Mapping Floral Resources in Montane Landscapes Using Unmanned Aerial Systems and Two-step Random Forest Classifications

Abstract

Monitoring floral biodiversity is a critical step in understanding terrestrial ecosystems. However, manual methods to quantify flowering vegetation are costly in time and personnel. In large landscapes, these limited methods may not capture the spatial and temporal variation of floral resources. Recent advances in sensors and unmanned aerial vehicle (UAV) platforms offer opportunities to characterize the dynamic distribution of floral resources at the landscape level. In this study, UAV imagery and a multistep machine learning classification analysis were used to quantify floral resources in nonagricultural environments, where topography, vegetation, and inflorescence size were variable. Seven flowering species covering an area of 2 138 m² were classified throughout our study, equaling 0.5% of the overall landscape. We determined the period of flowering for important species based on the temporal changes of the floral area classified from UAV images. Models performed well considering the extreme rarity of flowers in the UAV images. The flower class in the land cover classification models performed well with an average sensitivity of 0.77 and average specificity of 0.99. Individual flower classes also performed well with the majority of flower classes receiving sensitivity and specificity values of over 0.90. The use of UAVs is a feasible method for characterizing floral resources in nonagricultural settings. Classifications would benefit from a more robust and comprehensive UAV and floral resource sampling plan, to better characterize the variability of floral resources in UAV imagery.