Data Products

Methodology
Version 2.0.2 - March 2024

hello@[Link]
Kanop, Paris, France
Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Table of contents
Abstract............................................................................................................................................3
Graphical abstract........................................................................................................................... 4
I. Introduction...................................................................................................................................5
II. Data products.............................................................................................................................. 7
III. Data sources...............................................................................................................................8
IV. Methods.................................................................................................................................... 12
Image selection optimization................................................................................................. 12
Optical multispectral Imagery.......................................................................................... 12
Optimal months selection................................................................................................ 12
Cloud filtering.................................................................................................................... 12
SAR pre-processing...........................................................................................................13
Inference.................................................................................................................................. 14
Post-calibration....................................................................................................................... 15
Gaussian process with field measurements................................................................... 15
Height corrections............................................................................................................ 15
Forest-cover.............................................................................................................................16
Allometric relationships..........................................................................................................16
Belowground biomass calculation.................................................................................. 17
Total biomass calculation................................................................................................ 19
Carbon stock calculation..................................................................................................19
CO2 equivalent calculation...............................................................................................19
Biodiversity indicators.............................................................................................................19
Rao’s Q diversity index...................................................................................................... 19
Confidence interval................................................................................................................. 19
V. Validation................................................................................................................................... 21
GEDI L2A comparison.............................................................................................................22
Global Forest Canopy Height comparison.............................................................................23
GEDI L4B comparison............................................................................................................. 23
Airborne LiDAR- PALSAR-derived AGB comparison..............................................................24
VI. References................................................................................................................................25

Victor Allory (PhD), Louis De Vitry, Colette Gelas (PhD), Coby Strell, Romain Fau. Detailed
methodology and validation of Kanop's Screening and Standard data products for measuring
ecosystem services provided by nature-based solutions. 2024.

No part of this document may be reproduced in any form or any means without the prior
written consent of Kanop.

© RAIL SAS 2024 Page 2/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Abstract
In this document, Kanop presents a comprehensive and detailed methodology delineating the
processes utilized to generate precise estimates of ecosystem services provided by
nature-based solutions. Kanop leverages cutting-edge scientific approaches, including
aboveground biomass and height models, to derive a multitude of key performance
indicators.

The Kanop data production process is fully automatic and starts with the selection of remote
sensed images. Optimal optical and Synthetic Aperture Radar (SAR) satellite data are
carefully chosen during the dry season to mitigate the impact of moisture and cloud cover,
ensuring the reliability and accuracy of the generated data. Following selection, these images
undergo thorough pre-processing, especially in the case of SAR, to prepare the data for
subsequent modeling processes.

Kanop then utilizes these pre-processed images to estimate primary indicators such as
canopy height, tree height, canopy cover, and living aboveground biomass. These indicators
are estimated with state-of-the-art trained models using airborne and spaceborne LiDAR, well
recognized reference datasets.

The resulting data is then subject to post-calibration to ensure alignment with user-provided
field measurements, employing a Gaussian process. This calibration remains effective even
when field measurements are from different years or nearby areas of interest.

From the primary indicators are derived the secondary indicators. This is done by using
regional specific allometric equations and definitions recommended by standards. At Kanop,
we know each project is unique so we have made allometric relationships and definitions
customizable by the platform user. These secondary indicators are forest cover, living
belowground biomass, total biomass, carbon stocks, and their CO2 equivalents

Through the standard and screening data products, Kanop facilitates affordable access to
the most precise data within the digital Measurement, Verification, and Reporting market for
nature-based projects.

Kanop's objective is to deliver an efficient, science-driven, and transparent solution for

accurately measuring the ecosystem services of nature-based projects in a scalable manner.
We firmly believe that these components are essential to foster growth and trust within the
voluntary carbon market.

© RAIL SAS 2024 Page 3/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Graphical abstract

Figure 1: Framework of the standard data product

© RAIL SAS 2024 Page 4/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

I. Introduction
Climate change poses unprecedented challenges, from extreme weather events to rising sea
levels, threatening ecosystems, livelihoods, and economies worldwide. In this critical
juncture, forests and natural ecosystems emerge as crucial allies. These vital ecosystems
play a multifaceted role, not only as carbon sinks, sequestering vast amounts of atmospheric
carbon dioxide, but also as biodiversity hotspots and sources of livelihood for millions of
people (Mondal and Palit, 2022). Recognizing their significance, concerted efforts towards
sustainable forest management and conservation are essential components of any
comprehensive climate mitigation strategy (Canadell and Raupach, 2008; Keenan, 2015).

