Climate Change Atlas ‐ Dominant vegetation in the Hawaiian Islands

Taken from the DMP:

"Georeferencing of plots and determination of site conditions using GIS. Accuracy of plot abundance values assessed with vegetation maps, satellite imagery, and expert consultation. Ten native and five invasive species of trees, shrubs and ferns were selected based on their potential ecological importance in communities, as well as how much field data was available to analyze for this study.

For baseline climate condition variables, we used surface temperature and rainfall estimates (30 year baseline period, 1978–2007) developed by Giambelluca et al. (2013) with grid cell resolution of 250m. To estimate climate change responses, we used future projected climate variables developed by the International Pacific Research Center through dynamical downscaling of climate changes based on Coupled Model Inter-comparison Project 3 (CMIP3) Special Report on Emissions Scenarios (SRES) A1B for end of century (based on Zhang et al. 2012). Plant species abundance models were developed using four baseline predictors with matching future projected climate variables: 1) simplified pioneer substrate (developed from geology maps by Price et al. (2012)) as a proxy for primary succession; 2) mean annual temperature (bio1); 3) wet season rainfall (bio18); and 4) dry season rainfall (bio19). Environmental variables were resampled to a consistent projection (WGS1984) with 250m grid cell resolution.

We examined data distribution and data gaps, and correlations between variables were assessed using a Pearson correlation coefficient. To estimate relative importance of predictors, we used species-specific response curves and variable importance plots. Individual model fit was assessed using residuals against fitted values plots, null and residual deviance values, and model comparisons using relative Akaike information criterion (AIC). We examined predictive performance of modeling approaches by first using 10 fold cross-validation then area under the curve (AUC) with an independent evaluation set.

Final models were built with a continuous second degree polynomial GLM model with second degree polynomial substrate variable removed, and no interactions. We used a 70/30 (train/test) random split using an evaluation data set that was independent of the training data, with 20 replicates. Multiple test statistics were used for validating model responses and a strong assessment of model performance. For final models in baseline climate conditions, we used AUC and Spearman rank correlation coefficient (ρ) values."