AMP_V2_R2: Generalised dissimilarity model of compositional turnover in amphibian species for continental Australia at 9 second resolution using ALA data extracted February 2014

Compositional turnover patterns in amphibian species across continental Australia were derived using Generalised Dissimilarity Modelling (GDM). These models use best-available biological data extracted from the Atlas of Living Australia current to 27th February 2014 and spatial environmental predictor data compiled at 9 second resolution (with novel climate seasonality predictors, undersampling covariates and >3 species aggregated per 9-second grid cell). The models were developed to underpin continental assessments of biodiversity significance and identify gaps in biological surveys. GDM is a statistical technique that models the dissimilarity in composition of species between pairs of surveyed locations, as a function of environmental differences between these locations. The compositional dissimilarity between a given pair of locations can be thought of as the proportion of species occurring at one location that do not occur at the other location (averaged across the two locations) - ranging from ‘0’ if the two locations have exactly the same species through to ‘1’ if they have no species in common. GDM effectively weights and transforms the environmental variables such that distances between locations in this transformed multidimensional environmental space now correlate, as closely as possible, with the observed biological compositional dissimilarities between these same locations. Once a GDM model has been fitted to the biological data from the sampled locations using environmental predictor data, it can be used to predict compositional dissimilarity values for sites lacking biological data, based purely on their mapped environmental attributes. For this purpose, a set of GDM-scaled environmental grids are produced for use in subsequent spatial assessments of biodiversity significance. This collection includes the source biological and environmental data, the GDM-fitted model, the GDM-scaled environmental predictors for the fitted-model which comprises substrate (constant) and 1990-centred climates, and a derived classification. Projections using past and future climates are not included here (available upon request). This model was used in the AdaptNRM series of reports by Williams et al. (2013) and Prober et al. (2014). \nLineage: 1. Biological data for amphibians were extracted from the Atlas of Living Australia (27th February 2014), at the species taxonomic rank (matched to the census of Australian fauna), excluding alien/introduced taxa. Unmatched or uncertain taxa were excluded. Locations with a geoaccuracy > +/-1000m were excluded; locations that lacked a geoaccuracy estimate were included. Approximately 1.5M site-pairs (of potentially millions) were generated using Biodiverse software (in groups of 0.0025 degrees latitude and longitude) and Perl scripts (courtesy S Laffan and D Rosauer) as a stratified random sample of 85 IBRA v7 continental Australia bioregions (85 x 85 strata) weighted by 10% for selection of site-pairs within bioregions. Only grid cells with at least 2 species occurrences represented were used. A response weight variable was generated as the sum of the number of species at the two sites in the pair, irrespective of species in common). Two independent sets of site pairs were generated of the same size and site-pair selection parameters – for model “training” and “test” (validation). 2. The biological data grouped into 0.0025 degrees of latitude and longitude (the same as the raster grid centroids) using Biodiverse software (Laffan et al, 2010) were exported and summarized to generate the number of (unique) species in each group, the number of original (unique) latitude and longitude per group, and the number of records (unique species latitude and longitude). Three “under-sampling” covariates were subsequently generated for inclusion as candidate covariate predictors in the GDM models. The covariate calculation is detailed in the report by Williams et al. 2010 (see link in related materials, https://publications.csiro.au/rpr/pub?list=BRO&pid=csiro:EP102983). 2. Environmental data were compiled from best available sources of geology, soil, landform (including DEM derivatives) and climate. Climate data were derived using ANUCLIM v6.1 software with the 1990-centred (30 year average) surfaces and version 3.1, 9 second digital elevation model for Australia. Climate predictors were generated as the monthly minimum or maximum values (including a range of seasonality predictors as described in Williams et al, 2012; IJGIS, 26:2009). The ratio of topographically-shaded and slope/aspect-corrected incoming shortwave radiation relative to the unshaded radiation on a horizontal surface was used to adjust both monthly radiation and maximum temperature. Details pertaining to these calculations are published in Reside et al. 2013 (see link in related materials http://www.nccarf.edu.au/publications/climate-change-refugia-terrestrial-biodiversity). 3. GDM models were fitted using .NET software v3.x (courtesy G Manion, NSW Office of Environment and Heritage), selecting environmental predictors using backward elimination by testing the partial contribution of each predictor and removing the least significant predictor until all predictors explained at least 0.05% of the model deviance in the presence of all other included predictors. These models excluded Geographic Distance between site pairs. The potential for a 4th spline to better define the shape of the predictors was tested selectively for the dominant predictors, using the model fit criterion of at least 0.05% additional partial deviance explained. Following these tests the significance of the predictors was again tested using backward elimination. Details of the resulting fitted model and the input data table are provided with the data package. 4. Transformed grids for the environmental predictor variables were generated for the final fitted model. The climatic predictors were replaced with past (1960-centred 75 year average) and six future (2050-centred 30 year averages) scenarios and the transformed grids generated also. The covariate predictors are not included in the set of transformed grids, and so assumed to have optimal values. Extrapolation error grids are also provided.