In harnessing the potential of forests to address the challenges of climate change, the
carbon market offers a promising avenue for incentivizing responsible stewardship of these
invaluable natural resources. The establishment of mechanisms such as REDD+ (Reducing
Emissions from Deforestation and Forest Degradation) provides financial incentives for
developing countries to conserve and sustainably manage their forests, thereby reducing
carbon emissions and promoting biodiversity conservation. Moreover, the emergence of
voluntary carbon markets further expands opportunities for forest carbon projects, allowing
businesses and individuals to offset their emissions by investing in forest conservation and
restoration initiatives.

However, the effective implementation of

forest carbon projects relies heavily on
robust Measurement, Reporting, and
Verification (MRV) systems to accurately
quantify carbon stocks and emissions
reductions. Given the complex and
dynamic nature of forest ecosystems,
ensuring the integrity and credibility of
carbon accounting processes is
paramount. This necessitates the
development of comprehensive reporting
frameworks and methodologies that
adhere to rigorous standards and are
transparent, verifiable, and compatible
with international guidelines.

In response to these challenges, digital

MRV solutions have emerged as powerful
tools for enhancing the efficiency and
reliability of forest carbon monitoring and
reporting. Leveraging advances in remote
sensing, geospatial technologies, and
data analytics, these innovative solutions enable real-time monitoring of forest cover
changes, carbon stocks, and emissions, providing stakeholders with timely and actionable

© RAIL SAS 2024 Page 5/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

insights. By automating data collection, analysis, and reporting processes, digital MRV
solutions streamline workflow management, reduce administrative burden, and improve data
accuracy and transparency.

Nevertheless, the effectiveness of digital MRV solutions hinges on the adoption of detailed
and transparent methodologies that account for the unique characteristics of different forest
types and ecosystems. Standardized protocols and guidelines are essential for ensuring
consistency and comparability across projects, while stakeholder engagement and
capacity-building efforts are crucial for promoting understanding and trust in MRV
processes. Moreover, fostering collaboration and knowledge-sharing among governments,
civil society organizations, and the private sector is essential for driving continuous
improvement and innovation in forest carbon accounting practices.

In such a context, Kanop provides a comprehensive methodology document that describes

how we measure and analyze environmental impacts with precision. Leveraging
state-of-the-art artificial intelligence (AI) and numerous satellite images, Kanop provides a
scalable and accurate assessment of ecosystem services. Our methodology is built on a
foundation of remote sensing data, using a mix of multispectral, SAR, and space-borne LiDAR
data, which allows us to monitor and evaluate changes in land use and biomass with high
precision. By combining this data with advanced machine learning models, we generate
detailed insights into the carbon sequestration and biodiversity of forested areas.

© RAIL SAS 2024 Page 6/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

II. Data products

Kanop's services are divided into two data products based on the same models, but which
differ in terms of indicators computed, number and sources of remote sensing data used and
resolution.

Table 1: Kanop data products

Screening Standard
Canopy cover (%) Canopy cover (%)
Forest cover (%) Forest cover (%)
Aboveground biomass (Mg/ha) Aboveground biomass (Mg/ha)
Indicators Belowground biomass (Mg/ha) Belowground biomass (Mg/ha)
(stock and changes) Carbon & CO2eq. (Mg/ha) Carbon & CO2eq. (Mg/ha)
Rao’s Q diversity index Rao’s Q diversity index
Tree height (m)
Canopy height (m)
Recalibration with
- Yes, optional.
field measurements
Uncertainty - 95% confidence interval
Resolution 25 m 10 or 25 m
Global coverage from 2013 Global 25 m coverage from 2013
Availability
Global 10 m coverage from 2017
LiDAR LiDAR
Airborne LiDAR Airborne LiDAR
GEDI GEDI

