Predicting Tree Species Proportions Using ALS and

Sentinel-2 Data With Deep Learning And Data

Fusion

Brent A. Murray 1, Nicholas C. Coops 1, Joanne C. White 2, Adam Dick 3, Ahmed Ragab 4, 5

1 Department of Forest Resources Management, University of British Columbia; 2Canadian Forest Service (Pacific Forestry Centre), Natural

Resources Canada; 3Canadian Forest Service (Atlantic Forestry Centre), Natural Resources Canada; 4CanmetENERGY, Natural Resources

Canada; 5Department of Mathematics and Industrial Engineering, Polytechnique Montreal

Silvilaser 2025 –October 1, 2025

Background

2

Economic Ecological Sociocultural

Tree species are critically important for understanding

various ecosystem services.

Species Composition

3

Species composition has

traditionally been derived from

photographic interpretation.

•Polygons >10 ha

Plot level assessment may

provide the information

required for some management

practices (White et al. 2013).

White, J. C., Wulder, M. A., Varhola, A., Vastaranta, M., Coops, N. C., Cook, B. D., Pitt, D., & Woods, M. (2013). A best practices guide for generating forest inventory attributes from airborne laser scanning data using an area-based approach. The Forestry Chronicle,

89(06), 722–723. https://doi.org/10.5558/tfc2013-132

[0.3, 0.0, 0.0, 0.5, 0.0, 0.2, 0.0, 0.0, 0.0]

BF BW CE LA PJ PO PT SB SW

Enhanced Forest

Inventories

4

EFI’s provide vital structural information

of forests (White et al. 2016).

Accurate, large area tree species

mapping with ALS data is challenging,

but would be a fundamental

integration into an EFI (Coops et al.

2022; White et al. 2025).

Coops, N. C., Tompalski, P., Goodbody, T. R. H., Achim, A., & Mulverhill, C. (2022). Framework for near real-time forest inventory using multi source remote sensing data. Forestry: An International Journal of Forest Research, cpac015.

https://doi.org/10.1093/forestry/cpac015

White, J. C., Coops, N. C., Wulder, M. A., Vastaranta, M., Hilker, T., & Tompalski, P. (2016). Remote Sensing Technologies for Enhancing Forest Inventories: A Review. Canadian Journal of Remote Sensing, 42(5), 619–641. https://doi.org/10.1080/07038992.2016.1207484

White, J. C., Tompalski, P., Bater, C. W., Wulder, M. A., Fortin, M., Hennigar, C., Robere-McGugan, G., Sinclair, I., & White, R. (2025). Enhanced forest inventories in Canada: Implementation, status, and research needs. Canadian Journal of Forest Research, 55, 1–37.

https://doi.org/10.1139/cjfr-2024-0255

Previous Work

5

Studies have looked predicting

plot level species composition

data using remotely sensed data.

•ALS Data (Murray et al. 2024)

•Multispectral Imagery (Bolyn

et al. 2022)

Bolyn, C., Lejeune, P., Michez, A., & Latte, N. (2022). Mapping tree species proportions from satellite imagery using spectral–spatial deep learning. Remote Sensing of Environment, 280, 113205. https://doi.org/10.1016/j.rse.2022.113205

Murray, B. A., Coops, N. C., Winiwarter, L., White, J. C., Dick, A., Barbeito, I., & Ragab, A. (2024). Estimating tree species composition from airborne laser scanning data using point-based deep learning models. ISPRS Journal of Photogrammetry and Remote Sensing, 207,

282–297. https://doi.org/10.1016/j.isprsjprs.2023.12.008

Objective

6

Predict tree species proportions by

fusing ALS data and Sentinel-2 imagery

using deep learning

Questions

7

1

How can multiple remote

sensing data types be fused

with deep learning to predict

tree species proportions?

2

How does integrating

multitemporal multispectral

composites into a deep

learning model affect tree

species proportion prediction?

Question 1

8

2

Question 1

9

2

To what extent would the

integration of multitemporal

multispectral image

composites into a deep

learning based fusion model

impact the accuracy of tree

species proportion predictions,

particularly for broadleaf

species?

Question 1

10

2

Questions

11

1

How can multiple remote

sensing data types be fused

with deep learning to predict

tree species proportions?

2

How does integrating

multitemporal multispectral

composites into a deep

learning model affect tree

species proportion prediction?

Question 2

12

2

To what extent would the

integration of multitemporal

multispectral image

composites into a deep

learning based fusion model

impact the accuracy of tree

species proportion predictions,

particularly for broadleaf

species?

Broadleaf species have shown to

have poorer prediction accuracies

compared to coniferous species

(Beloiu et al. 2023; Bolyn et al. 2022;

Murray et al. 2024).

Species exhibit different phenological

responses throughout the year and

may help to distinguish between

species.

Incorporating multitemporal

Sentinel-2 imagery could help identify

key spectral-temporal patterns.

Beloiu, M., Heinzmann, L., Rehush, N., Gessler, A., & Griess, V. C. (2023). Individual Tree-Crown Detection and Species Identification in Heterogeneous Forests Using Aerial RGB Imagery and Deep Learning. Remote Sensing, 15(5), 1463. https://doi.org/10.3390/rs15051463

Bolyn, C., Lejeune, P., Michez, A., & Latte, N. (2022). Mapping tree species proportions from satellite imagery using spectral–spatial deep learning. Remote Sensing of Environment, 280, 113205. https://doi.org/10.1016/j.rse.2022.113205

Murray, B. A., Coops, N. C., Winiwarter, L., White, J. C., Dick, A., Barbeito, I., & Ragab, A. (2024). Estimating tree species composition from airborne laser scanning data using point-based deep learning models. ISPRS Journal of Photogrammetry and Remote Sensing, 207,

282–297. https://doi.org/10.1016/j.isprsjprs.2023.12.008

Study Site

13

2

Romeo Malette Forest (RMF) –

630,000 ha, with 582,430 ha

being forested.

Dominant tree species:

•Black Spruce

•Jack Pine

•Trembling Aspen

•Paper Birch

Treaty 9 territory (James Bay

Treaty).

Traditional lands of the

Ojibway, Cree and Anisininew,

as well as the Métis Nation of

Ontario.

Methods –Model Architecture

14

2

Methods –Model Architecture

14

2

Methods –Model Architecture

14

2

For More Information

Methods –Model Architecture

14

2

Methods –Model Architecture

14

2

Methods –Modified U-Net

15

2

Methods –Model Architecture

16

2

Methods –Model Architecture

16

2

Methods –Model Architecture

18

2

[0.3, 0.0, 0.0, 0.5, 0.0, 0.2, 0.0, 0.0, 0.0]

BF BW CE LA PJ PO PT SB SW

Results

17

Higher weights

were given to the

ALS features for

all species.

Results

18

Overall

R2: 0.58

RMSE: 0.14

19

Results

20

Black Spruce

exhibited the

highest RMSE values

while Balsam Fir and

White Spruce were

the lowest.

Results

21

Overall

OA: 0.65

F1: 0.62

Results

22

Results

23

Improvements in

predicting

broadleaf species

compared to ALS

alone by ~8% with

some as high as

31%.

24

Brent Murray

@brntmrry.bsky.social

brntmrry@student.ubc.ca

Thank You

25