UAVs are being used to take large numbers of tropical forest canopy photos at Harapan Rainforest Ecosystem Restoration Concession in Indonesia and at the Atewa Range Forest Reserve in Ghana. UAVs capture extremely high resolution imagery on demand where previously this could only be achieved at vast cost, using manned aircraft and state of the art sensors. This is a fantastic opportunity for technology to help study these complex systems and provide useful information about forest structure and species composition so that the areas most in need of conservation or restoration are identified. Supported by the Cambridge Conservation Initiative Collaborative Fund for Conservation, the project aims to develop low-cost methods for measuring forest quality and directing restoration management.

UAVs are able to measure forest structure through a technique known as structure from motion (SFM), which builds 3D forest models (point clouds) from images taken of the mapped area from large number of locations, which have a high degree of overlap (Figure 1, below). The data produced are similar to those obtained by LiDAR (using a position and orientation corrected scanning laser) but at a fraction of the cost. Many automated techniques are already available for analysing LiDAR data and these can be applied directly to SFM. However, SFM point clouds tend to have far fewer ground returns than those obtained from LiDAR and as a result tend to produce erroneous digital terrain models (DTM), and consistently under-estimate canopy height, particularly where forest canopies are most dense (Zahawi et al., 2015, Wallace et al., 2016).

Figure 1. A 3D model produced from structure from motion. (A) A recovering secondary forest at the left and top of the image, and oil palm at the right of the image. (B) A false colour image showing an automated ground classification, with the ground shown in brown, outliers shown in pink, and all other pixels shown in white. (C) The canopy height model produced from the 3D model by subtraction of the digital elevation model from the digital surface model. 

Canopy height is positively correlated with a number of forest quality metrics, including successional status, biomass and carbon, and is a useful proxy of forest quality. Canopy height is also a powerful starting point for taking a tree-level approach for assessing forest quality (e.g. tree size distributions and individual tree carbon accounting). This may be achieved by segmenting canopies into individual tree crowns with algorithms that use the allometric relationship with height or a watershed approach (that progressively flood a topology, starting from local minima and draw boundaries where pools of water meet). Both of which are somewhat dependent upon the accuracy of the canopy height model (CHM) (Jing et al., 2012). This highlights the importance of assessing the uncertainty of DTMs and CHMs, which is a function of canopy density and the mapping data (e.g. % overlap).

A simple but costly solution is to obtain LiDAR data prior to the use of UAVs but there would be huge value in being able to improve the accuracy of UAV DTMs, either through better software tools or sensors. Another untested possibility is to purposefully overexpose mapping images so that the understory pixels are well-exposed, which may enable better DTMs be constructed.

Species composition is another important forest quality metric which may vary somewhat independently of forest structure. Hyperspectral data (from manned aircraft and satellites) are being harnessed in increasingly complex ways to predict forest diversity and species composition (Asner and Martin, 2008, Fricker et al., 2015). At present UAVs usually only carry multispectral cameras (e.g. RGB or RGB+NIR) but can data from these sensors be harnessed to extract meaningful estimates of species identity or diversity? The advantage of RGB sensors is that they can be extremely high resolution and potentially contain a lot of colour variation and textural pattern information that may be of use. Vectors of values generated from typical image analysis transforms/convolutions can be extracted from mapping images and input, alongside training information (i.e. species labels) to machine learning algorithms which attempt to learn the best classifications for the given vectors. In theory these trained algorithms can then be used to automatically classify new data sets (Figure 2).

Figure 2. (A) An aerial image taken from a UAV at 200 m above the ground, which is (B) marked up using unique RGB triplets to code for species identity; Macaranga gigantea in orange and Bellucia pentamera in blue. (C) After training a machine learning algorithm with approximately 20 marked up images the ability for the algorithm to accurately identify species is still rather limited.

The problem is that forest mapping data are usually highly variable in terms of the angle of image capture, the size of the features, the nature of incident light etc., which can make classification extremely challenging, even for tree species which are rather conspicuous to the human eye. One way of minimising this variation is to take a tree-centred approach to classification. This requires that tree crowns are first segmented from images (as described above) and then feature vectors are extracted for these. In order to use the best quality data it may be necessary to extract the raster information from mapping photos or orthophotos (orthorectified mapping photos; these are usually produced automatically from SFM software) rather than the orthomosaic (a combination of many mapping photos, which can often be quite noisy and contain many artefacts from the mosaicking process) but currently this requires some clever GIS processing and may well be associated with cumulative errors in translating the tree position from the 3D model back to the 2D orthophotos.

To summarise, UAVs have huge potential as tools for measuring forest quality, in terms of tracking the recovery of forest structure and species composition. However, their relatively recent explosion in popularity has led to a glut of data and a lack of dedicated tools and expertise for analysing them. Canopy height maps can be generated relatively easily but their quality is highly dependent upon the density of true ground returns and care needs to be taken to ensure these are accurate. Automated tree species identification may be possible in the near future but at present is challenging even when using only multispectral sensors.

About the Contributors

Tom Swinfield is a Conservation Scientist working for the RSPB Centre for Conservation Science whose work focuses on measuring tropical forest quality and techniques for enhancing forest recovery.

David Coomes is the Head of the Forest Ecology and Conservation Group at the University of Cambridge which uses state of the art remote sensing technologies to investigate how human pressures are changing forests around the world in order to better direct conservation policy.

Jeremy Lindsell is Conservation Science Manager at A Rocha International where he is responsible for the Tropical Forest Programme which involves habitat monitoring in both protected and unprotected areas.

Carola-Bibane Schönlieb is Head of the Cambridge Image Analysis group at the Department of Applied Mathematics and Theoretical Physics, University of Cambridge.

Tuomo Valkonen is a Lecturer in applied mathematics at the University of Liverpool

Ping Zhong is Associate Professor with the School of Electronical Science and Engineering at the National University of Defense Technology, Changsha City, Hunan Province, PR China.

Rhett Harrison is Professor of Forest Ecology and Conservation at the Consultative Group on International Agricultural Research.