A view from above: Unmanned aerial vehicles (UAVs) provide a new tool for assessing liana infestation in tropical forest canopies

1. Tropical forests store and sequester large quantities of carbon, mitigating climate change. Lianas (woody vines) are important tropical forest components, most conspicuous in the canopy. Lianas reduce forest carbon uptake and their recent increase may, therefore, limit forest carbon storage with global consequences for climate change. Liana infestation of tree crowns is traditionally assessed from the ground, which is labour intensive and difficult, particularly for upper canopy layers. 2. We used a lightweight unmanned aerial vehicle (UAV) to assess liana infestation of tree canopies from above. It was a commercially available quadcopter UAV with an integrated, standard three-waveband camera to collect aerial image data for 150 ha of tropical forest canopy. By visually interpreting the images, we assessed the degree of liana infestation for 14.15 ha of forest for which ground-based estimates were collected simultaneously. We compared the UAV liana infestation estimates with those from the ground to determine the validity, strengths, and weaknesses of using UAVs as a new method for assessing liana infestation of tree canopies. 3. Estimates of liana infestation from the UAV correlated strongly with ground-based surveys at individual tree and plot level, and across multiple forest types and spatial resolutions, improving liana infestation assessment for upper canopy layers. Importantly, UAV-based surveys, including the image collection, processing,


| INTRODUC TI ON
Tropical forests and their canopies play a crucial role in the maintenance and provision of unique biodiversity and essential ecosystem services to all life on Earth (Lowman & Schowalter, 2012;Ozanne et al., 2003). One of the most important ecosystem services that tropical forests provide is their ability to store and sequester carbon (Pan et al., 2011). Managing tropical forests for carbon sequestration, therefore, provides a key opportunity to mitigate some of the effects of climate change resulting from increasing atmospheric CO 2 concentrations (Canadell & Raupauch, 2008).
Lianas (woody vines) are conspicuous components of tropical forests, where they peak in abundance, biomass, and species richness (Schnitzer & Bongers, 2002). Lianas use the structural biomass of trees to deploy leaves in the canopy, thus investing relatively more resources in producing an extensive leaf canopy than in woody tissue. Lianas, therefore, disproportionately contribute to the forest canopy: liana leaves can comprise up to 30% of forest leaf area but only up to 5% woody stem biomass of tropical forests (van der Heijden, Schnitzer, Powers, & Phillips, 2013). Liana abundance and biomass have increased over the last few decades (Schnitzer & Bongers, 2011). Consequently, lianas have proliferated in the forest canopy, indicated by an increase in their contribution to leaf productivity as well as in the number of tree crowns infested (Ingwell, Wright, Becklund, Hubbell, & Schnitzer, 2010;Wright, Calderón, Hernandéz, & Paton, 2004). Partly due to their extensive canopies, lianas aggressively compete with trees, reducing tree growth (Ingwell et al., 2010;van der Heijden & Phillips, 2009), fecundity (e.g., Kainer, Wadt, & Staudhammer, 2014), survival (Ingwell et al., 2010;Phillips, Vásquez Martínez, Monteagudo Mendoza, Baker, & Núñez Vargas, 2005) and, consequently, forest biomass and net carbon uptake (van der Heijden, Powers, & Schnitzer, 2015). Lianas pose a particular problem for managed forests, where they can substantially hinder carbon sequestration and forest restoration (e.g., Marshall et al., 2017). Liana cutting is often used to enhance carbon uptake (Marshall et al., 2017;van der Heijden et al., 2015); however, this is expensive and labour intensive to perform over large extents.
The ability to identify where liana management would be most beneficial would therefore help target management of tropical forests.
Being able to accurately monitor the presence and degree of liana infestation in forest canopies over time and space is, therefore, important for determining whether and where liana impacts are high and/or may be increasing, particularly in managed tropical forests. Due to practical difficulties in accessing tropical forest canopies (Nakamura et al., 2017), assessing liana canopy infestation is traditionally done by ground-based surveys (e.g., van der Heijden, Feldpausch, Herrero, van der Velden, & Phillips, 2010). These are labour-and time intensive, and consequently limited in their spatial and temporal coverage, and lianas, therefore, remain understudied in tropical forests (Marvin, Asner, & Schnitzer, 2016). Furthermore, the stratified nature of tropical forests often limits the visibility of canopy and emergent tree crowns, affecting the reliability of ground-based estimates for them. As these larger trees tend to store and sequester the most carbon and, due to high light conditions in their crown, often harbour lianas (van der Heijden, Healey, , reliable assessment of liana infestation for top-of-the-canopy trees is especially important. Assessing lianas from a vantage point above the canopy should be feasible using remote sensing platforms which offer views of the canopy with much less obscuration by vegetation than possible from the ground (Nadkarni, Parker, & Lowman, 2011). However, satellite and many airborne platforms generally provide data too coarse in temporal or spatial resolution for this task, too expensive at very fine resolutions, and frequently suffer from cloud obscuration, especially in moist forests. Workarounds exist: using hyperspectral and LiDAR sensors, the Carnegie Airborne Observatory was able to accurately map heavy liana infestation across the forests of Panama (Marvin et al., 2016). The use of such sensors is very expensive and restricted to specialists, however, prohibiting their accessibility to the majority of researchers and forest managers. Furthermore, such remote sensing campaigns are typically carried out as one-time operations, so frequent monitoring is difficult (Xue & Su, 2017).
Unmanned aerial vehicles (UAVs) with sensors overcome most of the aforementioned limitations of remote sensing platforms (Cunliffe, Brazier, & Anderson, 2016). They can acquire remotely sensed data from relatively inaccessible environments, and thus are useful for measuring and (long-term) monitoring of forest canopies (Kachamba, Ørka, Gobakken, Eid, & Mwase, 2016;Paneque-Gálvez, McCall, Napoletano, Wich, & Koh, 2014;Zahawi et al., 2015;Zhang et al., 2016). Additionally, UAVs can capture data at even finer temporal and spatial resolutions than satellite and manned-airborne remote sensing (Nakamura et al., 2017). This is especially important because visually distinguishing lianas from trees requires ultra-fine resolution (mm-cm) image data: liana leaves grow among the leaves forest management where knowledge of the location and change in liana infestation can be used for tailored, targeted, and effective management of tropical forests for enhanced carbon sequestration (e.g., REDD+ projects), timber concessions, and forest restoration.

