This study evaluates the applicability of TESSERA embeddings for identifying dominant species in highdiversity steppe communities. Remote sensing has traditionally been effective for monocultures and forests (accuracy 85-95%), but yields poor results in species-rich grasslands due to spectral mixing. The research utilized field data from 87 plots and 128-dimensional TESSERA embeddings generated from annual time series of Sentinel-1 and Sentinel-2. Random Forest models demonstrated high efficiency for three of the four analyzed species (Festuca valesiaca Schleich. ex Gaudin, Stipa lessingiana Trin. & Rupr, Elymus repens (L.) Gould) with ROC-AUC >0.83, substantially outperforming traditional methods (R²≤0.4). Spatial analysis confirmed the ecological interpretability of predictions. Results open opportunities for cost-effective biodiversity monitoring across large territories.

Strelnikov I.I., Ostapko V.M., Ibatulina Yu.V.