1 of 25

Global restoration monitoring: the challenge of counting every tree on the planet

Dr Josh Veitch-Michaelis, DS3Lab / Restor Eco

2 of 25

About Restor

  • Founded by ETH Zurich's Crowther Lab, the first version of Restor was developed in collaboration with Google in 2020 and aimed at supporting restoration practitioners around the world.

  • Online platform that lets users and stakeholders explore restoration sites globally
  • Restor brings transparency, connectivity, and ecological insights to restoration and conservation efforts around the world.�

https://restor.eco

3 of 25

Restor

4 of 25

Why is transparency and monitoring important?

Many high-profile restoration projects fail.

�In many cases there is no monitoring protocol in place

There is often no transparency and money is wasted

Who should verify pledges?

… fewer than 2 percent of them had survived. The other 98 percent had died or were washed away …. “It was a complete disaster,” agrees Jim Enright, former Asia coordinator of the U.S.-based nonprofit Mangrove Action Project. “But no one that we know of from Guinness or the record-planting proponents have carried out follow-up monitoring.” Guinness has not responded to requests for comment.

The most mangrove trees planted within one hour by a team is 1,009,029 and was achieved by Governor Lray Villafuerte of the El Verde Movement and the people of San Rafael of the municipality of Ragay, Camarines Sur, Philippines, on 8 March 2012.

Guinness World Record Organisation

“Phantom Forests”, Yale Environment 360

5 of 25

Monitoring protocols

6 of 25

Individual Tree Crown Delineation (TCD)

Supports Restor’s mission for science and transparency

Google.org project focuses on AI applications for monitoring

Task: Instance segmentation

Predict masks over all trees in an input RGB image

Why?

Necessary for numerous indicators of restoration progress/health

Transparent monitoring (pledge verification, etc)

Example display showing delineated tree crowns and annotated downstream products

7 of 25

Data modalities for TCD

Green 550±40 nm

Red 660±40 nm

Red Edge 735±10 nm

NIR 790±40 nm

Citrus trees in multi-spectral (Parrot Sequoia)

LIDAR point cloud (DJI)

RGB Images

Also hyper-spectral, thermal, etc

8 of 25

What’s been done?

  • Pre-deep-learning-era
    • Lots of rule-based methods
    • Habitat-specific, hard to scale globally
  • Deep learning
    • Generally works a lot better
    • Requires applicable training data
  • DeepForest (Weecology, University of Florida)
    • Only US sites
    • High quality documentation, open-source
    • Bounding boxes, not masks
  • Lots of smaller studies on single-species sites, e.g. plantations
    • Works very well if your site only has a few species
    • Label a few examples and fine-tune model

DeepForest uses LIDAR-derived CHM to initially

Label RGB imagery

9 of 25

Our Tree Mask Dataset

Current dataset using Open Aerial Map

  • Around 11k drone surveys, resampled to 10 cm/px, 2048 x 2048 tiles
  • Image extent varies from 1 to 1000 tiles, around 450k total
  • Test data: stratified subset chosen for geographic and habitat diversity
  • Train data: in principle everything else

Criteria for inclusion

  • Reasonable quality: no strong blur, stitching/warping artifacts, etc

Over 400k individually labelled trees in our train/test/val split

About $5 per image labelled, takes hours!

A variety of real images from our data. Mixtures of urban, natural and plantation type forests. Some easy to label, some not!

10 of 25

Primary labelling challenge: closed canopy

  • Imagery often contains regions with both open and closed canopy
  • Closed canopy is difficult to manually segment in RGB
  • Choosing only open canopy images massively reduces available data
  • Current strategy: label both
  • Hierarchical approach, perform secondary processing on complex regions
  • Also:
    • Differentiating between plants/bushes that look like trees
    • How sure are you it’s one tree?
    • Ground truth is very difficult to obtain

Open Canopy

Closed Canopy

11 of 25

Tree detection workflow

In: orthomosaic produced from user-submitted images - many ha

TCD Instance Segmentation Model (Mask-RCNN)

Out: 10 cm map showing closed canopy and open canopy/individual tree locations

Canopy coverage

Tree count, stem density, etc.

Crown statistics

Spectral diversity

Derived products

Predict over tiles and merge

Open

canopy

+ Others

12 of 25

Example model results

The model can also identify closed canopy (yellow)

Performance on open canopy in high-contrast

environments is good (as expected)

13 of 25

Example model results on a large image using tiling

Almost all trees are detected, even small ones, but

some issues in tile merging (false positive canopy)

An unseen acquired using the monitoring kit (Mick dos Santos, ERC)

14 of 25

Model performance (test dataset)

mAP is standard, but not a good metric because it doesn’t penalise false positives (overcounting)

Probabilistic Data Quality (Hall et. al, WACV 2020). Best confidence threshold is around 0.6-0.7

Around 75 %

accuracy - comparable to literature

15 of 25

Label QA and lessons learned (ongoing)

Labellers are not experts (so QA is required)

Does this matter?

