ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep

Learning and Aerial Imagery

Gyri Reiersen1,2 , David Dao2 , Björn Lütjens3 , Konstantin Klemmer4,5 , Kenza Amara2 , Attila
Steinegger6 , Ce Zhang2 , Xiaoxiang Zhu1
1
Technical University of Munich
2
ETH Zurich
3
Massachusetts Institute of Technology
4
University of Warwick
arXiv:2201.11192v2 [cs.CV] 27 Nov 2024

5
New York University
6
WWF Switzerland
[email protected], [email protected]
Abstract of global anthropogenic emissions and contributes to driv-
ing up atmospherical carbon levels (IPCC 2019). Forests,
Forest biomass is a key influence for future climate, and especially tropical forests, also provide habitats for 80% of
the world urgently needs highly scalable financing schemes, land-based biodiversity and with the increasing risk and fre-
such as carbon offsetting certifications, to protect and restore
forests. Current manual forest carbon stock inventory meth-
quency of wildfires, droughts, and extreme weather, forest
ods of measuring single trees by hand are time, labour, and ecosystems are under severe pressure (Shi et al. 2021).
cost intensive and have been shown to be subjective. They To avoid planetary tipping points (Rockstöm et al. 2009)
can lead to substantial overestimation of the carbon stock and and maintain a stable and livable climate, mankind urgently
ultimately distrust in forest financing. The potential for im- need to reduce carbon emissions until 2050 and restore es-
pact and scale of leveraging advancements in machine learn- sential ecosystems (IPCC 2021). Forests and natural carbon
ing and remote sensing technologies is promising, but needs sequestration are important climate change mitigation strate-
to be of high quality in order to replace the current forest gies (Canadell and Raupach 2008) with a biophysical miti-
stock protocols for certifications.
gation potential of 5,380 MtCO2 per year on average until
In this paper, we present ReforesTree, a benchmark dataset 2050 (IPCC 2019).
of forest carbon stock in six agro-forestry carbon offsetting
sites in Ecuador. Furthermore, we show that a deep learning- Forestry is a large industry and the causes of deforesta-
based end-to-end model using individual tree detection from tion are mostly economically driven (FAO 2020) (Geist and
low cost RGB-only drone imagery is accurately estimating Lambin 2001). For the last 20 years, major conservation ef-
forest carbon stock within official carbon offsetting certifica- forts have been underway to mitigate and safeguard against
tion standards. Additionally, our baseline CNN model out- these losses. One of the global financing strategies is car-
performs state-of-the-art satellite-based forest biomass and bon offsets (Blaufelder et al. 2021). Initially, it started as
carbon stock estimates for this type of small-scale, tropical the Clean Development Mechanism (CDM) under the Kyoto
agro-forestry sites. We present this dataset to encourage ma- Protocol, allowing governments and business organizations
chine learning research in this area to increase accountabil- from industrialized countries to invest in forestry in devel-
ity and transparency of monitoring, verification and reporting
(MVR) in carbon offsetting projects, as well as scaling global
oping countries by buying carbon credits to offset industrial-
reforestation financing through accurate remote sensing. ized emissions (FAO 2020) Several other independent bod-
ies have later developed official standards for verifying and
certifying carbon offsetting projects, such as the Gold Stan-
Introduction dard (GS) and the Verified Carbon Standard (VERRA). The
certification process for forest carbon offsetting projects is
The degradation of the natural world is unprecedented in hu- capital and labour intensive, especially due to the high cost
man history and a key driver of the climate crisis and the of manual monitoring, verification and reporting (MVR) of
Holocene extinction (Ceballos and Ehrlich 2018). Forests the forest carbon stock.
play a significant role in the planet’s carbon cycle, directly
impacting local and global climate through its biogeophysi- The carbon offsetting market is rapidly increasing and
cal effects and as carbon sinks, sequestering and storing car- expected to grow by a factor of 100 until 2050 due to
bon through photosynthesis (Griscom et al. 2017). high demand and available capital (Blaufelder et al. 2021).
However, the main obstacle is limited supply of offsetting
However, since the year 2000, we have lost 361 million ha
projects as forest owners lack upfront capital and market ac-
of forest cover, equivalent to the size of Europe, mainly in
cess (Kreibich and Hermwille 2021).
tropical areas (Hansen et al. 2013). This accounts for 18%
Recent research investigations (Badgley et al. 2021; West
Copyright © 2022, Association for the Advancement of Artificial et al. 2020) have shown that the current manual forest carbon
Intelligence (www.aaai.org). All rights reserved. stock practices systematically overestimate forestry carbon
offsetting projects with up to 29% of the offsets analyzed,
totaling up to 30 million tCO2e (CO2 equivalents) and worth
approximately $410 million. The overestimation was identi-
fied to come from subjective estimations and modeling of
the carbon stock baseline and of the project’s additionally
and leakage reporting. There is thus a need for higher quality
carbon offsetting protocols and higher transparency and ac-
countability of the MVR of these projects (Haya et al. 2020).
