Estimating tree species composition with lidar and

point-based deep learning

Brent A. Murray 1, Nicholas C. Coops 1, Lukas Winiwarter 1, 2, Joanne C. White 3, Adam Dick 4, Ignacio Barbeito 1, Ahmed Ragab 5, 6

1 Department of Forest Resources Management, University of British Columbia; 2Photogrammetry Research Group, Department of Geodesy and Geoinformation, TU

Wien; 3Canadian Forest Service (Pacific Forestry Centre), Natural Resources Canada; 4Canadian Forest Service (Canadian Wood Fibre Centre, Natural Resources

Canada; 5 CanmetENERGY, Natural Resources Canada; 6 Department of Mathematics and Industrial Engineering, Polytechnique Montreal

Symposium on Systems Analysis in Forest Resources –May 15, 2024

Background and Introduction

Increased knowledge of forest attributes, including tree

species, can help to improve forest management

practices in Canada

Commercial Value Habitat, Biodiversity,

Species at Risk

Cultural Practices

Methods for Estimating Tree Species

Photographic Interpretation

Use fine-scale aerial images and a variety of

cues to estimate species composition.

Automated Classification

Regression and machine learning

approaches to estimate species groups

(functional groups or leading species)

Single Photon Lidar

Single photon lidar (SPL) data was acquired in Ontario,

Canada to support Forest Resources Inventory (FRI)

development

•Forest Cover

•Tree Density

•Tree Height

•Terrain and Elevation

•Hydrology

•Transportation

Minimum point density is 25 points/m and was classified

according to the Ontario Lidar Classifications

Modeling Species

From Lidar

Estimating tree species from lidar is a

difficult task

•Lack of spectral information

•Individual tree extraction is difficult

and species estimation of each

crown is problematic

•Challenging in mixed plots

Can advanced computing approaches

such as deep learning help in the

assessment of tree species composition

at the plot level?

Objective

To estimate tree species proportions using ALS

data and existing forest inventory data with

adapted point-based deep learning techniques.

Site Summary

•Romeo Malette Forest (RMF)

•630,000 ha, with 582,430 ha

being forested

•Dominant tree species:

•Black Spruce

•White Spruce

•Jack Pine

•Trembling Aspen

•Paper Birch

•First Nations:

•Flying Post First Nation

•Matachewan First Nation

•Mattagami First Nation

•Taykwa Tagamou Nation

•Wahgoshig First Nation

•Métis Nation of Ontario

Modeling Approach

Forested

Polygons

> 10,000 m2

20 Polygons

With Same

Composition

Code

400 m2 Plots

Generated

50 m from

Boundary

Plots Within

Disturbed

Areas

Removed

(NTEMS)

SB70 BF10 CE10 PT10

[BF, BW, CE, LA, PT, PJ, PO, SB, SW]

[0.1, 0.0, 0.1, 0.0, 0.1, 0.0, 0.0, 0.7, 0.0]

SB70 BF10 CE10 PT10

Point Clouds

Extracted

Labelled With

Species

Composition

Vector (SCV)

Point Clouds

Down-

Sampled to

7168 Points

Height

Normalized

and Filtered

> 2 m

Model Results

The model achieved an R2adj

of 0.61 and an RMSE of 0.14

Model Results

Species

RMSE

MAE

Balsam Fir

0.053

0.039

Paper Birch

0.171

0.124

Cedar Species

0.122

0.069

Larch

0.115

0.073

Jack Pine

0.172

0.089

Populus Species

0.109

0.021

Trembling Aspen

0.192

0.117

Black Spruce

0.210

0.168

White Spruce

0.025

0.007

Model Results

Leading Species

•F1 –0.63

•OA –0.67

Coniferous and

Deciduous

•F1 –0.85

•OA –0.85

Model Results

77% of estimated

values were ± 10%

89% of estimated

values were ± 20%

Thank You

Brent Murray

@BrentAMurray

brntmrry@student.ubc.ca

To Paper