Floristic vs. Dominant Classification Approaches Applied to Geospatial Modeling of Mixed and Broadleaf Forest Types in the Northwestern Caucasus (Russia)

The Caucasus Mountains are recognized as a global center of biodiversity but currently face significant risks of degradation due to intensified economic development and the effects of climate change. Forest inventory and mapping are essential for biodiversity conservation in the Caucasus region. Geospatial modeling is a common method of thematic mapping, but its reliability depends heavily on the initial classification of reference data used for model training. Modern vegetation science features various classification approaches, most of which were developed independently of digital mapping practices and are rarely assessed for their suitability in geospatial modeling. To fill this gap, we classified the same dataset of vegetation relevés from mixed and broadleaf forests in the northwestern Caucasus using two approaches, based on floristic and dominant concepts, and compared the predictive performance of geospatial models trained on these datasets. We considered multiple types of geospatial variables, including optical satellite imagery, a digital elevation model (DEM), and bioclimatic and soil features, to evaluate their informativeness for spatial differentiation of the resulting forest types and to identify optimal variable combinations for modeling via multistage feature selection. We trained several models using different variable sets and machine learning methods for both classifications and evaluated their accuracy via nested cross-validation. The forest types produced by the two approaches scarcely matched, and the selected variable sets for model training differed accordingly. Unexpectedly, bioclimatic and soil variables were more effective than DEM- and satellite-derived variables, despite their coarser spatial resolution. Floristic-based geospatial models outperformed dominant-based models in terms of forest-type separability and predictive accuracy. Therefore, a floristic classification approach may be preferable for forests with complex species composition, both ecologically and in terms of the reliability of geospatial modeling and the derived mapping results.