There are three key aspects that are important for the use
of remote sensing in MVR of forest carbon stock. One as-
pect is financial; using available and accessible technology
and sensors to lower the cost and upfront capital require-
ments for forest owners to get certified, especially in low
and middle-income countries. The second aspect is reducing
subjectivity in estimating carbon stock and increasing trust-
worthiness and transparency in the carbon offsetting certifi-
cation protocols. And lastly, the solutions need to be scalable Figure 1: Drone imagery of each site of the ReforesTree
due to the urgency of financing forest restoration, especially dataset with a resolution of 2cm/px. The red dots are the
in tropical regions. locations of the trees measured in field surveys, plotted to
Various verification bodies, new ventures, and academia make clear that the coverage of drone images were larger
are currently developing remote sensing technologies to au- than the field measured area.
tomate parts of the certification process of forestry carbon
offsetting projects (Narine, Popescu, and Malambo 2020;
Dao et al. 2019). Satellite imagery is increasing in qual-
ity and availability and, combined with state-of-the-art deep To summarize, with ReforestTree, we contribute the
learning and lidar, promises to soon map every tree on earth following: 1) the first publicly available dataset of trop-
(Hanan and Anchang 2020) and to enable forest above- ical agro-forestry containing both ground truth field data
ground biomass and carbon to be estimated at scale (Saatchi matched with high resolution RGB drone imagery at the in-
et al. 2011; Santoro et al. 2021). Compared to current man- dividual tree level and 2) a methodology for reducing the
ual estimates, these advancements reduce time and cost and current overestimation of forest carbon stock through deep
increase transparency and accountability, thus lowering the learning and aerial imagery for carbon offsetting projects.
threshold for forest owners and buyers to enter the mar-
ket (Lütjens, Liebenwein, and Kramer 2019). Nevertheless, Related Work
these algorithms risk additionally contributing to the system-
atic overestimation of carbon stocks, not reducing it, and Deep Learning for Remote Sensing
are not applicable for small-scale forests, below 10,000 ha In recent years, deep learning (DL), and especially deep con-
(White et al. 2018), (Global Forest Watch 2019). volutional neural networks (CNN) are increasing in popular-
Accurately estimating forest carbon stock, especially for ity for image analysis in the remote-sensing community (Ma
small scale carbon offset projects, presents several interest- et al. 2019), (Zhu et al. 2017). With the increase in compu-
ing machine learning challenges, such as high variance of tation power, larger datasets, transfer learning, and break-
species and occlusion of individual tree crowns. There are throughs in network architecture, DL models have outper-
many promising approaches, such as hyperspectral species formed conventional image processing methods in several
classification (Schiefer et al. 2020), lidar-based height mea- image tasks such as land use and land cover (LULC) classifi-
surements (Ganz, Käber, and Adler 2019) and individ- cation, segmentation and detection. Examples of deep super-
ual tree crown segmentation across sites (Weinstein et al. vised learning in remote sensing are the prediction of wild-
2020b). However, these applications have been developed fires (Yang, Lupascu, and Meel 2021), detection of invasive
mainly on datasets from temperate forests and, to the knowl- species (Bjorck et al. 2021). CNNs offer feature extraction
edge of the authors, there is no publicly available dataset of capabilities in recognizing patterns in both spatial and tem-
tropical forests with both aerial imagery and ground truth poral data, even with low resolution inputs. With recent ad-
field measurements. vances in meta and few shot learning these models can be
Here, we present ReforesTree, a dataset of six tropical trained and generalized on larger datasets and fine-tuned for
agroforestry reforestation project sites with individual tree local variance.
crown bounding boxes of over 4,600 trees matched with
their respective diameter at breast height (DBH), species, Manual Forest Inventory
species group, aboveground biomass (AGB), and carbon The standardized forest carbon stock inventory consists of
stock. This dataset represents ground truth field data mapped manually measuring and registering sample trees of a project
with low-cost, high-resolution RGB drone imagery to be site. Tree metrics such as diameter at breast height (DBH),
used to train new models for carbon offsetting protocols and height, and species are then put through scientifically devel-
for benchmark existing models. oped regression models called allometric equations to cal-
 S ITE N O . OF N O . OF S ITE TOTAL TOTAL
NO . T REES S PECIES A REA AGB CO2 E
1 743 18 0.51 8 5
2 929 22 0.62 15 9
3 789 20 0.48 10 6
4 484 12 0.47 5 3
5 872 14 0.56 15 9
6 846 16 0.53 12 7
TOTAL 4463 28 3.17 66 40

Figure 2: The standard procedure for calculating the correct Table 1: Overview of the six project sites in Ecuador,
amount of carbon offsets to be certified for a reforestation as gathered in field measurements. Aboveground biomass
project. The tree metrics are collected from manual forest (AGB) is measured in metric tons and area in hectares.
inventory.

for standard forestry inventory tasks such as individual tree

crown detection (Weinstein et al. 2019), lidar-based height
culate the aboveground biomass (AGB) as seen in Figure 2.
estimation (Ganz, Käber, and Adler 2019), and species clas-
The total biomass of a forest is the total AGB added with
sification (Miyoshi et al. 2020; Schiefer et al. 2020; Mäyrä
the below-ground biomass (BGB), calculated using a root-
et al. 2021), using deep learning models and aerial imagery.
to-shoot ratio specific to the forest type and region (Ma et al.
