An objective methodology for potential vegetation reconstruction constrained by climate

Reconstructions of modern Potential Natural Vegetation (PNV) are widely used in climate modelling and vegetation survey as a starting point for studies (historical changes of land-use, past or future vegetation distribution modelling, etc.). A PNV distribution is often related to vegetation models, which are based on empirical relationships between vegetation (or pollen data in paleoecological studies) and climate. Vegetation models are used to directly simulate a PNV distribution or to correct vegetation types derived from remotely-sensed observations in human-impacted regions. Consequently, these methods are quite subjective and include biases from models. This article proposes a new approach to build a high-resolution PNV map using a statistical model.As vegetation is a nominal variable, our method consists in applying a multinomial logistic regression (MLR). MLR build statistical relationships between BIOME 6000 data covering Europe and several climatological variables from the Climate Research Unit (CRU).The PNV reconstructed by MLR appears similar to those reconstructed from remotely-sensed data or simulated by a vegetation model (BIOME 4) except in southern Europe with the establishment of warm-temperate forests. MLR produces a realistic PNV distribution, which is the closest to BIOME 6000 data and provides the vegetation distribution in each grid-cell of our map. Moreover, MLR allows us to compute an uncertainty index that appears as a convenient tool to highlight the regions lacking some data toimprove the PNV distribution. The MLR method does not suffer any dynamic biases or subjective corrections and is a fast and objective alternative to the other methods. MLR provides an independent reference for vegetation models that is entirely based on vegetation and climatological data.

► We describe a statistical method to reconstruct potential vegetation from climatological data.
► We compare our results with recent studies and a vegetation model output.
► Our final map appears the most realistic in good agreement with data providing advantages by the use of probabilities.
► Our method appears as a fast, useful and robust tool in vegetation modelling.