Instructions must be simple in order to get consistent results

Can ~enforce consistency in post-processing

Trade-off between price/image/time incentives

Tricky, save budget for critical images (test set), active learning, etc.

ML pre-labelling very useful to save time …

… but relies on labellers to understand when the model is correct or not

Our internal web app to QA results from the model versus ground truth

16 of 25

Downstream products

Requires geo-referencing

  • Canopy area statistics
  • Density
  • Metrics derived from canopy area
    • Trunk diameter, if a model exists
    • Allometry
    • Other proxy measurements
  • Temporal analysis at tree-level

Requires 3D from LIDAR/Photogrammetry

  • Tree height/volume

Doesn’t require geo-referencing

  • Number of trees in a plot
    • If images cover entire polygon
    • Open canopy…
  • Spectral diversity
  • Crown morphology/shape
  • Canopy openness/ground cover
  • Temporal analysis at plot-level

With more spectral information

  • Species
  • RGB as a proxy?

17 of 25

Scaling to satellite

  • Repeat drone data acquisition is not scalable, but satellite imagery is

  • Data Science Masters Project: performance analysis of our TCD model as a function of image spatial resolution

  • Can we get good results at 0.5 m? 1 m? 3 m? How does the model transfer to different sensors?

  • ETH has access to Planet’s archive of satellite imagery, other data will be provided by Restor

Zurich, yesterday in 3m resolution - yup, it was a cloudy day (copyright Planet)

18 of 25

What raw data is required (ideally)?

Cheap/Free

High temporal

resolution

High spatial

resolution

6-12 days

10 m

Free

Daily(!)

3 m

$$

0.3-0.5 m, $$$

(Strong) Opinion - satellite imagery is the only practical solution for transparent, high-resolution, high-temporal frequency monitoring at the global scale

High resolution - < 1m/px to see most mature open canopy trees

High time resolution - monthly-yearly

Low cost - could be government funded, or provided as a data layer

19 of 25

What raw data is required (ideally)?

Cheap/Free

High temporal

resolution

High spatial

resolution

6-12 days

10 m

Daily(!)

3 m

0.3-0.5 m

Beckenhof

STF

ESA Sentinel 2A

~10 m Captured 2022-10-18

20 of 25

What raw data is required (ideally)?

Cheap/Free

High temporal

resolution

High spatial

resolution

6-12 days

10 m

Daily(!)

3 m

0.3-0.5 m

Beckenhof

STF

Planetscope

~3 m Captured 2022-10-18

21 of 25

What raw data is required (ideally)?

Cheap/Free

High temporal

resolution

High spatial

resolution

6-12 days

10 m

Daily(!)

3 m

0.3-0.5 m

Beckenhof

STF

ESRI Wayback/Maxar Worldview

0.3 m(?) Captured 2022-10-12

22 of 25

Upcoming Work

  • Evaluate on user-submitted data (field surveys) and against OAM archive

  • More field work: Restor has several ongoing field studies and is working with partners to get data for representative sites

  • ETH Data Science Masters, Fall 2022, 9 students working on derived products and satellite scaling

  • Continued label QA re-train on cleaned dataset in near future�
  • Initial model live in platform and public data release

Typical full OAM image, 4 gigapixels!

23 of 25

Live tree crown prediction demo

Imagery from Justdiggit

Site: Bumila, Central Tanzania

Input size: 1024 x 1024 px

Hardware: Nvidia GTX3090

Prediction speed: ~5-7 fps

Model: Resnet50-Mask-RCNN

Credit: inTime.de/Justdiggit

https://docs.google.com/presentation/d/18HaHhAPR4Pq3nOlFcnWbcguuMDmtVjLhPdydlPWmO2I/edit

24 of 25

Live tree canopy prediction demo

Imagery from Justdiggit

Site: Chaludewa, Central Tanzania

Input size: 1024 x 1024 px

Hardware: Nvidia GTX3090

Prediction speed: ~7-9 fps

Model: UNet-Resnet34

Credit: Justdiggit

25 of 25

Thank you! Any questions?

Feel free to email: jveitch@inf.ethz.ch / josh@restor.eco

Visit: restor.eco!