This shows high potential for combining high-resolution
2021).
imagery with deep learning models as a method for accu-
The procedure how to calculate the correct amount of car- rate carbon stock estimation for small-scale reforestation
bon offsets (CO2 e) to be certified for a project is standard- projects (Sun and Liu 2019).
ized through (Pearson, Walker, and Brown 2005) as shown
in Figure 2. The (CO2 e), also known as the baseline forest As most tropical forests are situated in low to middle in-
carbon stock, is equivalent of the total biomass divided by come countries, without access to hyperspectral, lidar and
two. Despite being prone to error propagation (Petrokofsky other more advanced sensors, the models need to be devel-
et al. 2012; Malhi et al. 2004) and shown to systematically oped using available technologies. A trade-off for accuracy
overestimate carbon stock (Badgley et al. 2021), this is cur- and data availability is basic high-resolution RGB drone im-
rently the standardized forest inventory method for certifica- agery. Drone imagery (1-3cm/px resolution), combined with
tion of forestry projects. CNN, has previously been used to directly estimate biomass
and carbon stock in individual mangrove trees (Jones et al.
Related Methods and Models 2020) or indirectly by detecting species or tree metrics such
as DBH or height (Nåfält 2018; Omasa et al. 2003), achiev-
The following are three types of methods to estimate forest ing an accuracy similar to manual field measurements. And
carbon stock remotely, adapted from (Sun and Liu 2019); by leveraging multi-fusion approaches (Du and Zare 2020;
1) inventory-based models, based on national and regional Zhang 2010), e.g. combining low-resolution satellite, high-
forest inventories and regression models, are known to over- resolution drone imagery, and field measurements and con-
estimate due to over-representations of dense commercial textual ecological or topological data, and multi-task learn-
forests in the data, (Global Forest Watch 2019). 2) Satellite- ing (Crawshaw 2020), e.g. tree metrics and carbon storage
based models leveraging datasets from optical remote sens- factors as auxiliary tasks, these models can replace and scale
ing, synthetic aperture radar satellites (SAR), and lidar (Li- the existing manual forest inventory.
DAR) to create global aboveground biomass and carbon
maps (Santoro et al. 2021; Saatchi et al. 2011; Spawn, Sul- There are several datasets for tree detection and clas-
livan, and Lark 2020). 3) Ecosystem-based models using to- sification from drone imagery such as the NEON dataset
pography, elevation, slope, aspect, and other environmental (Weinstein et al. 2020a), or the Swedish Forest Agency
factors to construct statistical models and quantitatively de- mainly from temperate forests from the US or Europe. To
scribe the process of forest carbon cycle to estimate forest our knowledge, there are no publicly available datasets in-
carbon stock(Ma et al. 2021). cluding both field measurements and drone imagery of het-
The most scalable and affordable of these methods are, erogeneous tropical forests.
evidently, satellites-based models. Nevertheless, these mod-
els and global maps are yet to estimate carbon stock at lo-
cal scale and provide accurate estimates of highly heteroge- Dataset and Method
neous and dense forest areas due to their low resolution of
30-300m (Bagheri, Shataee, and Erfanifard 2021). An indi- The ReforesTree dataset consists of six agro-forestry sites
vidual tree-based model that takes the individual overstory in the central coastal region of Ecuador. The sites are of dry
trees into account can provide this accuracy, especially if tropical forest type and eligible for carbon offsetting certi-
fused with geostatistical and satellite data. fication with forest inventory done and drone imagery cap-
In recent years, researchers have achieved high accuracy tured in 2020. See Table 1 for information on each site.
Forest Inventory Data and Drone Imagery
Field measurements were done by hand for all live trees
and bushes within the site boundaries and include GPS loca-
tion, species, and diameter at breast height (DBH) per tree.
Drone imagery was captured in 2020 by an RGB camera
from a Mavic 2 Pro drone with a resolution of 2cm per pixel.
Each site is around 0.5 ha, mainly containing banana trees
(Musaceae) and cacao plants (Cacao), planted in 2016-2019.

AGBf ruit = 0.1466 ∗ DBH 2.223 (1)

2.13
AGBmusacea = 0.030 ∗ DBH (2)
AGBcacao = 0.1208 ∗ DBH 1.98 (3)
2
AGBtimber = 21.3 − 6.95 ∗ DBH + 0.74 ∗ DBH (4)
The aboveground biomass (AGB) is calculated using
published allometric equations for tropical agro-forestry,
namely Eq.1 for fruit trees, including citrus fruits (Segura,
Kanninen, and Suárez 2006), Eq.2 banana trees (Van Noord-
wijk et al. 2002), Eq.3 for cacao (Yuliasmara, Wibawa, and
Prawoto 2009), and Eq.4 for shade trees (timber) (Brown
and Iverson 1992). These are commonly used in global cer-
tification standards. The carbon stock is calculated through Figure 3: The raw data and data processing pipeline for the
the standard forest inventory methodology using a root-to- ReforesTree dataset, resulting in labels matched to bounding
shoot ratio of 22%, which is standard for dry tropical refor- boxes per tree.
estation sites (Ma et al. 2019).

Data Processing and Method

The raw data is processed in several steps as seen in Figure 3.
The goal of this process is to have a machine learning ready
dataset that consists of matched drone image of an individual
tree with the trees labels, such as AGB value. All the drone
images have been cropped to fit tightly the boundaries of the
field measured areas. The details of this cropping process,
and the code repository, are in the Appendix.
Initially the RGB orthomosaics are cut into 4000×4000
tiles and sent through DeepForest, a python package for
predicting individual tree crowns from RGB imagery (We-
instein et al. 2019), fine-tuned on some manually labelled
bounding boxes from the sites. Afterwards, the bounding
boxes containing more than 80% white were filtered out, e.g.
bounding boxes lying on the border of the drone imagery,
and manually labeled to banana and non-banana, due to the
easily recognizable characteristics of banana trees, resulting
in clear bounding boxes of all trees as shown in Figure 4.