Optical Optical
Landsat-8 Sentinel-2
Landsat-8
SAR
Data inputs PALSAR SAR
PALSAR-2 Sentinel-1
NiSAR (late 2024) PALSAR
Biomass (2025) PALSAR-2
NiSAR (late 2024)
Biomass (2025)

Ground
Field measurements

© RAIL SAS 2024 Page 7/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

The Screening data product enables users to assess large quantities of land at a lower price
point, which makes this product particularly suitable for screening operations and feasibility
studies. The Standard data product provides uncertainty information, which can be improved
upon by recalibrating models with field measurements. This product is well suited to build
baselines and performance benchmarks, for MRV, Scope 3 reporting, communicating impact
to stakeholders, and more.

III. Data sources

Several data sources are used either in the training of our deep learning models or for the
estimation of key performance indicators for requested areas (henceforth “inference”).

More precisely, we can differentiate between data used for:

● Inference & model training: Satellite data (optical and SAR) automatically fetched
within the project area and its surroundings, used as a data input either for the
inference (estimation of key performance indicators) or for the training of the model.

● Post-calibration: When available, reliable and representative field measurements are

used to recalibrate the results of the models.

© RAIL SAS 2024 Page 8/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Table 2: Data sources, uses and descriptions

Data Use Description

Optical satellite data
Sentinel-2 - Inference Description: Sentinel-2 is a satellite mission developed by the European Space Agency (ESA)
& model training for global monitoring. It provides high-resolution optical imagery for various applications
including land cover mapping, agricultural monitoring, and disaster management.
Type of Data: Optical imagery
Resolution: 10 meters to 60 meters, depending on the spectral band
Bands: 13 spectral bands
Frequency: Approximately every 5 days
Date Range: Operational since June 2015
Landsat-8 - Inference Description: Landsat-8 is a satellite mission operated by the United States Geological Survey
& model training (USGS) and NASA. It captures medium-resolution optical imagery of the Earth's surface for
scientific research, resource management, and environmental monitoring.
Type of Data: Optical imagery
Resolution: 15 meters to 100 meters
Bands: 11 spectral bands
Frequency: Approximately every 16 days
Date Range: Operational since February 2013
SAR satellite data
Sentinel-1 - Inference Description: Sentinel-1 is a satellite mission (2 sensors) by the European Space Agency (ESA)
& model training designed for all-weather, day-and-night radar imaging. It provides data for applications such as
monitoring land use change (like agriculture or forestry), ice and ocean surfaces, land
deformation, and disaster mapping.
Type of Data: C Band Synthetic Aperture Radar (SAR) imagery
Resolution: Varies depending on the mode (from 5 meters to 40 meters)

© RAIL SAS 2024 Page 9/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Polarization: dual polar (VV and VH) with some exception area where single-polarized (HH or
VV) or quad-polarized (HH, HV, VH, VV) are available
Frequency: Approximately every 6 to 12 days
Date Range: Operational since April 2014
PALSAR & - Inference Description: PALSAR (Phased Array type L-band Synthetic Aperture Radar) and PALSAR-2 are
PALSAR-2 & model training satellite missions by the Japan Aerospace Exploration Agency (JAXA), providing L-band SAR
data for land and disaster monitoring, forestry, and topographic mapping.
Type of Data: L band Synthetic Aperture Radar (SAR) imagery
Resolution: Varies depending on the mode (from 7 meters to 100 meters)
Polarization: dual-polarized (HH and HV)
Frequency: Varies
Date Range: PALSAR: Operational from January 2006 to May 2011. PALSAR-2: Operational
since January 2014
NISAR - Inference Description: NASA-ISRO Synthetic Aperture Radar (NISAR) is a joint satellite mission between
(not available yet, & model training NASA and the Indian Space Research Organisation (ISRO). It aims to provide global
planned late 2024) observations for understanding natural hazards, ecosystem disturbances, and climate
dynamics.
Type of Data: L and S band Synthetic Aperture Radar (SAR) imagery
Polarization: quad-polarized (HH,HV,VH,VV)
Resolution: Expected to be around 5 to 30 meters
Frequency: Scheduled launch in 2024, revisit frequency will be determined later.
Additional training data
Airborne LiDAR - Model training Description: Airborne LiDAR (Light Detection and Ranging) is a remote sensing technology that
data uses laser pulses to measure the distance between the sensor and the Earth's surface. It is
used for generating high-resolution digital elevation models, mapping terrain, and vegetation
structure.
Type of Data: LiDAR waveform data
Resolution: Sub-meter to several meters

