Three-dimensional thermal characterization of forest canopies using UAV photogrammetry

Clare Webster, Matthew Westoby, Nick Rutter, Tobias Jonas
2018 Remote Sensing of Environment  
10 Measurements of vegetation structure have become a valuable tool for ecological 11 research and environmental management. However, data describing the thermal 3D 12 structure of canopies and how they vary both spatially and temporally remain sparse. 13 Coincident RGB and thermal imagery from a UAV platform were collected of both a 14 standalone tree and a relatively dense forest stand in the sub-alpine Eastern Swiss 15 Alps. For the first time, SfM-MVS methods were used to develop 3D RGB and
more » ... thermal 16 point clouds of the two sites with point densities of 35,245 and 776 points per m 2 , 17 respectively, compared to 78 points per m 2 for an airborne LiDAR dataset of the same 18 area. Despite the low resolution of the thermal imagery compared to RGB photosets, 19 forest structural elements were accurately resolved in both point clouds. 20 Improvements in the quality of the thermal 3D data were gained through the 21 application of a distance filter based on the proximity of these data to the RGB 3D 22 point data. Vertical temperature gradients of trees were negative with increasing height 23 2009), energy balance (Musselman et al. 2013) and radiative transfer modeling 63 (Essery et al. 2008a), as well as search and rescue logistics (Rudol and Doherty 2008). 64 Remotely sensed data describing 3D forest structures have been retrieved using 65 airborne or terrestrial light detection and ranging methods (LiDAR; Kankare et al. 2013; 66 Liang et al. 2012; Lucas et al. 2008; Srinivasan et al. 2014; Wagner et al. 2008). LiDAR 67 data can be acquired across large (> 50,000 ha) areas in a series of repeat over-68 flights. However, the commission of LiDAR flights or data purchase can exceed 69 USD$20,000 per flight (Erdody and Moskal 2010), particularly when data at high 70 spatial and temporal resolutions are required. More recently, improvements in the 71 affordability and accessibility of lightweight unmanned aerial vehicle (UAV, or 'drone') 72 technology has facilitated low-cost methods of low-altitude (< 150 m flying height) 73 photographic and videographic data collection in a range of environments (e.g. Cohen 74 4 et al. 2005; Dandois and Ellis 2013; Faye et al. 2016; Morgenroth and Gomez 2014). 75 The deployment of lightweight fixed-wing or multi-rotor UAV systems with on-board 76 digital imaging sensors facilitates the collection of remotely sensed data at increasingly 77 high spatial and temporal resolutions. Further advances in the development of flight 78 planning software now facilitate GPS-guided flight repeatability. 79 The recent emergence of a new generation of digital photogrammetric and computer 80 vision-based algorithms for reconstructing 3D scene topography from 2D input 81 imagery, popularly known as 'Structure-from-Motion' (SfM) has revolutionized the field 82 of 3D data acquisition and analysis (e.g. Carrivick et al. 2016; James and Robson 83 2012; Snavely et al. 2008; Westoby et al. 2012), and originates from advances in the 84 computer vision community (e.g. Spetsakis and Aloimonos, 1991; Boufama et al., 85 1993; Szeliski and Kang, 1994). Unlike conventional photogrammetric techniques, 86 SfM methods identify matching features in overlapping digital images and use this 87 information as input to an iterative bundle adjustment which simultaneously solves for 88 the interior and exterior camera parameters and generates a sparse 3D point-cloud. 89 This process can be enhanced through the use of input imagery which has been 90 geotagged using GPS technology. SfM algorithms are commonly used in conjunction 91 with multi-view stereo methods (SfM-MVS) to increase 3D point densities, typically by 92 an order of magnitude or more (Carrivick et al. 2016; James and Robson 2012; 93 Westoby et al. 2012), whilst the addition of ground control points (GCP) with known 94 xyz positions in the scene facilitates the georegistration of SfM-derived 3D data. 95 A number of recent studies have employed SfM-MVS methods to derive 3D models of 96 forest canopy structure from RGB imagery acquired from UAVs (e.g. Dandois and Ellis 97 2010; Dandois and Ellis 2013; Mlambo et al. 2017). Example applications of SfM-MVS 98 157 processing to ensure accurate surface temperatures are calculated (e.g. Faye et al. 158 2016). Additionally, the collection of data across the forest stand during winter when 159
doi:10.1016/j.rse.2017.09.033 fatcat:3hzqrvvphnculahbwagw2qmjdi