Modeling monthly mean temperatures for the mountain regions of Taiwan by generalized additive models

Taiwan is a mountainous island with a mountain flora that has a high degree of endemism. In order to facilitate the assessment of Taiwan’s mountain flora integrity under various climate change scenarios, the objective of this study was to construct the monthly mean temperature models for the island’s mountain regions (elevation ≥ 1000 m a.s.l.) based on data from 43 meteorology stations. We used a generalized additive modeling approach to construct the models, with elevation, easting, and northing co-ordinates, as the predictors. The effects of elevation were modeled mainly by linear or second-order polynomial functions, whereas the effects of location were modeled nonparametrically by thin plate regression splines. We used the leave-one-out cross-validation to validate the models, with mean errors, mean absolute errors, and root-mean-square of normalized errors as measures of generalization reliability. Spatial interpolations were then performed at a 1-km resolution. Judging by the adjusted R2 and the proportion of deviance explained, all fitted models performed well. Monthly mean temperatures during the period of April to August decreased linearly with increasing altitude. The observed lapse rates were between 4.9 and 5.6 °C/km, with higher rates during the late spring and summer, and lower rates in the fall and early winter. For the first 3 calendar months, the lapse rates were between 3.2 and 3.6 °C/km, but rose by 0.4–0.5 °C/km for each km of increase in elevation. The three performance measures suggested that the monthly predictions were reliable. The predictions were more reliable for regions from which most of the data were collected and less reliable at the edges of the mountain regions. Model analyses and spatial interpolation results suggested that the models realistically reflected the influence of the geographic location, terrain characteristics, dominant regional climatic features, and their interactions. Predictions also confirmed the existence of the Massenerhebung effect in central Taiwan. Our models also suggest that the alpine florae of the central and the southern regions could be more vulnerable to future winter temperature increases than that of the northern region. Although this study suffered from the lack of data from remote areas, a relatively dense network with quality stations and a good representation along the altitudinal gradient enabled this study to develop reliable models over a complex terrain. Future model improvements can be achieved by placing temporary monitoring points in less represented areas and then spatially interpolating between the temporary stations and the permanent ones.