Methods
For this project, only interpolated data was used. This data was generated by Mitchell and Jones for the 1901-2001 period based on the CRU-TS 3.1 dataset and updated to 2013 with CRU-TS 3.33 by Dr. Andreas Hamman of University of Alberta. This data is available as a downloadable software at http://www.ualberta.ca/~ahamann/data/climatesa.html, and can be used to generate monthly, seasonal and annual climate variables, needing only latitude, longitude, and elevation references (Hamann and Wang, 2005).
Using two different resolution files (36km2 and 8km2), seasonal variables were generated for the 1961 – 1990 period and for the year 2050, using an average of 15 AOGCMs under the RCP 4.5 concentrations trajectories adopted by IPCC in 2014 (Moss et al., 2008). To define a biome for each point of these files, the classification defined by Olsen et al., (2001) was used.
Grouping of regions
Using the 36 km2 resolution file, differences between the seasonal variables for the year 2050 and the 1961-1990 period were calculated. Using Euclidean distances and Ward clustering in the R Statistical Software, groups that show the type and level of affectation that each region is going to experience were formed.
Comparison of methods of modelling of distribution
To predict the changes in the distribution of biomes for the Neotropic, two methods were implemented and compared using the 8km2 resolution file: Linear Discriminant Analysis (LDA) from the MASS package (Venables and Ripley, 2002) for the R statistical software and Random Forest (RF) (Liaw and Wiener, 2002), also for R. First, the models were trained using a subsample of 1000 points for each biome class and the seasonal variables for the 1961 – 1990 period. Next, to compare these two methods in terms of accuracy, the trained models were used to reclassify the subsample and the total error for each model and a matrix of errors of misclassification were obtained. Finally, to assess the level of change in the suitable habitat of the biomes of the Neotropic for the future, the trained models were used to classify the complete datasets of seasonal variables for the 1961 – 1990 period and for the year 2050.
For this project, only interpolated data was used. This data was generated by Mitchell and Jones for the 1901-2001 period based on the CRU-TS 3.1 dataset and updated to 2013 with CRU-TS 3.33 by Dr. Andreas Hamman of University of Alberta. This data is available as a downloadable software at http://www.ualberta.ca/~ahamann/data/climatesa.html, and can be used to generate monthly, seasonal and annual climate variables, needing only latitude, longitude, and elevation references (Hamann and Wang, 2005).
Using two different resolution files (36km2 and 8km2), seasonal variables were generated for the 1961 – 1990 period and for the year 2050, using an average of 15 AOGCMs under the RCP 4.5 concentrations trajectories adopted by IPCC in 2014 (Moss et al., 2008). To define a biome for each point of these files, the classification defined by Olsen et al., (2001) was used.
Grouping of regions
Using the 36 km2 resolution file, differences between the seasonal variables for the year 2050 and the 1961-1990 period were calculated. Using Euclidean distances and Ward clustering in the R Statistical Software, groups that show the type and level of affectation that each region is going to experience were formed.
Comparison of methods of modelling of distribution
To predict the changes in the distribution of biomes for the Neotropic, two methods were implemented and compared using the 8km2 resolution file: Linear Discriminant Analysis (LDA) from the MASS package (Venables and Ripley, 2002) for the R statistical software and Random Forest (RF) (Liaw and Wiener, 2002), also for R. First, the models were trained using a subsample of 1000 points for each biome class and the seasonal variables for the 1961 – 1990 period. Next, to compare these two methods in terms of accuracy, the trained models were used to reclassify the subsample and the total error for each model and a matrix of errors of misclassification were obtained. Finally, to assess the level of change in the suitable habitat of the biomes of the Neotropic for the future, the trained models were used to classify the complete datasets of seasonal variables for the 1961 – 1990 period and for the year 2050.