Projected vegetation redistribution: Australia - 9second gridded projection to 2050, pre-clearing extents of 77 Major Vegetation Sub-groups using kernel regression with GDM-scaled environments for Vascular Plants (GDM: VAS_v5_r11; CMIP5: MIROC5 RCP 8.5)

***UPDATED***\nThis collection contains 9-second gridded datasets (ESRI binary float format in GDA94) showing the projected future (2050-centred) potential vegetation redistribution of 77 Major Vegetation Sub-groups (MVS classes) for continental Australia based on their pre-clearing distribution patterns and correlation with baseline ecological environments (c.1990 climates, substrate and landform). The pre-clearing vegetation patterns and classification derive from version 4.1 of “Australia - Estimated Pre1750 Major Vegetation Groups - NVIS Version 4.1 (Albers 100m analysis product)” developed by the Australian Government Department of the Environment and collaborating State agencies. A kernel regression was used with c.155,000 locations of training classes for the 77 MVS classes attributed with 17 GDM-scaled environmental predictors for Vascular Plants representing baseline ecological environments. The training class data input to the kernel regression is provided with this package. The GDM-scaled environmental predictors are available with the “VAS_v5_r11” data package. Using the 1990 baseline training MVS class data, and without constraining the prediction to pre-existing map boundaries, the kernel regression projected to 2050 the distribution of the 77 Major Vegetation Sub-groups using 2050-centred (30 year average) future climates derived from the MIROC5 global climate model for the emission scenario defined by a representative concentration pathway of 8.5. The kernel regression generates unconstrained probabilities varying in the range from 0 and up to 1 for each of the 77 MVS classes. \n\nThe data are provided as 9-second (approximately 250m), ESRI binary float grid format in GDA94. Each class is denoted “UNCON###”, where the number refers to the code originally assigned to that MVS class by the supplier. A lookup table linking the MVS classes to the output codes and descriptive title is provided. Generalised representations of the vegetation classes derived from the individual class probabilities as the maximum probability in any grid cell are provided separately (see related information). \n\nThere are three dataset packages in this series: 1) 1990 predictions of MVS classes; 2) 2050 CanESM2 RCP 8.5 predictions of MVS classes; 3) 2050 MIROC5 RCP 8.5 predictions of MVS classes. \n\nThis dataset series and its use is described in the AdaptNRM Guide “Helping biodiversity adapt to climate change: a community-level modelling approach”, available online at: www.adaptnrm.org\nLineage: Predictive models of vegetation classes were derived using the two-step process originally developed for individual species distribution modelling with GDM (described in Elith et al. 2006). The first step uses a Generalised Dissimilarity Model (GDM) of vascular plants (VAS_V5_R11) to derive a set of scaled environmental variables for current (e.g. 1990 baseline) and future climates (e.g. 2050). The second step applies this data in a kernel regression to predict each vegetation class using training data derived from the pre-clearing mapping of 77 Major Vegetation Sub-groups. The training data comprised c.155,000 locations defined by randomly sampling within each vegetation class, proportional to their observed areal extent. These locations were then attributed with the baseline values of the GDM-scaled environmental variables. Separate kernel regressions were then run for the baseline and future climate scenarios using the baseline training data. In this way, the future distribution of each vegetation class was projected based on its affinity with present-day ecological environments. \n\nAt any location (grid cell), the kernel regression considers the surrounding relative density of training sites of the target vegetation class as a proportion of other types and generates a predicted probability for that class for the focal grid cell. A probability surface for the predicted proportions, varying from 0 to 1, is generated for each of the 77 mapped Major Vegetation Sub-groups. This method is infrequently used in ecology because of the need first to scale and reduce the dimensionality of the predictor variables (Lowe 1995). The GDM step reduces dimensionality (by choosing the variables to use) and scales the predictor variables using similarity-decay functions which equate to the multivariate distances expected by kernel regression. The kernel regression thus incorporates interactions by modelling ecological distances and vegetation class densities within a truly multivariate predictor space, with no assumption of additivity. \n\nKernel regression aims to optimise model performance in terms of the accuracy of predictions at any single location according to the area predicted for each class. The predicted proportions of common vegetation types are typically greater than for rarer vegetation types. Therefore, cell by cell, the class with the maximum probability selected to represent spatially varying vegetation class mosaics on a single map (essentially one dimension) will often be the common type, at the expense of locally rare and nationally rare types. Therefore, the best way to view the results, and to inform planning, is the individual probability surfaces. These properly reveal where the rarer vegetation types have a likelihood of persistence. Higher probabilities associated with other vegetation types at the same location can be viewed as a measure of the extent to which those other vegetation types may compete. However the outcome, at least in the medium term, may be more driven by the extant occurrence of ecosystems and their ability to persist under marginal conditions. \n