本数据集基于广义相异性建模（Generalised Dissimilarity Modelling, GDM）方法，推导得到澳大利亚大陆两栖动物物种的组成更替格局。本模型采用截至2014年2月27日的澳大利亚生物图集（Atlas of Living Australia）中公开的最优生物数据，以及分辨率为9弧秒的空间环境预测变量数据（包含新型气候季节节律预测变量、欠采样协变量，且每9弧秒栅格单元内聚合了至少3个物种的分布记录）。该模型旨在支撑大陆尺度生物多样性重要性评估，并识别生物调查的空白区域。

广义相异性建模是一种统计方法，以调查样点对之间的环境差异为自变量，对样点对间的物种组成相异性进行建模。给定样点对之间的组成相异性，可以理解为：一个样点中存在但另一样点中不存在的物种占比（以两个样点的平均值计），取值范围为0到1：当两个样点的物种组成完全相同时取值为0，当二者无共有物种时取值为1。GDM通过对环境变量进行加权与变换，使得样点在变换后的多维环境空间中的距离，能够尽可能贴合观测得到的对应样点间的生物组成相异性。当利用环境预测变量数据对采样样点的生物数据拟合得到GDM模型后，即可仅基于待预测样点的空间环境属性，对无生物调查数据的样点的组成相异性进行预测。

为此，本研究生成了一套经GDM标准化的环境栅格数据，用于后续的生物多样性重要性空间评估。本数据集集合包含原始生物与环境数据、拟合完成的GDM模型、适用于该拟合模型的经GDM标准化的环境预测变量（包含恒定不变的基质数据与以1990年为基准期的气候数据），以及派生得到的分类结果。本数据集未包含基于过去与未来气候情景的投影结果（可按需获取）。该模型已被Williams等人（2013）与Prober等人（2014）应用于AdaptNRM系列报告的相关研究中。

数据集生成谱系：
1. 两栖动物生物数据提取自2014年2月27日版澳大利亚生物图集，分类阶元限定为物种水平（匹配澳大利亚动物区系普查名录），剔除外来/引入类群，未匹配或分类地位存疑的类群亦被排除。地理定位精度大于±1000米的样点被剔除，无地理定位精度信息的样点予以保留。研究采用分层随机抽样方法，以澳大利亚临时生物地理区域划分第7版（IBRA v7）的85个大陆生物区为抽样层（共85×85个抽样层），按10%的权重在各生物区内选取样点对；使用Biodiverse软件与Perl脚本（由S Laffan与D Rosauer提供）生成约150万个样点对（潜在总量可达数百万），样点按0.0025度经纬度分组。仅保留至少包含2个物种分布记录的栅格单元。构建响应权重变量，其取值为样点对中两个样点的物种总数之和，无需考虑共有物种。基于相同的样点选择参数，生成两套独立的等规模样点对数据集，分别用于模型"training"与"test（验证）"。
2. 使用Biodiverse软件（Laffan等，2010）将生物数据按0.0025度经纬度（与栅格网格质心一致）进行分组，导出并汇总得到各分组内的（唯一）物种数、每个分组的原始（唯一）经纬度数量，以及记录数（唯一物种-经纬度组合数）。随后生成3个"欠采样"协变量，作为GDM模型的候选协变量预测变量。协变量的计算方法详见Williams等人2010年的报告（相关材料链接：https://publications.csiro.au/rpr/pub?list=BRO&pid=csiro:EP102983）。
2. 环境数据整合自最优公开数据源，涵盖地质、土壤、地形（包含数字高程模型（Digital Elevation Model, DEM）衍生变量）与气候数据。气候数据通过ANUCLIM v6.1软件生成，采用以1990年为基准期的30年平均气候表面，以及澳大利亚9弧秒版本3.1数字高程模型。气候预测变量由逐月最低或最高气温等指标构成（包含一系列气候季节节律预测变量，详见Williams等，2012；《国际地理信息科学期刊》，26卷：2009年）。采用地形遮蔽与坡度/坡向校正后的入射短波辐射与水平表面无遮蔽辐射的比值，对月辐射量与最高气温进行校正。相关计算细节已发表于Reside等人2013年的研究（相关材料链接：http://www.nccarf.edu.au/publications/climate-change-refugia-terrestrial-biodiversity）。
3. 使用.NET软件v3.x（由新南威尔士州环境与遗产办公室G Manion提供）拟合GDM模型，采用向后剔除法筛选环境预测变量：依次检验每个预测变量的偏贡献度，剔除最不显著的变量，直至所有保留的预测变量在与其他已纳入变量共同作用时，至少能解释模型偏差的0.05%。本模型未纳入样点对间的地理距离变量。针对主要预测变量，额外检验是否可通过第4次样条函数更好地刻画其变量形态，判定标准为额外解释至少0.05%的偏偏差。完成上述检验后，再次通过向后剔除法验证各预测变量的显著性。最终拟合模型与输入数据表的详细信息已随本数据集包一同提供。
4. 为最终拟合的GDM模型生成环境预测变量的变换栅格数据。将气候预测变量替换为过去（以1960年为基准期的75年平均）与6个未来情景（以2050年为基准期的30年平均），并分别生成对应的变换栅格。协变量预测变量未纳入本次变换栅格集，默认其取值为最优值。本数据集同时提供外推误差栅格数据。