To fuse the tree information extracted from the ground
measurements with the bounding boxes of the trees de-
tected, we used OneForest, a recent machine learning ap-
proach for fusing citizen data with drone imagery. To re-
Figure 4: Bounding box annotations per tree, as a result
move noise introduced in both GPS locations, OneForest
of fine-tuned DeepForest tree crown detection and manual
uses a greedy optimal transport algorithm. This is a known
cleaning. Red boxes represent banana trees and blue boxes
coupling method to map between two GPS positions (center
represent other species.
of bounding box from drone imagery and GPS location of
tree from field data). Developed by Villani (Villani 2003),
the methods finds the minimal distance between two distri-
butions via a convex linear program optimizing for a match-
ing that moves the mass from one distribution to the other
with minimal cost. The cost is usually defined as the eu-
clidean distance or the Kulback-Leibler divergence between
the distributions. The optimum, i.e. the minimal distance be- S ITE F IELD GFW S PAWN S ANTORO BASELINE
NO . DATA 2019 2020 2021 (O URS )
tween the two distributions, is called the Wasserstein metric.
1 8 99 97 36 7
Baseline CNN Model 2 15 108 130 42 8
3 10 36 206 15 15
With a dataset of matched bounding boxes and tree labels, 4 5 5 102 32 9
we fine-tuned a basic pre-trained CNN, ResNet18 (He et al. 5 15 73 352 12 11
2015) with a mean-square-error loss to estimate individual 6 12 26 91 72 15
tree AGB. The results were satisfying despite the simple TOT. 66 331 413 89 65
baseline model, and proves that the individual tree estima-
tion from drone imagery has potential.
Fourteen images were identified as being larger than the Table 2: The benchmark results from comparing different
expected crown size of a tree, and they were center cropped models for estimating AGB with the forest inventory of the
at 800×800. To preserve the crown size information, the ReforesTree sites. All numbers are given as AGB in Mg.
smaller images were zero-padded up to 800×800, before all GFW is (Global Forest Watch 2019), Spawn is (Spawn, Sul-
images were resized to fit the network architecture. livan, and Lark 2020), Santoro is (Santoro et al. 2021). All
The dataset has is unbalanced with regards to species, of these three are satellite-based. Lastly, the baseline CNN
of which 43% is cacao and 32% is banana. Additionally, is our drone-based model.
due to the trees being planted between 2016-2019, many of
the trees have similar size (e.g. DBH) and half of the trees
have DBH between 7-10cm. The training dataset consisted Conclusions and Future Work
of equal number of samples of species and DBH, and from
the different project sites. We introduce the ReforesTree dataset in hopes of encour-
aging the fellow machine learning community to take on
Experiments the challenge of developing low-cost, scalable, trustworthy
With the emerging new biomass maps and forest stock esti- and accurate solutions for monitoring, verification and re-
mation models, we used the ReforesTree dataset to bench- porting of tropical reforestation inventory. We also present
mark these maps and compare with our baseline CNN model an outlined methodology for creating an annotated machine
for AGB estimation. We compared the maps taken from learning dataset from field data and drone imagery, and
(Global Forest Watch 2019), (Spawn, Sullivan, and Lark train a baseline CNN model for individual tree aboveground
2020), and (Santoro et al. 2021). The Global Forest Watch’s biomass estimation. This methodology includes a data pro-
Above-Ground Woody Biomass dataset is a global map of cessing pipeline leveraging a fine-tuned tree crown detection
AGB and carbon density at 30m×30m resolution for the algorithm and an optimal transport matching algorithm for
year 2000. It is based on more than 700,000 quality-filtered reduction of GPS noise.
Geoscience Laser Altimeter System (GLAS) lidar observa- The ReforesTree dataset of field measurements and low-
tions using machine learning models based on allometric cost, high-resolution RGB drone imagery represents the
equations for the different regions and vegetation types. The trade-off for accuracy and data availability of remote sensing
second dataset from (Spawn, Sullivan, and Lark 2020) is of forest carbon stock in tropical regions. It can be used to
a 300m×300m harmonized map based on overlayed input train new or benchmark existing models for MVR of carbon
maps. The input maps were allocated in proportion to the offsetting reforestation protocols. Remote inventory of small
relative spatial extent of each vegetation type using ancillary scale tropical reforestation projects comes with several eco-
maps of tree cover and landcover, and a rule-based decision logical challenges, such high biodiversity, level of canopy
schema. The last, and most recent 100m×100m dataset from closure, and topology. This dataset is a start to develop a
(Santoro et al. 2021) is obtained by spaceborne SAR (ALOS generalized model that can be fine-tuned on local scale. Fu-
PALSAR, Envisat ASAR), optical (Landsat-7), lidar (ICE- ture work will investigate ways to improve the methodology
SAT) and auxiliary datasets with multiple estimation proce- and reduce error in the machine learning ready dataset, and
dures with a set of biomass expansion and conversion factors increase the explainability to have a trustworthy and trans-
following approaches to extend ground estimates of wood parent model. Additionally, we see further potential in fus-
density and stem-to-total biomass expansion factors. ing satellite and other available geoecological data layers as
As seen in Table 2, all of the available global AGB maps well as leveraging the multiple labels available (e.g. DBH,
have a tendency to overestimate the ground truth measure- species) as auxiliary tasks in a multitask learning problem.
ments up to a factor of ten. These are not encouraging results As the world is rapidly approaching planetary doom, we
showing that these maps are far from being accurate enough need to collaborate across disciplines to implement and
to be used in remote sensing of forest carbon stock at a small scale the climate mitigation strategies available. Restoration
scale, as is the case for the ReforesTree dataset. of forests is one of our most important climate mitigation
Our baseline model, on the other hand, has a slight ten- strategies. And by reducing the overestimation of carbon
dency of underestimating the biomass. The model has an offsets, we can allow every man on earth who owns a tree
evident advantage, to be trained on the dataset, but these ini- to participate in climate action. Biodiverse and sustainable
tial results show promise for the individual tree estimation forestry can provide hope not only the for the machine learn-
approach using drone imagery for forest carbon inventory. ing community, but also beyond.
