Integration of UAV-sensed features using machine learning methods to assess species richness in wet grassland ecosystems

Abstract

Wet grasslands are crucial components of terrestrial ecosystems, known for their biodiversity and provision of ecosystem services such as flood attenuation and carbon sequestration. Given their ecological significance, monitoring biodiversity within these landscapes is of utmost importance for effective conservation and management strategies. This study, conducted in a wet grassland of Brandenburg, Germany, utilized unmanned aerial vehicles (UAVs) to facilitate the estimation of species richness by the integration of remotely sensed canopy features such as canopy height (CH), spectral data (Vegetation Indices, VI), and texture features (Gray-Level Co-occurrence Matrix, GLCM) using two machine learning methods (Partial Least Square regression (PLS) and Random Forest (RF)). Data was collected over two growing seasons under three different grass cutting regimes, employing multispectral sensors to capture detailed vegetation characteristics. The analysis revealed that the performance of the machine learning methods varied with the feature combinations. Models combining VI and GLCM features demonstrated the highest predictive accuracy, particularly in frequently cut grasslands, as indicated by higher R² values (up to 0.52) and lower root mean square errors (rRMSE as low as 34.9 %). RF models generally outperformed PLS models across different feature sets, with the CH + VI + GLCM combination yielding the best results. These findings underscore the potential of spectral and textural data to effectively capture the ecological dynamics of wet grasslands, providing valuable insights into biodiversity patterns.