School of Agriculture, Food and Ecosystem Sciences - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemImproving species distribution models using extreme value theory and climate dataset ensembles( 2020)The development of climate datasets at fine spatial and temporal scales has commonly been driven by the need to better understand vegetation distributions and ecological systems. While a wide range of global, national and regional climate datasets have been developed over the last two decades, they are rarely compared directly in the ecological literature. This thesis evaluates a range of climate interpolation techniques and investigates how the spatial and temporal characteristics of climate datasets may be utilised to improve the predictive performance of plant species distribution models (SDM). A series of spline-based and geostatistical methods for interpolating temperature variables are first compared across Victoria, southeast Australia. Secondary predictors (thermal remote sensing data and local topographic indices) which indirectly capture mesoscale microclimate and cold air drainage regimes were found to improve monthly mean minimum temperature interpolations by up to 39%. Thermal remote sensing data only reduced root mean square error (RMSE) by up to 6% for maximum temperature across Victoria and was most effective during the summer months. The interpolation methods used in southeast Australia were subsequently transferred to the Royal Himalayan Kingdom of Bhutan to validate their effectiveness in a novel climate. In Bhutan, the predictive performance of minimum temperature interpolations was also improved considerably (up to 23% reduction in RMSE) when using thermal remote sensing data and local topographic indices as spatial covariates. Thermal remote sensing data also reduced the RMSE for maximum temperature interpolations by up to 16% in Bhutan. Interannual variability of climate extremes were used to evaluate how the temporal characteristics of climate may be used to improve the predictive performance of SDMs. Generalised Extreme Value (GEV) distributions were fitted to monthly climate data to generate variables which account for the skewed distribution of extremes. Models incorporating interannual variability (drawn from a range of expected return intervals) improved predictive performance compared to models using seasonal extremes only for 28 of 37 species assessed. Iteratively fitting models using alternate expected return intervals typically acted on the leading and trailing edges of current distributions, indicating that such methods may be useful for model calibration and characterising climate-driven source-sink population dynamics. The impact of spatial disparities in climate on the predictive performance of plant SDMs was evaluated using three distinct datasets developed for Victoria as part of this research, in addition to two global datasets (WorldClim v1 and v2). Individual models were compared against one another and as ensembles to explore the potential for alternate predictions to complement one another. The Victorian datasets demonstrated a significant improvement over the original WorldClim dataset (up to 17.3% mean increase in D2) and trended towards an improvement relative to WorldClim v2; however, no significant differences were found when comparing the alternate Victorian datasets. Multi-model ensembles achieved a mean increase of up to 13.8% and 29.2% in D2 relative to individual models when using regional and global datasets, respectively. Ensembles provide a pragmatic method to improve the predictive performance of SDMs and allow a trade-off between the uncertainties and potential biases embedded in competing climate datasets.