 Acknowledgments FAO. 2020. Global Forest Resources Assessment 2020:
The authors are thankful for the guidance and advice by Main report. FAO. ISBN 978-92-5-132974-0.
our academic collaborator (Prof. Dava Newman, Prof. Lynn Ganz, S.; Käber, Y.; and Adler, P. 2019. Measuring
H Kaack, Prof. Thomas Crowther and the CrowtherLab), Tree Height with Remote Sensing—A Comparison of Pho-
non-governmental institutions (BrainForest, WWF Switzer- togrammetric and LiDAR Data with Different Field Mea-
land, Restor), Isabel Hillman, Simeon Max, Microsoft AI surements. Forests, 10: 694.
for Earth, and support from the local community in Ecuador. Geist, H. J.; and Lambin, E. F. 2001. What drives tropical
Lastly, we extend our sincere gratitude to Autumn Nguyen deforestation?: a meta-analysis of proximate and underly-
and Sulagna Saha for their significant contributions to this ing causes of deforestation based on subnational case study
work. Their thorough review process led to substantial im- evidence. LUCC International Project Office, University of
provements in both the manuscript and the underlying code- Louvain.
base. Their detailed technical analysis and implementations
have enhanced the robustness and reliability of our research. Global Forest Watch. 2019. Aboveground Live Woody
A comprehensive report of their contributions can be found Biomass Density. Dataset Accessed: 30.11.2021.
in our technical documentation: https://gainforest.substack. Griscom, B. W.; Adams, J.; Ellis, P. W.; Houghton, R. A.;
com/p/improving-reforestree-correcting. Lomax, G.; Miteva, D. A.; Schlesinger, W. H.; Shoch, D.;
Siikamäki, J. V.; Smith, P.; Woodbury, P.; Zganjar, C.; Black-
References man, A.; Campari, J.; Conant, R. T.; Delgado, C.; Elias, P.;
Gopalakrishna, T.; Hamsik, M. R.; Herrero, M.; Kiesecker,
Badgley, G.; Freeman, J.; Hamman, J. J.; Haya, B.; Trug- J.; Landis, E.; Laestadius, L.; Leavitt, S. M.; Minnemeyer,
man, A. T.; Anderegg, W. R.; and Cullenward, D. 2021. Sys- S.; Polasky, S.; Potapov, P.; Putz, F. E.; Sanderman, J.; Sil-
tematic over-crediting in California’s forest carbon offsets vius, M.; Wollenberg, E.; and Fargione, J. 2017. Natural
program. bioRxiv. climate solutions. Proceedings of the National Academy of
Bagheri, R.; Shataee, S.; and Erfanifard, S. Y. a. 2021. Sciences, 114(44): 11645–11650.
Canopy Based Aboveground Biomass and Carbon Stock Es-
Hanan, N. P.; and Anchang, J. Y. 2020. Satellites could soon
timation of Wild Pistachio Trees in Arid Woodlands Us-
map every tree on Earth. Nature, 587.
ing GeoEye-1 Images. Journal of Agricultural Science and
Technology, 23(1). Hansen, M. C.; Potapov, P. V.; Moore, R.; Hancher, M.; Tu-
rubanova, S. A.; Tyukavina, A.; Thau, D.; Stehman, S. V.;
Barenne, V.; Bohl, J. P.; Dekas, D.; and Engelmann, T. 2022.
Goetz, S. J.; Loveland, T. R.; Kommareddy, A.; Egorov, A.;
Tropical Forest Carbon Stock Estimation using RGB Drone
Chini, L.; Justice, C. O.; and Townshend, J. R. G. 2013.
Imagery.
High-Resolution Global Maps of 21st-Century Forest Cover
Bjorck, J.; Rappazzo, B. H.; Shi, Q.; Brown-Lima, C.; Dean, Change. Science, 342(6160): 850–853.
J.; Fuller, A.; and Gomes, C. 2021. Accelerating Ecologi-
cal Sciences from Above: Spatial Contrastive Learning for Haya, B.; Cullenward, D.; Strong, A. L.; Grubert, E.; Heil-
Remote Sensing. Proceedings of the AAAI Conference on mayr, R.; Sivas, D. A.; and Wara, M. 2020. Managing un-
Artificial Intelligence, 35(17): 14711–14720. certainty in carbon offsets: insights from California’s stan-
dardized approach. Climate Policy, 20(9): 1112–1126.
Blaufelder, C.; Levy, C.; Mannion, P.; Pinner, D.; and We-
terings, J. 2021. McKinsey&Co: A Blueprint for Scaling He, K.; Zhang, X.; Ren, S.; and Sun, J. 2015. Deep Residual
Voluntary Carbon Markets to Meet the Climate Challenge. Learning for Image Recognition. arXiv:1512.03385.
