FLYING FOR POLLINATORS. Can UAV-based vegetation height and topography help predict pollinator abundance and diversity using machine learning?

Abstract

In this thesis, semi-natural grassland features are mapped using UAV-based remote sensing data to predict pollinator abundance and distribution with the Random Forest machine learning algorithm and linear regression. “The Height variation hypothesis” states that higher heterogeneity in the vegetation corresponds to higher pollinator diversity. This hypothesis is tested by using vegetation heterogeneity quantified by Rao’s Q CHM, and adding, for example, topographic features derived from UAV-based LiDAR data. The addition of predictors such as NDVI and coefficient of variation for digital terrain models to linear regression models shows an increase in the model's predictive power. Rao’s Q CHM's ability alone explains variations in pollinator abundance by only 1%. Rao's Q CHM is, however, a stronger predictor for Shannon's diversity and alone explains around 10% of the variation in Shannon’s diversity. The model's explanatory power rises with the addition of terrain and vegetation indices, and using a best subset of predictors, the best model explains 44.8% (Random Forest) and 69.0% (linear regression) for abundance and Shannon’s diversity, respectively. Making HVH a useful metric for predicting pollinator abundance and diversity when combined with other indices in semi-natural grasslands.