Van-Tho Nguyen is a postdoctoral fellow under Richard Fournier at Sherbrooke University. He completed his master’s degree in image processing at La Rochelle University, France, in 2014 and obtained his doctorate in plant and forestry biology from Agroparistech, France, in 2018. His Ph.D. research was on the estimation of the quality of standing tree and roundwood by detecting and analyzing the defects on the trunk surface from high-density terrestrial LiDAR data.
Van-Tho is working on Question 19 of Theme 3 for AWARE, building on the work of Abdelmounaime Safia on the development and validation of algorithms adapted to estimate tree attributes Terrestrial Laser Scanning (TLS). TLS offers a unique way to estimate structural attributes at individual trees and plot levels. Specific algorithms are needed to deal with TLS limitations, namely occlusion and non-uniform sampling. Van-Tho’s research project aims to develop algorithms to estimate structural attributes at individual trees and plot levels by taking into account the limitations of TLS. His research consists of three parts, first, he continues to develop the L-Vox software (a plugin of Computree) which allows estimation of the vertical structure of the forest. Second, he uses 3D distribution of canopy elements computed from L-Vox as a reference to develop the ability to achieve similar 3D mapping with aerial LiDAR scanning using the full waveform data. This allows an exact mapping of vertical canopy layers and detecting the presence of understory. Third, Van-Tho continues to develop an algorithm to isolate individual tree crowns from the TLS data. This algorithm was based on component connected labeling that was initialized by Joris Ravaglia.
Aside from his postdoctoral research project, Van-Tho’s other interests are in computer science, particularly in signal processing, artificial intelligence including deep learning, and how they can be applied to LiDAR data. In his personal time, Van-Tho enjoys spending it with his family and reading fictional and historical books.
Q21: SPECIES IDENTIFICATION AT THE INDIVIDUAL TREE LEVEL
Jean-François Prieur is a PhD student under the supervision of Richard Fournier (Université de Sherbrooke) and Daniel Kneeshaw (Université du Québec à Montréal). His research project (question 21 of theme 4) centers on the identification of tree species (within mixed hard- and softwood stands) at the individual tree crown (ITC) level using airborne laser scanning (ALS) data. Better species information will improve the accuracy of forest inventories and lead to a greater understanding of forest structure and dynamics. His current study areas are Petawawa Research Forest (PRF) located in Petawawa, ON, and Black Brook Forest (BBF), near Edmunston, NB. After developing his methods for large-scale processing of ALS data at BBF for our industrial partner J.D. Irving, Limited, he is now working with data from several ALS overflights at PRF spanning 6 years: standard ALS (2012), multispectral ALS (2016) and Single Photon Lidar (SPL) (2018). Standard ALS systems scan with one laser beam and using one wavelength while multi-spectral ALS systems scan with multiple beams each with their own wavelength (three beams in the case of the system used in this research, the Teledyne Optech Titan). His current work compares these three systems when used for species identification at the individual tree level.
Standard ALS has already been shown to have high accuracy when separating broadleaf crowns from conifer ones but accurate and robust species identification at the ITC level is still difficult in dense, mixed forests such as the ones found at PRF and BBF. Delineation of the individual crowns remains problematic in these environments, at least with the ITC software used in this project (SEGMA). Nevertheless, it is thought that the additional point density (more beams) and different wavelengths (which can be used to create pseudo-NDVI features) provided by MSL could further enhance the predictive capabilities of the random forest models used to classify previously delineated crowns. As for SPL, it is an emerging technology that uses photon counting methods to achieve equivalent nominal point densities from higher flight paths (590 km2/h vs 50 km2/h for standard systems at same altitude) giving it obvious spatial coverage advantages over standard ALS systems. It is, however, uncertain whether this advantage translates to the species identification task at the individual tree level.
The 2012 ALS data was used to delineate crowns at PRF. These crowns were then used in field work that occurred in 2014 and 2015 to assign species information to individual crowns. Approximately 1500 crowns were identified, covering 12 species, these were then used as training crowns to train the random forest classifier. Python scripts performed a spatial join between the SEGMA ITC crown shapefile and the ALS point cloud data to extract the ALS points specific to each crown. 3D and intensity features are then calculated for each crown using and used to train the classifier. The 2016 and 2018 data were processed using the crowns delineated in 2012 in order to avoid having to do field work after each lidar overflight. The methods to train the classifier were the same for all sensors.