K E Y W O R D S
drone, drone ecology, liana infestation, lianas, remote sensing, tropical forest canopy, unmanned aerial vehicles, visual image interpretation of the trees in which they are located, becoming embedded in the canopy, and are also heterogeneous in nature, as lianas are phylogenetically and functionally highly diverse (Burnham, 2004;Gentry, 1991). Because their physical traits vary, leaves of a given liana species may look very different or very similar to other liana species or the tree in which they are located. Additional textural context, such as leaf shape or arrangement, could allow improved discrimination of liana leaves from tree leaves in cases where spectral discrimination is unfeasible. Thus, UAVs and the enhanced spatial resolution they offer are potentially effective for assessing canopy-level liana infestation. However, even though UAVs have been used to study other canopy phenomena (e.g., Zahawi et al., 2015); thus far, they have not been used to study lianas.
Here, we examine, for the first time, the applicability of UAVderived image data to assess the presence and degree of liana infestation in tropical tree canopies, using ground-based observations as a benchmark. Specifically, we aim to assess the validity of utilizing a consumer-grade UAV and camera as a new method for collecting data on liana infestation, by (a) assessing interobserver bias in clas-   (Figure 1). Danum is characterized by lowland, evergreen dipterocarp forest, covering ~43,800 ha, and Sepilok by alluvial lowland dipterocarp, sandstone hill dipterocarp, and kerangas forests, covering ~4,300 ha. We surveyed 17 plots: eight 1-ha plots located in the Center for Tropical Forest Science (CTFS) 50-ha plot, three additional 10-ha plots (Berry, Phillips, Ong, & Hamer, 2008), and three 0.05-ha circular plots (Foody et al., 2001) in Danum and three 1-ha plots in Sepilok located in the alluvial, sandstone hill, and kerangas forests (Nilus, 2004).

