An accurate segmentation of individual trees using LiDAR suggests a wide variety of forest applications. From the exact height and width of individual trees it is trivial to quantify, for example, carbon storage or timber volumes. The identification of individual tree species could also be obtained by incorporating such LiDAR attributes as intensity or returned pulse width into analysis.This 2-part article was provided by Iurii Shendryk and originally published in the Aug/Sep 2016 Issue of Position magazine. To read part 1, please click here.
The Geospatial Analysis for Environmental Change (GAEC) lab at UNSW used accurately segmented trees to classify forest health using machine learning and field-measured tree crown dieback as indicators of eucalypt tree health. Although tree health is a subjective term, in the Barmah-Millewa forest it is best approximated by dieback levels (i.e. the proportion of dead branches to the total number of branches). This attribute was visually assessed in the field and grouped into three classes.
The LiDAR indices were then calculated for segmented tree crowns exploiting the full range of LiDAR attributes and tree geometry, and were used as predictor variables in object-oriented random forest classification. Random forest is a supervised non-parametric machine learning technique, and was particularly suitable for this study as it was able to achieve superior classification performance as compared to other ensemble learning algorithms with small training samples.To get stories like this delivered to your mailbox every week, subscribe to our weekly newsletter.
The results showed that tree health of individual eucalypt trees can be classified with an overall accuracy of 81%. The model was built from less than 100 field-measured trees and allowed classifying millions of LiDAR-measured trees, and could easily be scaled up.
Imagine an airborne system that fires lasers out of the bottom of the plane, sweeping across the ecosystem and providing real-time information on forest attributes resolving individual trees in high resolution 3D. Yes, this type of 3D mapping is possible these days and could be used to transform climate change and resource policy developments.
However, in order to reduce the cost of larger scale studies, GAEC team suggests using very high resolution multispectral satellite imagery as a possible substrate for extrapolating LiDAR-derived individual tree health. This is the aim of Shendryk’s follow up study, where he will use 800 square kilometres of DigitalGlobe Foundation granted very high resolution satellite imagery over the Barmah-Millewa forest to upscale his tree health maps.
Basically, individual tree crowns could be also identified in satellite imagery and their spectral signal related to the LiDAR or field measured tree health. Thus making the accounting system scalable and affordable.
Tree extraction represents one of the most difficult tasks when segmenting complex LiDAR scenes, due to their irregular and entangled shapes. Therefore, this newly developed algorithm could be an ideal addition for modelling of urban areas to produce, for example, seamless 3D models of entire cities. Moreover, recent advances in the field of multispectral (e.g. Optech Titan) and space borne LiDAR systems (e.g. NASAs GEDI) could be a starting point for the potential development of a global accounting system of trees as well as large scale bio-banking schemes.
Informing environmental decisions
Recognising the signs of unhealthy forest and teasing out the causes are important both for sustaining the services that humans rely on and for the effective conservation of forest ecosystems. The decrease in flooding has been frequently identified as the main cause of tree health decline in Australia’s floodplains.
To test this GAEC team overlaid the flooding frequency map derived from time series of Landsat imagery (1986-2011) with the LiDAR-derived tree health maps of the Barmah-Millewa forest. Their findings highlighted that trees located in infrequently flooded areas were most susceptible to dieback. This is crucial for management of, for example, environmental flows, as forest areas that require more water than others could be easily identified.
The GAEC team hopes that their methodology will act as a starting point for the introduction of forest health monitoring framework in Australia, enabling to prioritise areas for forest health promotion and conservation of biodiversity.
The potential users of such technology driven by smart algorithms span from governments to ecologists, as it promotes efficiency and large scale applicability.Iurii Shendryk is a PhD candidate at the University of New South Wales in Sydney, Australia. His research is focused on the integration of remote sensing, GIS and spatial statistics to explore interactions between species, environment and land use, with ongoing research and capacity building work focused on forest health monitoring.