Accessed 31.05.2021. IPCC. 2019. 2019: Summary for Policymakers. In Shukla,
Brown, S.; and Iverson, L. 1992. Biomass estimates for trop- P.; Skea, J.; Buendia, E. C.; Masson-Delmotte, V.; Pörtner,
ical forest. World Res. Rev., 4: 366–383. H.-O.; Roberts, D. C.; Zhai, P.; Slade, R.; Connors, S.; van
Diemen, R.; Ferrat, M.; Haughey, E.; Luz, S.; Neogi, S.;
Canadell, J. G.; and Raupach, M. R. 2008. Managing Forests Pathak, M.; Petzold, J.; Pereira, J. P.; Vyas, P.; Huntley,
for Climate Change Mitigation. Science, 320: 1456–1457. E.; Kissick, K.; Belkacemi, M.; and Malley, J., eds., Cli-
Ceballos, G.; and Ehrlich, P. 2018. The misunderstood sixth mate Change and Land: an IPCC special report on climate
mass extinction. Science, 360: 1080.2–1081. change, desertification, land degradation, sustainable land
Crawshaw, M. 2020. Multi-Task Learning with Deep Neural management, food security, and greenhouse gas fluxes in
Networks: A Survey. terrestrial ecosystems, 7–11.
Dao, D.; Cang, C.; Fung, C.; Zhang, M.; Pawlowski, N.; IPCC. 2021. Climate Change 2021: The Physical Science
Gonzales, R.; Beglinger, N.; and Zhang, C. 2019. GainFor- Basis. Contribution of Working Group I to the Sixth As-
est: Scaling Climate Finance for Forest Conservation using sessment Report of the Intergovernmental Panel on Climate
Interpretable Machine Learning on Satellite Imagery. ICML Change. Cambridge University Press.
Climate Change AI workshop 2019. Jones, A. R.; Raja Segaran, R.; Clarke, K. D.; Waycott, M.;
Du, X.; and Zare, A. 2020. Multiresolution Multimodal Goh, W. S. H.; and Gillanders, B. M. 2020. Estimating Man-
Sensor Fusion for Remote Sensing Data With Label Uncer- grove Tree Biomass and Carbon Content: A Comparison of
tainty. IEEE Transactions on Geoscience and Remote Sens- Forest Inventory Techniques and Drone Imagery. Frontiers
ing, 58. in Marine Science, 6: 784.
Kreibich, N.; and Hermwille, L. 2021. Caught in between: C.; Schellnhuber, H.; Nykvist, B.; Wit, C. A. D.; Hughes,
credibility and feasibility of the voluntary carbon market T.; S. van der Leeuw, H. R.; Sörlin, S.; Snyder, P. K.;
post-2020. Climate Policy, 21(7): 939–957. R. Costanza, U. S.; Falkenmark, M.; Karlberg, L.; Corell,
Lütjens, B.; Liebenwein, L.; and Kramer, K. 2019. Machine R. W.; Fabry, V. J.; Hansen, J.; Walker, B.; Liverman, D.;
Learning-based Estimation of Forest Carbon Stocks to in- Richardson, K.; Crutzen, P.; and Foley, J. 2009. Planetary
crease Transparency of Forest Preservation Efforts. 2019 boundaries:exploring the safe operating space for humanity.
NeurIPS Workshop on Tackling Climate Change with Ma- Ecology and Society, 14: 32.
chine Learning. Saatchi, S. S.; Harris, N. L.; Brown, S.; Lefsky, M.;
Ma, H.; Mo, L.; Thomas W. Crowther, D. S. M.; van den Mitchard, E. T. A.; Salas, W.; Zutta, B. R.; Buermann, W.;
Hoogen, J.; Stocker, B. D.; Terrer, C.; and Zohner, C. M. Lewis, S. L.; Hagen, S.; Petrova, S.; White, L.; Silman, M.;
2021. The global distribution and environmental drivers and Morel, A. 2011. Benchmark map of forest carbon stocks
of aboveground versus belowground plant biomass. Nature in tropical regions across three continents. Proceedings of
Ecology & Evolution, 5: 1110–1122. the National Academy of Sciences, 108(24): 9899–9904.
Ma, L.; Liu, Y.; Zhang, X.; Ye, Y.; Yin, G.; and Johnson, Santoro, M. 2018. GlobBiomass - global datasets of forest
B. A. 2019. Deep learning in remote sensing applications: biomass.
A meta-analysis and review. ISPRS Journal of Photogram- Santoro, M.; Cartus, O.; Carvalhais, N.; Rozendaal, D.
metry and Remote Sensing, 152: 166–177. M. A.; Avitabile, V.; Araza, A.; de Bruin, S.; Herold,
Malhi, Y.; Phillips, O. L.; Chave, J.; Condit, R.; Aguilar, S.; M.; Quegan, S.; Rodrı́guez-Veiga, P.; Balzter, H.; Car-
Hernandez, A.; Lao, S.; and Perez, R. 2004. Error propaga- reiras, J.; Schepaschenko, D.; Korets, M.; Shimada, M.;
tion and scaling for tropical forest biomass estimates. Philo- Itoh, T.; Moreno Martı́nez, A.; Cavlovic, J.; Cazzolla Gatti,
sophical Transactions of the Royal Society of London. Series R.; da Conceição Bispo, P.; Dewnath, N.; Labrière,
B: Biological Sciences, 359(1443): 409–420. N.; Liang, J.; Lindsell, J.; Mitchard, E. T. A.; Morel,
Miyoshi, G. T.; Arruda, M. d. S.; Osco, L. P.; Marcato Ju- A.; Pacheco Pascagaza, A. M.; Ryan, C. M.; Slik, F.;
nior, J.; Gonçalves, D. N.; Imai, N. N.; Tommaselli, A. Vaglio Laurin, G.; Verbeeck, H.; Wijaya, A.; and Willcock,
M. G.; Honkavaara, E.; and Gonçalves, W. N. 2020. A Novel S. 2021. The global forest above-ground biomass pool for
Deep Learning Method to Identify Single Tree Species in 2010 estimated from high-resolution satellite observations.
