Background: Satellite remote sensing of vegetation, its structure, functional and physiological properties is essential for studying and monitoring ecosystems. However, optical and multi-temporal satellite data for the tropical Andes is limited by the presence of clouds and requires processing due to the rough topography. Unmanned Aerial Vehicles (UAV), equipped with low cost RGB cameras offer a cheap alternative, since they can provide data of high spatial and temporal resolution, hardly affected by clouds. Using UAV images, 3D models of the vegetation surface can be reconstructed using Structure from Motion (SfM) algorithms. Features extracted from these 3D point clouds can be used to estimate the space-filling by vegetation and can provide additional information on forest structure.

Objective: Here, we assess the usefulness of 3D features extracted from UAV sensed 3D point clouds to predict vegetation structure.

Methods: We acquired UAV images of a complex structured Andean landscape along a gradient from monoculture crops of coffee and cacao, over diversified agroforestry systems to conserved Andean forest of the Yariguíes National Natural Park. We characterized vegetation structure in the midstory and overstory layers over 34 plots using the Point Centered Quarter Method (PCQM) and hemispherical photography. We estimated tree density, height, basal area, leaf area index (LAI), canopy closure and above-ground biomass (AGB). From 3D reconstructions we generated canopy height models (CHM) and derived features based on the height of the points, like maximum height (Hmax), height distributions, skewness, entropy, vertical complexity index (VCI), surface area and volume spanned up by the CHM. Finally, we used different univariate and multivariate models to predict vegetation structure from these 3D features.

Results: Overall, vegetation structure derived from 3D features allowed to discriminate between agroforestry plantations of coffee, cacao and forest. Using a redundancy analysis, we found that 71% of the variance of vegetation structure can be explained by 3D features. Furthermore, multiple linear models showed significant relationships of predictor features like Hmax, VCI and surface area with on ground variables such as AGB, height in the mid- and overstory, also including density of shade trees.

Conclusion: 3D point clouds derived from UAV sensing using standard optical sensors are useful to predict stand structure and provide an opportunity to study the horizontal and vertical structure of vegetation. In the future, these 3D features could be contrasted with biophysical characteristics and biological inventories of forests and strengthen biodiversity studies using UAV remote sensing.