【更新版】本数据集集合包含以GDA94坐标系存储的9秒分辨率栅格数据集（ESRI二进制浮点格式），展示了澳大利亚大陆77个主要植被亚组（Major Vegetation Sub-groups，MVS类）以清理前分布格局为基础，结合基线生态环境（约1990年气候、基质与地形）的相关性，所预测的以2050年为中心的未来潜在植被再分布情况。

本次研究使用的清理前植被格局与分类体系，源自澳大利亚环境部及合作州级机构开发的《澳大利亚——1750年前估算主要植被群——NVIS版本4.1（Albers 100米分析产品）》（"Australia - Estimated Pre1750 Major Vegetation Groups - NVIS Version 4.1 (Albers 100m analysis product)"）第4.1版。

研究采用核回归（kernel regression）方法，以约15.5万个训练样点为基础，针对77个MVS类进行建模；这些训练样点对应77个维管植物（Vascular Plants）的17个经GDM缩放的环境预测变量，用于表征基线生态环境。

本次核回归所用的训练样点数据已随本数据包一同提供；经GDM缩放的环境预测变量可从"VAS_v5_r11"数据包获取。

本研究采用1990年基线MVS类训练数据，且未将预测结果约束于现有地图边界范围内；通过使用以2050年为中心的30年平均未来气候数据（源自MIROC5全球气候模型，对应典型浓度路径8.5（Representative Concentration Pathway 8.5，RCP8.5）排放情景），核回归模型完成了77个主要植被亚组的2050年分布预测。

核回归模型为77个MVS类分别生成取值范围为0至1的无约束概率值。

本数据集以GDA94坐标系存储的9秒（约250米）分辨率ESRI二进制浮点栅格格式提供。每个MVS类均以"UNCON###"命名，其中###代表供应商最初为该MVS类分配的代码。本数据包附带了将MVS类与输出代码、描述性标题相关联的查找表。此外，本数据包还单独提供了基于各栅格单元内最大概率值，由单类概率生成的植被类群广义表征结果（详见相关信息）。

本数据集系列共包含3个数据包：1）MVS类的1990年预测结果；2）基于CanESM2模型、RCP8.5情景的2050年MVS类预测结果；3）基于MIROC5模型、RCP8.5情景的2050年MVS类预测结果。

本数据集系列及其使用方法详见AdaptNRM指南《助力生物多样性适应气候变化：一种社区级建模方法》（"Helping biodiversity adapt to climate change: a community-level modelling approach"），可通过网址www.adaptnrm.org在线获取。

数据集溯源：植被类群预测模型采用两步法构建，该方法最初由Elith等人于2006年提出，用于基于GDM的单物种分布建模。

第一步，利用维管植物的广义相异模型（Generalised Dissimilarity Model，GDM，VAS_V5_R11），为当前（如1990年基线）和未来（如2050年）气候生成一组缩放后的环境变量。

第二步，将该数据应用于核回归模型，利用源自77个主要植被亚组清理前分布制图的训练数据，对每个植被类群进行预测。

训练数据包含约15.5万个样点：在每个植被类群内按其实际分布面积比例进行随机采样生成样点，随后为每个样点赋予经GDM缩放的基线环境变量值。

随后，利用基线训练数据分别针对基线气候情景与未来气候情景运行核回归模型。通过该方式，每个植被类群的未来分布可基于其与当前生态环境的亲和性进行预测。

在任意栅格单元位置，核回归模型会将目标植被类群的周边训练样点相对密度（以其他类群样点占比为参照）纳入考量，并为该栅格单元生成对应植被类群的预测概率。针对77个已制图的主要植被亚组，模型均会生成一张取值范围为0至1的预测比例概率表面。

该方法在生态学研究中应用较少，原因在于其需首先对预测变量进行缩放并降维（Lowe，1995）。GDM步骤通过筛选待使用变量实现降维，并利用相似性衰减函数对预测变量进行缩放，该函数与核回归所需的多元距离相匹配。

因此，核回归模型可在真正的多元预测变量空间中，通过对生态距离与植被类群密度进行建模来纳入交互效应，且无需假设变量间具有可加性。

核回归模型旨在针对每类植被的预测面积，优化任意单点预测的准确性以提升模型整体性能。通常而言，常见植被类群的预测占比会高于稀有植被类群。

因此，若在单张地图（本质为单维度空间）中按栅格单元选取最大概率类群以表征空间异质性的植被类群镶嵌格局，最终结果通常会以常见类群为主，而牺牲局域稀有乃至全国性稀有类群。因此，解读本数据集结果并为规划提供参考的最佳方式，是查看单类植被的概率表面：这类结果可准确展现稀有植被类群具备存续可能性的区域。

同一栅格单元内其他植被类群的较高概率值，可用于衡量这些类群的潜在竞争程度。不过至少在中期尺度上，植被分布的最终结果可能更多取决于当前生态系统的存续现状，以及它们在边缘生境下的存续能力。