UAV-Based Hyperspectral Images. Remote Sensing, 12(8). Earth System Science Data, 13(8): 3927–3950.
Mäyrä, J.; Keski-Saari, S.; Kivinen, S.; Tanhuanpää, T.; Schiefer, F.; Kattenborn, T.; Frick, A.; Frey, J.; Schall,
Hurskainen, P.; Kullberg, P.; Poikolainen, L.; Viinikka, A.; P.; Koch, B.; and Schmidtlein, S. 2020. Mapping forest
Tuominen, S.; Kumpula, T.; and Vihervaara, P. 2021. Tree tree species in high resolution UAV-based RGB-imagery by
species classification from airborne hyperspectral and Li- means of convolutional neural networks. ISPRS Journal of
DAR data using 3D convolutional neural networks. Remote Photogrammetry and Remote Sensing, 170: 205–215.
Sensing of Environment, 256: 112322. Segura, M.; Kanninen, M.; and Suárez, D. 2006. Allometric
Narine, L. L.; Popescu, S. C.; and Malambo, L. 2020. Us- models for estimating aboveground biomass of shade trees
ing ICESat-2 to Estimate and Map Forest Aboveground and coffee bushes grown together. Agroforestry Systems, 68:
Biomass: A First Example. Remote Sensing, 12(11). 143–150.
Nåfält, S. 2018. Estimating above ground biomass in a Salix Shi, H.; Tian, H.; Lange, S.; Yang, J.; Pan, S.; Fu, B.; and
plantation using high resolution UAV images. Student thesis Reyer, C. P. O. 2021. Terrestrial biodiversity threatened
series INES, Lund University:8963727. by increasing global aridity velocity under high-level warm-
ing. Proceedings of the National Academy of Sciences of the
Omasa, K.; Qiu, G. Y.; Watanuki, K.; Yoshimi, K.; and United States of America (PNAS), 18: 36.
Akiyama, Y. 2003. Accurate Estimation of Forest Carbon
Stocks by 3-D Remote Sensing of Individual Trees. Envi- Spawn, S.; Sullivan, C.; and Lark, T. e. a. 2020. Harmo-
ronmental Science & Technology, 37. nized globadgleyl maps of above and belowground biomass
carbon density in the year 2010. Sci Data, 7: 112.
Pearson, T.; Walker, S.; and Brown, S. 2005. Sourcebook
for BioCarbon Fund Projects. Accessed 15.09.2021 URL: Sun, W.; and Liu, X. 2019. Review on carbon storage esti-
https://winrock.org/document/sourcebook-for-land-use- mation of forest ecosystem and applications in China. Forest
land-use-change-and-forestry-projects/. Ecosystems, 7: 4.
Petrokofsky, G.; Kanamaru, H.; Achard, F.; Goetz, S. J.; Van Noordwijk, M.; Rahayu, S.; Hairiah, K.; Wulan, Y.;
Joosten, H.; Holmgren, P.; Lehtonen, A.; Menton, M. C. S.; Farida, A.; and Verbist, B. 2002. Carbon stock assessment
Pullin, A. S.; and Wattenbach, M. 2012. Comparison of for a forest-to-coffee conversion landscape in Sumber-Jaya
methods for measuring and assessing carbon stocks and car- (Lampung, Indonesia): from allometric equations to land use
bon stock changes in terrestrial carbon pools. How do the change analysis. Science in China, 45.
accuracy and precision of current methods compare? A sys- Villani, C. 2003. Topics in optimal transportation. 58.
tematic review protocol. Environmental Evidence, 1: 6. American Mathematical Soc.
Rockstöm, J.; Steffen, W.; K. Noone, Á. P.; Chapin, Weinstein, B. G.; Marconi, S.; Bohlman, S.; Zare, A.; Singh,
F. S.; Lambin, E.; Lenton, T. M.; Scheffer, M.; Folke, A.; Graves, S. J.; and White, E. 2020a. NEON Crowns: a
remote sensing derived dataset of 100 million individual tree
crowns. bioRxiv.
Weinstein, B. G.; Marconi, S.; Bohlman, S.; Zare, A.; and
White, E. 2019. Individual tree-crown detection in RGB im-
agery using semi-supervised deep learning neural networks.
Remote Sensing, 11(11): 1309.
Weinstein, B. G.; Marconi, S.; Bohlman, S. A.; Zare, A.;
and White, E. P. 2020b. Cross-site learning in deep learn-
ing RGB tree crown detection. Ecological Informatics, 56:
101061.
West, T. A. P.; Börner, J.; Sills, E. O.; and Kontoleon, A.
2020. Overstated carbon emission reductions from volun-
tary REDD+ projects in the Brazilian Amazon. Proceed- Figure 5: This figure represents the count of species family
ings of the National Academy of Sciences, 117(39): 24188– groups for each of the sites. All sites have trees of all species
24194. family groups, but cacao and banana are over represented.
