Open AccessNew regionally modelled soil layers improve prediction of vegetation type relative to that based on global soil models
Author(s) -
Cramer Michael D.,
Wootton Lara M.,
Mazijk Ruan,
Verboom George Anthony
Publication year - 2019
Publication title -
diversity and distributions
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.918
H-Index - 118
eISSN - 1472-4642
pISSN - 1366-9516
DOI - 10.1111/ddi.12973
Subject(s) - edaphic , soil texture , vegetation (pathology) , environmental science , range (aeronautics) , biodiversity , spatial variability , soil science , ecology , spatial distribution , spatial ecology , physical geography , geography , soil water , biology , remote sensing , mathematics , statistics , medicine , pathology , materials science , composite material
Aim High‐resolution spatial soil data are crucial to species distribution modelling for fundamental research and conservation planning. Recent globally modelled soil layers (e.g. SoilGrids) have transformed distribution modelling, but may fail to represent regional soil characteristics accurately. We hypothesize that in the Cape biodiversity hotspot of South Africa, the use of global soil layers has led to underestimation of the importance of edaphic factors as determinants of species’ and vegetation distributions. We present a series of new, regionally modelled layers to address this deficiency. Location Greater Cape Floristic Region (GCFR, South Africa). Methods We georeferenced edaphic characteristics from literature and other sources and used boosted regression trees (BRT) to associate edaphic characteristics with spatially explicit topographic, climatic, soil texture and biotic variables. Multinomial BRTs were used to predict mapped vegetation types from the collated edaphic and other data. Results BRTs reliably predicted pH (92% of variance), Na (87%), K (85%), electrical conductivity (81%) and P (73%), but were less accurate for total N (55%) and total C (61%), for which data were sparser. Soil clay and pH values differed markedly in range and in spatial variation from those in SoilGrids. Using our new edaphic layers, we were able to accurately predict spatial distributions of vegetation types within the GCFR (multi‐class AUC = 0.96). The multinomial BRT predicted vegetation less well when based on SoilGrids data alone (AUC = 0.84). Main conclusions The more faithful representation of soil properties in our model is attributable both to its use of ca. 10‐fold more samples, and to its regional focus. Our model of edaphic characteristics captures important edaphic variability that is vital for understanding plant and consequently faunal distributions, with wide‐ranging conservation implications. Ongoing development of global syntheses of soil data requires more samples, especially in areas with high spatial heterogeneity and extreme edaphic conditions.