Global mapping of potential natural vegetation: an assessment of Machine Learning algorithms for estimating land potential

Abstract

Potential Natural Vegetation (PNV) is the vegetation cover in equilibrium with climate, that would exist at a given location non-impacted by human activities. PNV is useful for raising public awareness about land degradation and for estimating land potential. This paper presents results of assessing Machine Learning Algorithms (MLA) for operational mapping of Potential Natural Vegetation (PNV). The following MLA were considered: neural networks (nnet package), random forest (ranger), gradient boosting (gmb), K-nearest neighborhood (class) and cubist. Three case studies were considered: (1) global distribution of biomes based on the BIOME 6000 data set (8057 modern pollen-based site reconstructions), (2) distribution of forest tree taxa in Europe based on detailed occurrence records (1,546,435 ground observations), and (3) global monthly Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) values (30,301 randomly-sampled points). A stack of 160 global maps representing biophysical conditions over land, including atmospheric, climatic, relief and lithologic variables, were used as explanatory variables. The overall results show that random forest gives the overall best performance. he highest accuracy for predicting BIOME 6000 classes (20) was estimated to be between 33% (with spatial Cross Validation) and 68% (simple random subsetting), with the most important predictors being total annual precipitation, monthly temperatures and bioclimatic layers. Predicting forest tree species (73) resulted in mapping accuracy of 25%, with the most important predictors being monthly cloud fraction, mean annual and monthly temperatures and elevation. Regression models for FAPAR (monthly images) gave an R-square of 90% with most important predictors being total annual precipitation, monthly cloud fraction, CHELSA bioclimatic layers and month of the year, respectively. Further developments of PNV mapping could include using GBIF records to map global distribution of plant species at different taxonomic levels. This methodology could also be extended to dynamic modeling of PNV, so that future climate scenarios can be incorporated. Global maps of biomes, FAPAR and tree species at 1 km spatial resolution are available for download via http://dx.doi.org/10.7910/DVN/QQHCIK.

Biomes of the world under climate change scenarios: increasing aridity and higher temperatures lead to significant shifts in natural vegetation

“…To avoid issues with harmonization of nomenclature between biomes, Harrison (2017) provide a standardized classification legend that can be globally applied (32 biomes in total) and a megabiome classification legend (8 megabiomes in total). While the megabiome system implies a necessary loss of information due to generalization, the original standardized classification system devised by Harrison (2017) has been considered too detailed and location-specific to be used for global modeling (Hengl et al, 2018). We adopted the 20 classes (Fig.…”

“…In contrast, for the hard class map we used the probability maps to calculate a per-pixel confidence metric. Contrary to Hengl et al (2018), we chose not to use the per-pixel entropy (Shannon, 1948) but the margin of victory (Calderón-Loor et al, 2021) The margin of victory is defined as the difference between the first and the second highest class probability value in a given pixel. Potential values in this case would go from 0 (i.e.…”

“…While for data-model comparison we used the original 20 classes classification system from Hengl et al (2018), to compare the model outputs we translated the classes in the IUCN Global Ecosystem Typology (Keith et al, 2020). This system classifies biomes based on functional characteristics and their structural role in the ecosystems rather than on climate, species distribution or vegetation patterns.…”

“…In the case of Zevallos and Lavado-Casimiro (2022), they used a Random Forest model and achieved higher levels of accuracy; however, they trained it on a smaller (6 bioclimatic) set of predictor variables, used a 80:20 train test split and did not use any spatial partitioning. Considering the huge differences in accuracy in the results obtained by Hengl et al (2018) in their Random Forest model with and without spatial partitioning, there's the risk that predictions from Zevallos and Lavado-Casimiro (2022) may have been optimistic; on the other hand, their analysis is focused on just one country and not on a global scale, so it is still possible to reach high levels of accuracy on a limited study area.…”

“…Two common approaches here are: (1) use actual ground observations i.e. monitoring stations to fit spatiotemporal models (Hengl et al, 2018), and (2) use complete Earth observation data cubes and then basically all pixel combinations to visualize and model relationships (Mahecha et al, 2020). An advantage of the data-based modeling is that it is often computationally less demanding than process-based modeling and it can be extended by adding more covariates (Beigaitė et al, 2022).…”