White, A. E.; Lutz, D. A.; Howarth, R. B.; and Soto, J. R.
2018. Small-scale forestry and carbon offset markets: An
empirical study of Vermont Current Use forest landowner
willingness to accept carbon credit programs. PLOS ONE, Aligning Drone Images with Field Boundaries
13(8): 1–24. A key issue identified in the ReforesTree pipeline was the
Yang, S.; Lupascu, M.; and Meel, K. S. 2021. Predict- mismatch between the drone imagery boundaries and the
ing Forest Fire Using Remote Sensing Data And Machine field data boundaries. To address this, we implemented the
Learning. arXiv:2101.01975. following steps to align the drone images with the field mea-
surements for the six agroforestry sites. The code for this is
Yuliasmara, F.; Wibawa, A.; and Prawoto, A. 2009. Carbon
in this reforestree-correction repository.
stock in different ages and plantation system of cocoa: al-
lometric approach. Pelita Perkebunan (a Coffee and Cocoa 1. GeoDataframe Creation: We converted the field data,
Research Journal), 26. which included the longitude and latitude of each point,
Zhang, J. 2010. Multi-source remote sensing data fusion: into a GeoDataFrame using the geopandas library. This
status and trends. International Journal of Image and Data allowed us to create point geometries that were easy to
Fusion, 1. visualize and manipulate. The field data points, visual-
ized as red dots in Figure 6, served as the starting refer-
Zhu, X. X.; Tuia, D.; Mou, L.; Xia, G.-S.; Zhang, L.; Xu, F.; ence.
and Fraundorfer, F. 2017. Deep Learning in Remote Sens-
2. Boundary Extraction using Alpha Shape: To capture
ing: A Comprehensive Review and List of Resources. IEEE
the boundary of the field data, we used the alphashape li-
Geoscience and Remote Sensing Magazine, 5(4): 8–36.
brary to create a convex hull around the points. By choos-
ing an alpha value of 15000, similar to the value used
Technical Appendix by (Barenne et al. 2022), we generated a tight boundary
Raw data cleaning around the field data points.
3. Overlay and Crop Drone Imagery: Using the rasterio
All 28 species were divided into 6 species family groups: library, we overlapped the generated alphashape bound-
banana, cacao, fruit, timber, citrus and other. ary onto the drone imagery (in TIFF format). We then
The field data was manually collected as a standard man- cropped the unnecessary parts of the image, outside the
ual forest inventory, potentially leading to human errors, boundary, replacing them with white pixels. This step is
missing values and outliers. illustrated by the transition from the third to fourth im-
The dataset needed to reflect the ground truth. Therefore ages in Figure 6.
it was important not to remove trees from the dataset un-
necessarily. All missing DBH values were given a value 4. Adjusting Image Boundaries: Finally, after cropping,
based on the average DBH of the same species for the year we identified the bounds of the non-white pixels in the
it was planted. Of the 28 species, only 3 species (in total images and adjusted them to ensure they fit a square
25 trees) were missing DBH values: 23 lemon (citrus), one shape correctly. This was essential for integrating the im-
balsa (timber), one bariable (other) trees. These were given ages into the AGBench library. The final result can be
DBH values interpolated from the other trees in the same seen in the transition from the fourth to the last image in
family group and which were planted the same year. Figure 6.
Additionally, 8 banana trees that had DBH values larger
than 50cm, which is unrealistically high. Assuming that Benchmark of satellite-based AGB maps
there was a manual entry mistake, these values were ex- To benchmark the low resolution (LR) satellite-based maps,
changed with the maximum value of the banana trees for we fitted it to the high resolution (HR) drone imagery over-
the year planted. lapping the GPS coordinates.
 • (Spawn, Sullivan, and Lark 2020): Global Aboveground
and Belowground Biomass Carbon Density Maps for the
Year 2010 with 300x300m resolution.
• (Santoro 2018): GlobBiomass - Global Datasets of Forest
Biomass with 100x100m resolution for the year 2010.
Figure 6: This figure shows the alignment process between Baseline CNN
the drone images and field boundaries. The field data points
(red dots) were used to create the alphashape, which was We trained the model on a single GPU of the type GeForce
overlaid onto the drone imagery to crop unnecessary areas RTX 3090. The learning rate used was 1e-3 , batch size of 64
and ensure accurate alignment. for 30 epochs achieving a root square mean loss (RMSE) of
0,1.

Figure 7: This figure represents the different steps in the

benchmark analysis and how we calculated the total AGB
amount from the satellite-based maps for the ReforesTree
sites. This is taken from site no. 0. The values represented in
the image is AGB density.

The calculation of the total AGB was done in five steps,

illustrated in Figure 7
1. cropping the LR satellite map with a padding around
the polygon of the site to reduce computation intensity
(Satellite Raw)
2. linearly interpolating the values for this map and resize
the map with the same HR pixel resolution as the drone
imagery (Satellite Interpolated)
3. cropping the map further fitting with the GPS locations
(max/min) of the drone imagery
4. filtering out the site area by removing all pixels in the
satellite-based map, that are outside of the drone imagery,
coloured white (Satellite Filtered)
5. lastly, multiplying the AGB mean density of the filtered
map with the project site area to get the total AGB
We analysed the following three maps:
• (Global Forest Watch 2019): Aboveground Woodly
Biomass with 30x30m resolution for the year of 2000.