Q21: SPECIES IDENTIFICATION AT THE INDIVIDUAL TREE LEVEL
Jean-François Prieur is a PhD student under the supervision of Benoît St-Onge (Université du Québec à Montréal) and Richard Fournier (Université de Sherbrooke). His research project (question 21 of theme 4) centers on the identification of tree species (within conifer dominated (fir/spruce) and mixed hard- and softwood stands) at the individual tree level using a combination of lidar data and optical imagery. 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, Ontario, and Black Brook Forest (BBF), near Edmunston, New-Brunswick. The data he is currently working with comes from a 2016 overflight of PRF using the Teledyne Optech Titan multispectral lidar, and single channel lidar data acquired at BBF and PRF in 2012. Additional remote sensing data to be collected in Summer 2018 will also be incorporated in his research
He has taken a one-year leave of absence from his PhD (Fall 2017 to Fall 2018) for personal reasons, while continuing to support research efforts in his lab group. The time off, as well as the knowledge acquired when designing the ALS ITC feature extraction process for the multispectral PRF2016 dataset has given him the opportunity to reflect on the current state of our species identification research. This will help address some challenges encountered in the current processing workflow as well as to make it more efficient over large areas.
The ALS feature work of Budei et al. (2018), which used multispectral ALS data on manually delineated crowns in York Regional Forest, Ontario to achieve an accuracy of 76% for 10 species, was applied at PRF on ground-surveyed crowns (Murray Woods, 2014 and 2015) that were automatically delineated by SEGMA software. A result of 66% for 9 species was achieved using 3D and intensity features derived from multispectral ALS. These features are now being used in Parvez Rana`s postdoctoral research (Q23) in transfer learning; the generalization of models constructed on one forest so that we can apply them to another forest.
The overarching bjective in JF’s tree species identification methodology is to develop a deep learning approach and workflow for identifying species without the need to pre-identify any 3D or intensity classification features. One of the technical challenges related to species identification at the individual tree level is the efficient extraction of point cloud data (and subsequent 3D and intensity feature calculations) for each training crown at the calibration stage, and for all crowns of a given site at the classification stage.
Approximately 1500 training crowns were extracted from the lidar data for PRF as well as 2000 training crowns for BBF. At PRF, the average size of each flight strip is 7-8 GB and contains between 50-60 million points. Each mature tree is typically composed of 10-100 points. The amount of processing required on the ALS data to extract these individual crowns and calculate features is cumbersome as well as time consuming. The development of an efficient and operationally-deployable classification workflow (BBF is composed of over 3000 1km2 tiles for example) is one of the primary drivers of his research. Even the hundreds of features used in Budei et al. (2018) are not complete and the number of possible 3D and intensity features we can come up with are practically limitless. Convolutional Neural Networks (CNN) have the characteristic of deducing a dataset feature map on their own via training, this approach seems more elegant from a completeness standpoint.
Most of the current literature on topic uses large 3D CNN’s using voxels or FWF data in combination with an optical imagery source (usually hyperspectral). These large 3D CNN’s require long training times on high-end GPU’s like the Nvidia Titan. Our approach uses smaller 2D CNN’s (Python with Keras and Tensorflow modules) using multiband raster products derived from ALS data (CHM and intensity values namely). This approach, if successful has the advantage of completely removing the necessity to create, extract and calculate 3D and intensity features from the point cloud. The CNN is small enough to be trained on a CPU although a GPU is 5x faster and can be exploited. The following table shows preliminary results which are promising:
|PRF||# of Patches (training/validation)||Training time||MS ALS (I1/I2/I3)|
|Broad vs Needle||1155/281||< 1 min||87%|
|Spruce vs Fir||171/63||< 1 min||77%|
|Red pine vs White pine||213/53||< 1 min||76%|
|Red maple vs Hard maple||154/38||< 1 min||87%|
|White spruce vs Black spruce||139/34||< 1 min||63%|
|GPU: Quadro P3000 6GB|
LIESMARS 2nd international graduate student workshop, Wuhan University, Wuhan, China, July 2018
IGARSS 2018 (FR3.R11.5), Valencia, Spain, July 27 2018
Budei, B. C., St-Onge, B., Hopkinson, C. et Audet, F.-A. (2018). Identifying the genus or species of individual trees using a three-wavelength airborne lidar system. Remote Sensing of Environment, 204, 632‑647. doi:10.1016/j.rse.2017.09.037