Choosing suitable agroforestry species, varieties and seed sources for future climates with ensemble approaches
Choosing suitable agroforestry species, varieties and seed sources for future climates with ensemble approacheswca2014-1504 Roeland Kindt 1,*Eike Luedeling 2,Paulo van Breugel 1 3,Jens-Peter B. Lillesø 3,Katja Kehlenbeck 1,James Ngulu 1,Barbara Vinceti 4Hannes Gaisberger 4,Ian Dawson 1 5,Lars Graudal 3,Ramni Jamnadass 1,Henry Neufeldt 6 1Science Domain 3, 2Science Domain 4, World Agroforestry Centre, Nairobi, Kenya, 3Department of Geosciences and Natural Resource Management, University of Copenhagen, Copenhagen, Denmark, 4Forest Genetic Resources Conservation and Use, Bioversity International, Rome, Italy, 5James Hutton Institute, York, United Kingdom, 6Science Domain 6, World Agroforestry Centre, Nairobi, Kenya
Adaptation to future climates requires that we plan interventions on the basis of reliable models for predicting how existing and potential future agroforestry systems will perform. Because process-based models are currently only available for a limited number of trees, species distribution modelling (SDM) is currently the most sophisticated approach to project climate change impacts for the majority of species. (The climate analogue methodology provides an alternative approach, whereby we will provide some examples how the SDM and climate analogue methods can be complementary.)
SDM is based on statistical inference to determine environmental niches (of species, varieties or seed sources), which then allow distribution maps to be drawn both in environmental and geographic space. The power of SDM has recently increased through the introduction of machine-learning algorithms, the application of ensemble approaches and the availability of high resolution raster data sets. Ensemble approaches are founded on weighted averaging of predictions from a large suite of algorithms, including maximum entropy-, boosted regression tree- and random forest-methods. As the options for modifying weights can result in an infinite number of ensemble models, we developed a statistical method for tuning input weights and a suitability mapping approach based on the number of algorithms that predict presence–absence. These methods have been integrated into the BioversityR package, including outputs that can be immediately scrutinized and shared through Google Earth.
We will provide examples to demonstrate our approach from recent studies in Africa, Asia and Latin America. Included are future suitability investigations of timber species in Latin America and food tree species in Burkina Faso, Africa. We will also show how information on potential natural vegetation can be combined with SDM approaches to improve seed sourcing strategies such that they better consider climate change. Results from transect studies of mango variety turnover in Kenya will also demonstrate how production data can be included in ensemble suitability mapping methods.