| Ground-based data collection and liana assessments
We classified the liana load carried by each tree ≥10 cm DBH within the plots using two methods: (a) crown occupancy index (COI) and (b) percentage liana cover (%LC). The COI expresses liana load in the tree crown on a simple 5-point ordinal scale: (0) no lianas in the crown, (1) 1%-25%, (2) 26%-50%, (3) 51%-75%, and (4) >75% of the crown covered by liana leaves (Clark & Clark, 1990). This index is widely used in liana research and accurately measures liana loads at both the individual tree-and site level with little interobserver bias (van der Heijden et al., 2010). %LC is a more detailed estimate, expressed as the mean of four compass quadrants into which the tree crown is visually split and percentage of the crown covered by lianas estimated to the nearest 5% (cf. Marvin et al., 2016).
The plot corners (or midpoints for the 0.05-ha plots) and individual trees ≥10 cm DBH within each plot were georeferenced using a handheld GPS unit (Garmin eTrex Vista HCx), allowing individual tree crowns to be identified and cross-referenced in the UAV images. We also assigned each tree ≥10 cm DBH a value indicating the light level its crown received using the crown illumination index (CII) (Clark & Clark, 1992). This ordinal scale index is more-or-less equivalent to canopy stature (1 = understorey, 2 = lower canopy, 3 = mid canopy, 4 = upper canopy, 5 = emergent); it helped identify individual trees on the UAV image data and allowed comparison across different tree canopy statures.
F I G U R E 1 Location of the 2 study sites and 17 plots, which are in the state of Sabah, Malaysia, on the island of Borneo, Southeast Asia. The orthomosaics created from the UAV survey (150 ha) are shown outlined in white on top of satellite imagery, and the plots (14.15 ha) outlined in yellow. Satellite imagery source: DigitalGlobe WorldView2 RGB imagery The CII, COI, and %LC values were assigned by two independent observers, who discussed their estimates in the field and agreed final values for each tree. In one of the Danum plots %LC data were not collected, and in the Sepilok plots we only collected data for higher canopy-level trees.

| UAV data collection and liana assessments
We acquired images of the forest canopy using a lightweight, agile, inexpensive, commercially available quadcopter UAV: a DJI Phantom 3 Advanced equipped with an integrated three-waveband (RGB) camera, mounted on a three-axis, gyro-stabilized gimbal. The highquality Sony EXMOR 1/2.3″ 12-megapixel camera has a narrow 94° field of view lens (35 mm format equivalent: 20 mm) reducing "fisheye" image distortion, and an f/2.8 aperture and 8s-1/8000s shutter High image overlap at ground-level was necessary to maintain adequate overlap for producing orthomosaics at canopy-level (for more information see Supporting Information Appendix S1.3). We identified canopy gaps large enough to allow the UAV to be launched/ landed, and manually piloted it through, to ensure maximal pilot control and minimal risk of collisions. The flights were conducted during calm conditions to prevent wind effects on leaves (McNeil, 2016) and, where possible, when there was even cloud cover to ensure diffuse radiation and minimize shadowing in the canopy-improving clarity in the images and aiding liana identification. All flights took place concurrently with the ground assessments in May and June 2016. Additional details on the UAV surveys, and our experiences and recommendations for using UAVs for research, are in Supporting Information Appendix S1.
In total, 6,094 and 1,344 images taken 30 and 60 m above the canopy were captured with spatial resolutions of ~10 mm/pixel and ~20 mm/pixel, covering ~150 ha and ~50 ha of forest, respectively, within which the plots cover 14.15 ha. The images were assembled to form a single two-dimensional orthorectified image (orthomosaic) for each plot, geo-referenced to the WGS84 UTM Zone 50N pro-  to automate liana infestation identification in tree canopies from RGB image data. Any trees for which ground data were collected, but which were obscured by larger trees and not visible on the UAV image data, were excluded from further analysis on individual tree levels but retained for plot-level comparisons of UAV and ground surveys.

