luke.fi if you want to join in or need more information! – Hopefully leads to model comparison paper • Results will be used to improve models: – CATIMO: Regrowth functions will be updated soon – BASGRA: Ongoing work to improve N responses – STICS: Planned upgrades to better simulation of plant reserve dynamics for improved regrowth and multiannual simulations 12 9.4.2015 © Natural Resources Institute Finland Thank you!