Counting trees using Machine Learning

on smallholder farmland in Tanzania

Sander de Haas Tom Eames

Chief Technical Officer PhD Candidate
Justdiggit, The Netherlands Vrije Universiteit Amsterdam, The Netherlands
Kim de Groot
Data Scientist
Lynxx, The Netherlands
Landscape restoration
Using simple techniques to bring back vegetation
Communication and media
Use modern channels to reach people and build a movement
Agroforestry
Smart pruning to turn tree stumps into trees
>6,300,000 trees
In just 3 years
Track tree numbers
Manual data collection by our 1200 trainers
Treetracker
Collect tree data on the ground
How to measure >6 million trees?
Spread out over >4 million hectares?
Track isolated trees in agricultural fields
Lots of noise in remote sensing signals
 Drone Sentinel
(5 cm) (10 m)
13 villages whole region
x 100 ha >4,000,000 ha
 Drone Planet SkySat Planet NICFI Sentinel
(5 cm) (50 cm) (3 m) (10 m)
13 villages 13 villages whole whole
x 100 ha x 2500 ha country country
Tree detection steps
Steps Goal
1. Count all trees using ML on drone data: Insight on tree count + size distribution,
All trees in 13 villages x 100 ha @ 5 cm including very small trees

2. Use this to train a ML model for SkySat data: Tree count for much larger area including
Most trees in 13 villages x 2500 ha @ 50 cm smaller trees

3. Detect tree cover change in NICFI data: Detect changes for the whole region
4 million hectares @ 3 m

→ accurate estimation of our impact

including tree count
Partnership

non-profit private sector university

Kim de Groot

Tree recognition on
drone imagery
From drone imagery to counting individual trees

Tree detection
The importance of labeling
The importance of labeling
Solving the half tree issue
Timing and contrast
Suggestions?

Calibrate Geotiff files

From drone imagery to counting individual trees

Final results
Bumila 2018 Bumila 2019 Bumila 2020

● TreeID 208  2.5m2 ● TreeID 208  2.4m2 ● TreeID 208  2.6 m2

● 4577 trees ● 12219 trees ● 7021 trees
● 0.65ha tree crowns ● 1.45ha tree crowns ● 0.9ha tree crowns
Tom Eames & Sanne Kuipers

Tree recognition in
Satellite imagery
Or: how to count
trees from space
My research

Remote sensing of biomass

Emissions = BA x FL x CC x EF
Tree/shrub biomass

Fig. 1 from
Brandt et
al., 2020
Method

3m
0.5m
Skysat 0.5m NICFI 3m

Skysat 0.5m RGB NICFI 3m RGB

Skysat 0.5m NICFI 3m

Skysat 0.5m ‘Truth’ NICFI 3m ML prediction

Skysat 0.5m NICFI 3m
High

Mid

Low

Skysat 0.5m ‘Truth’ NICFI 3m ML ‘confidence’

Conclusions & caveats
● In principle, objects >0.25m2 detectable. In practice, likely to be
error-prone <3m2 (also reported in Brandt et al.)
● Best results still reminiscent of fractional tree cover - better insight into
tree number at the expense of being ‘definite’
● Next steps: train/validate with drone data

Reference:
Brandt, M., Tucker, C. J., Kariryaa, A., Rasmussen, K., Abel, C., Small, J., … Fensholt, R.
2020. An unexpectedly large count of trees in the West African Sahara and Sahel.
Nature, 5877832, 7882. https://doi.org/10.1038/s4158602028245
Sander de Haas

Conclusions and
way forward
 Drone Planet SkySat Planet NICFI Sentinel
(5 cm) (50 cm) (3 m) (10 m)
all trees most trees at tree cover for
at field level regional level whole country
including small
trees and details
such as crown
size distribution

Accurate estimation
of program impact
including tree count
Conclusions & way forward
● Individual models for drone and satellite imagery show very good results.
● Next step: refine and combine to gain insight in total program impact.

● We hope to apply this approach for annual monitoring of current and

upcoming large landscape restoration projects.
● Depending on imagery available and partnerships.

For questions, more information or suggestions, please contact us:

Sander de Haas Kim de Groot Tom Eames

sander@justdiggit.org kim.degroot@lynxx.com t.c.eames@vu.nl