| RE SULTS
Liana load data were collected in 17 plots across 4 forest types, via both ground-and UAV-based methods, for more than 3,500 trees.
The ultra-fine spatial resolution (10 mm/pixel) of the UAV data ren-  (Figure 2b). This remained true when using the coarser ~20 mm/pixel spatial resolution.

| Reproducibility (between observers)
We assessed interobserver bias in classifying liana load from the UAV image data; bias in ground-based surveys was examined previously by van der Heijden et al. (2010)

| Reproducibility (against a benchmarked method)
There was high concordance of COI scores between UAV and ground surveys for the full dataset (Kendall's W = 0.947, p < 0.001, N = 3,555; Tables 2 and 3), with liana load scored the same on 71.1% of occasions. Classifications differed by one class for 26.1%, and by two or more classes for 2.8% of the trees. The most frequent differences between UAV and ground surveys (43.2% of trees that differed) were when COI was scored 0 (liana-free) by ground-based surveys and 1 (low infestation) by UAV surveys (Tables 2 and 3). Since canopies of taller trees can be hard to see from the ground due to the stratified nature of tropical forests canopies, we also compared ground and UAV survey results for tree crowns located in different canopy strata. We found strong agreement between ground-and UAV-derived COI values (Kendall's W > 0.9) for all canopy stature classes except emergent trees (Kendall's W = 0.750; Table 2; Supporting Information Appendix S3.4a-d). The most common differences were again when COI was classed as 0 by the ground survey and 1 by the UAV survey, especially in the higher canopy strata (this was up to seven times more likely for emergent trees). Agreement between the two methods was greater for tree crowns in the lower canopy layers, and for more heavily infested individuals in upper canopy strata (Table 2; Figure 5). This indicates TA B L E 1 Degree of concordance between different observers in independently assessing liana infestation of tree crowns in UAV image data using crown occupancy index (COI) and percentage liana cover (%LC). Differences in COI and %LC were assessed with Kendall's W and Spearman's rank tests, respectively  Figure 5) than ground-based estimates for almost all plots, with the difference more pronounced for plots with lower liana infestation.

| Efficiency
The UAV method was particularly time-efficient, with liana infestation assessment on average more than five times faster than the groundbased method, including both field and laboratory time (Table 4). Field campaigns are typically the most costly and time-limited phases of ecological research, and here the efficiency of the UAV survey over the ground survey is particularly enhanced. It reduces field time by 98.6% and the fixed costs of UAV hardware and software are recovered in the first 5.5 ha, with further UAV surveys costing 5.5% as much as the ground survey (Table 4). As fixed costs decrease with developments in UAV technology and popularity, the break-even point will occur even sooner than the 5.5 ha in this study. indicating high reproducibility of the UAV method. Additionally, the UAV method was much more time-efficient than the groundbased method, particularly in the field, and considerably more costefficient over multiple surveys (Table 4), with initial investment recouped within the first six plots. The UAV also remains cheaper than most suitable satellite or manned aerial survey image data.