The primary objective of this study is to ascertain whether the ITC tree species identification methods tested under linear ALS and SPL systems’ data perform similarly. We carried out this examination for different species groups ranging from hardwood/softwood to distinct species groups. A secondary objective was to determine if the increased number of species identification features, derived from the multispectral lidar, or the higher point density of the SPL provides us with greater classification accuracy than the standard linear ALS baseline.
Preliminary results are shown below in Table 1. A manuscript is in review for submission to a peer-reviewed journal (ISPRS Photogrammetry and Remote Sensing) and the results were accepted for presentation at the ISPRS 2020 Conference that was supposed to happen in Nice, France this June.
Table 1: RF classification average overall accuracy (20 classifications, correlated features > 0.9 removed) for the different datasets, feature subsets, and species groupings: Functional Groups (Tolerant hardwood, Intolerant hardwood, Other softwood, Pine and Spruce), 12 species (Ash (Black/White), Basswood, American Beech, Birch (White/Yellow), Eastern White Cedar, Balsam Fir, Eastern Larch, Maple (Red/Sugar), Red Oak, Pine (Red/White), Trembling Aspen and Spruce (Black/White)), 4 species (Maple, Pine, Poplar, Spruce)
Briefly, models based on SPL data show an overall accuracy that is slightly lower than the linear ALS systems. The SPL intensity features are more highly correlated than the ones calculated with both the ALS and MSL datasets which reduces the number of features available to train the classifier. There are differences as well in the structure of the SPL returns (distribution of 1st and 2nd returns for example which have been found to be important features for species identification) compared to standard lidar systems. These differences can explain the lower results for the SPL sensor.
MSL appears to be the most useful sensor for tree species identification at the individual tree level. Hardwood species classification accuracy is limited by the challenges of ITC delineation in dense hardwood-dominated forests. ITC delineation (and subsequent identification of training crowns in the field) in this type of environment is challenging and introduces errors of omission and commission when identifying tree crowns in these conditions. The relatively low number of training samples per class (30 crowns for some classes) at the 12 species level is also a constraint to classification accuracy. Improvements in ITC delineation is needed to further improve the results as we drill down to species.
The Hardwood/Softwood, functional group and 4 species classification accuracies, however, are acceptable (greater than 70%) across all three datasets. The results indicate that the additional features provided by MSL are more beneficial to classification accuracy than the additional point density provided by SPL. These additional features are also beneficial in the feature selection case since they provide a larger pool of features that the algorithm can choose from. Removing correlated features has a beneficial effect on the overall accuracy of models with fewer features, this is especially clear in the 4 species case. Parsimonious models have been found to better generalize our predictions across entire study areas.
Finally, we conclude that the ITC segmentation (and training crown field sampling) performed on the initial ALS (2012) dataset can be used to calculate features not only on the ALS dataset but on the two subsequent overflights (MSL (2016) and SPL (2018)) with broadly comparable results, even with SPL data.
Annual General Meeting
The final AWARE annual general meeting was held on February 19-20, 2020 followed by the first AWARE technology transfer session on February 21, 2020 in Toronto, Ontario. The AGM was held in the gothic revival Daniel’s building (originally built in 1875) of the Faculty of Architecture at the University of Toronto. The Technology Transfer session was held at the Double Tree by Hilton hotel in downtown Toronto. The AGM showcase day on February 20th was attended by 60 people, with 20 guests that came from different universities, forestry agencies and companies.
Downloadable copies of AGM presentation slides are available in the AWARE website.
In March this year, NSERC informed AWARE that the remaining unpaid balance of its funding will not be paid in full. This is due to the overall shortfall in in-kind contributions from industry partners that will result to a $27.5K reduction in the final NSERC installment. We are truly grateful to our industry partners for their improved involvement in the past year and their efforts to meet their original in-kind commitments despite challenging times.