Contents lists available at ScienceDirect Ecological Indicators journal homepage: www.elsevier.com/locate/ecolind Using orthoimages generated from oblique terrestrial photography to estimate and monitor vegetation cover Luke Wallace a,b, ⁎ , Daisy S. Saldias a , Karin Reinke a,b , Samuel Hillman a,b , Bryan Hally a,b , Simon Jones a,b a School of Science, RMIT University, Melbourne, Australia b Bushfire and Natural Hazards Cooperative Research Centre, Melbourne, Australia ARTICLE INFO Keywords: Percent vegetation cover Orthoimage Structure from motion Object based image analysis Prescribed burning ABSTRACT Percent vegetation cover is important variable used in understanding ecosystem processes, vegetation health and productivity. Downward looking images captured using a handheld camera have been demonstrated as a viable option for rapidly capturing in situ information to assess vegetation cover. This technique, however, is prone to perspective distortions biasing cover estimates towards taller vegetation elements. In this paper we present a new approach to generate imagery for use in vegetation cover estimation utilising multiple overlapping pho- tographs and structure from motion algorithms to produce a 3D point cloud representation of the target plot. This point cloud is then converted into an orthoimage consisting of four bands—red, green, blue and vegetation height—which is free from perspective distortions. The approach is trialled in two Eucalypt forests in South Eastern Australia to produce an estimate of change in cover of all vegetation elements following a prescribed burn. Orthoimages are generated with 2.5 mm resolution and classified using object-based image analysis and random forests into broad vegetation and fuel classes. Utilising this approach an overall classification accuracy of 81% is achieved with the resulting estimates of cover agreeing with visual point based interpretation to within 6% across all classes. 1. Introduction Percent Vegetation Cover (PVC) is defined as the percentage of the ground surface covered by vegetation elements from a vertical view point (Purevdorj et al., 1998). Understanding PVC and its variation is important as it is a key indicator of ecosystem processes, health and productivity (Liu and Treitz, 2016), as well, as a key variable used in understanding fuel hazard in fire prone environments (Gould et al., 2011). Traditional methods for measuring cover and composition of near-ground vegetation have been based on visual estimation techni- ques (Kennedy and Addison, 1987). However, such methods are sub- jective, prone to observer bias and not repeatable (Kercher et al., 2003; Baxendale et al., 2016; Spits et al., 2017). Point-intercept sampling techniques represent the primary objective methodology currently used to collect quantitative information de- scribing percent vegetation cover in the field (Bonham, 2013). These techniques estimate ground or understory vegetation cover through the proportion of times vegetation intercepts a vertical pointing device (laser pointer, thin rod or cross hairs) at a zenith of zero across a grid or transect (Held et al., 2015). A drawback of this technique often cited in the literature is that it is requires significant effort (labour and time) in order to create a statistically significant sample. Recently, downward-looking digital images collected over small plots, combined with image classification approaches such as object- based image analysis (OBIA) and rule based decision or machine learning, have proven successful in objectively quantifying percent vegetation cover in a range of environments (Luscier et al., 2006; Laliberte et al., 2007; Baxendale et al., 2016; Malenovskỳ et al., 2017). These approaches extract vegetation cover utilizing the information available in imagery; both spectral information contained in a pixel, and geometric information contained in objects representing con- tiguous groups of similar pixels (Luscier et al., 2006). Through the application of machine learning approaches on a set of training objects, the methodology has been shown capable of accurately distinguishing cover types such as green vegetation, senescent vegetation and bare soil/rock (Liu and Treitz, 2016; Malenovskỳ et al., 2017). The quality of the results obtained using these approaches have allowed them to be employed for validation of satellite based vegetation classifications (Malenovskỳ et al., 2017). A key requirement of the terrestrial digital cover approach is that https://doi.org/10.1016/j.ecolind.2018.12.044 Received 17 February 2018; Received in revised form 16 December 2018; Accepted 23 December 2018 ⁎ Corresponding author at: School of Science, RMIT University, Melbourne, Australia. E-mail address: luke.wallace2@rmit.edu.au (L. Wallace). Ecological Indicators 101 (2019) 91–101 1470-160X/ © 2018 Elsevier Ltd. All rights reserved. T