Floristic vs. Dominant Classification Approach Applied to Geospatial Modelling of Mixed and Broadleaf Forest Types in the North-Western Caucasus (Russia)

Geospatial modeling based on a combination of remote sensing data and machine learning methods currently is a common way for thematic mapping of forest ecosystems. However, the reliability of such mapping strongly depends on the initial classification of the reference data used for model training. In modern vegetation science, there are different classification approaches that coexist, but most of them were originally developed regardless of digital mapping practices, and were rarely analyzed on comparative suitability and benefits for geospatial modelling tasks. To fill this gap, we compared the performance of geospatial models for mixed and broadleaf forest types, provided by two different classification approaches, based on floristic and dominant concepts. Both classifications were applied on the same dataset of field plots, collected in the North-Western Caucasus region. We considered geospatial variables of different types, including optical satellite imagery, DEM and its derivatives, bioclimatic and soil features, to analyze their informativeness, and to obtain their optimal combinations via multistage feature selection. We trained the set of models with different variables’ sets and machine learning methods (Random Forest, CatBoost, LDA, and kNN) for both classifications, and evaluated their accuracy via nested cross-validation. As a result, forest types, obtained by different approaches, had very little matching, just like the compositions of the respective variables’ sets, selected for models’ training. Bioclimatic and soil variables were more effectie, than DEM-based and optical satellite-based ones, despite their coarser spatial resolution. Floristic-based geospatial models clearly outperformed the dominant ones in terms of forest types’ separability and potential predictive accuracy. Therefore, a floristic classification approach may be preferable for forests with complex species composition not only in terms of common ecological sense, but also in terms of reliability of geospatial modelling and its derivative mapping results.