| D ISCUSS I ON
TA B L E 2 Percentage of trees in each of the crown occupancy index (COI) classes for the ground (G) and UAV surveys, and the degree of concordance between the surveys (Kendall's W) for the full dataset (All trees) and the dataset partitioned by canopy strata. p < 0.001 for all comparisons N W  Supporting Information Appendix S3.4c,d). These tall trees store and sequester the most carbon and are the main commercial species. Liana-induced changes in them may, therefore, be an important mechanism affecting forest-level and tree-level carbon storage and sequestration, for which UAVs represent a particularly useful management tool. Successful liana management may also help to increase timber and fruit productivity, and carbon storage and sequestration of degraded forests (van der Heijden et al., 2015). In particular, UAVs increased our ability to detect low-level liana infestation in these trees, which is particularly difficult from the ground as they are often partly obscured by shorter canopy trees (Table 3; Figure 4c,d). Although lianas exert limited effects at low levels (<50% crown coverage; for example Ingwell et al., 2010), identifying them quickly is important as infestation progression is likely as lianas utilize each other to climb into the tree crown (Putz, 1984), stressing the importance of repeated surveys of liana infestation.

COI 0 (%) COI 1 (%) COI 2 (%) COI 3 (%) COI 4 (%)
Unmanned aerial vehicles answer this need, offering user- facilitates monitoring and assessment of the success of management practices after they are put in place (Zahawi et al., 2015), including changing where management efforts are concentrated, TA B L E 4 Efficiency comparison of the time taken and costs required to collect crown occupancy index (COI) and percentage liana cover (%LC) using ground-based and UAV-based methods for a 1-ha plot, including the fixed costs (field and UAV equipment and processing software) required for the whole campaign (14.15 ha for ground; 150 ha for UAV). All timings are reported in "person-hours" (i.e., time taken to collect data considering number of people required) and exclude time to walk to the plots. All costs exclude international travel and costs/ ha have been calculated based on number of days required at each plot with daily rates of £14.10 and £27.75 for accommodation/ subsistence and field assistant costs, respectively. Training time is not included as it is not measured per ha. We found 5 hr ground and 2 hr UAV training was sufficient. Ground-measurement training must take place in the field; UAV training can take place beforehand, although some flight training in a tropical forest is recommended. b We found it possible to survey three 1-ha plots, at two altitudes in a single day. To generate costs/ha, we have divided the daily accommodation/subsistence and field assistant costs by three.
as spatial patterns of liana infestation change over time. It will also help track temporal changes, not only in liana infestation but also in wider canopy phenomena, such as tree crown shape and area, and timber, fruit, and forest-level biomass productivity and phenology, over shorter time-scales than is possible with ground-based surveys. The ability to track temporal changes en- forest biodiversity, which is particularly important for the management of degraded forests (e.g., Marshall et al., 2017). With the emergence of a new platform and sensor capabilities, the opportunities for using UAVs in both liana, and canopy research more generally, will increase. Thus, the UAV method presented here offers a wealth of opportunities for forest canopy research and monitoring, including liana monitoring, over space and time to assist with tailored management of tropical forests, and forms a firm foundation for exploiting future advances.

| CON CLUS IONS
The recent proliferation of lianas, coupled with their large impacts on the carbon balance and cycle of tropical forests, has made it important to study liana infestation of tree canopies more comprehensively and frequently than feasible with current methods. Here, we show, for the first time, how capturing RGB images of tree canopies via an inexpensive, lightweight UAV can be used accurately and efficiently to assess liana infestation and help make such data collection more accessible. Liana infestation data derived from UAV image data are at least as accurate as traditional ground data, and superior in assessing liana infestation of tree crowns in upper canopy layers, enabling future advances in liana and tropical forest ecology research. The support for frequent surveys, data archiving, wealth of additional data captured, and larger geographical extent covered will enable more detailed monitoring of liana infestation and forest canopies over space and time with the potential to revolutionize both liana and canopy research. These advantages will be enhanced by rapidly developing protocols for UAV use in science (Duffy et al., 2018) and the potential for additional sensors offered by UAV platforms. UAVs also provide potential for tailored and targeted liana management protocols to effectively manage liana infestation to aid restoration of degraded forests, silvicultural systems, and projects designed to increase carbon storage and sequestration in tropical forests.

ACK N OWLED G EM ENTS
We