© RAIL SAS 2024 Page 10/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Spaceborne LiDAR - Model training Description: Global Ecosystem Dynamics Investigation (GEDI) Level 2A data is derived from
data spaceborne LiDAR aboard the International Space Station (ISS). It provides vertical profiles of
(GEDI L2A) forest structure and height globally.
Type of Data: LiDAR waveform data
Resolution: Around 25 meters, sparse data
Frequency: Approximately every 30 days
Date Range: Operational since April 2019
Aboveground - Model training Description: GEDI Level 4A data provides aboveground biomass (AGB) estimates derived from
biomass spaceborne LiDAR measurements. It offers information on forest biomass distribution and
(GEDI L4A) dynamics.
Type of Data: Aboveground biomass estimates
Resolution: 25 m, sparse data
Frequency: Approximately every 30 days
Date Range: Operational since April 2019
Field measurements
Field - Post-calibration Description: User-provided field measurements typically aggregate data collected from sample
measurements plots during field campaigns. Field measurements can be uploaded directly from our web app
or via the API in .csv or .geojson file formats.

© RAIL SAS 2024 Page 11/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

IV. Methods
Image selection optimization
Optical multispectral Imagery
For a more accurate representation of the Earth's surface, we opt for surface reflectance data
rather than top of atmosphere (TOA) reflectance data. Surface reflectance corrects for
atmospheric effects such as scattering and absorption by gasses and aerosols, facilitating
more consistent comparisons over time and across different locations.

Optimal months selection

Kanop uses IPCC climatic variables to identify the months with the least number of wet days
to maintain consistency in the model's output. The data used is the version 4 of the CRU TS
monthly high resolution gridded multivariate climatic dataset (Harris et al., 2020). This is
done to minimize the effect of humidity on optical and SAR satellite data (Liesenberg et al.,
2016). This optimized selection is applied on the recovering of Sentinel-1 and Sentinel-2 data,
for both training and inference purposes.

Cloud filtering
For a more accurate representation of the Earth's surface, we opt for surface reflectance data
rather than top of atmosphere (TOA) reflectance data. Surface reflectance corrects for
atmospheric effects such as scattering and absorption by gasses and aerosols, facilitating
more consistent comparisons over time and across different locations.

© RAIL SAS 2024 Page 12/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

We leverage advanced image pre-processing, to filter out cloudy imagery (especially present
in the tropics) and thus ensure the best quality end results. This optimisation concerns solely
optical satellite data and is achieved with the integrated Sentinel-2 cloud mask.

More precisely, “the cloud mask [...] contains both opaque clouds and cirrus clouds with an
indicator specifying the cloud type (cirrus or opaque cloud). Three spectral bands are used to
process this mask: a blue band (B1 (443 nm) or B2 (490 nm)), the band B10 (1375 nm) and a
SWIR band (B11 (1610 nm) or B12 (2190 nm)).”, its overall accuracy is around 90% (Baetens
et al., 2019). We process Landsat-8 data in a similar fashion to obtain cloudless images and
mosaics.

SAR pre-processing
Radar imagery is essential for our time series analysis due to its ability to penetrate clouds
and its correlation to forest structure characteristics.

A SAR Ground Range Detected (GRD) pipeline is used to pre-process Sentinel-1 and PALSAR
data. The pipeline uses two types of input data. For Sentinel-1, it uses sigma and gamma
PolSAR indicators (HV and VV polarisations), for Alos PALSAR, it uses the intensity (HV and
HH polarisations). The pipeline is divided in 2 different steps: data preprocessing and
indicators computation.

Data pre-processing
For Sentinel-1 data, a time series wrapping is achieved and a multi-temporal filtering. The
time series wrapping is necessary to apply multi-temporal treatment. It allows the application
of the same grid projection for the all multi-temporal dataset. We decided to project the data
on the grid of the first acquired data of each studied time series. The multi-temporal filtering
consists of a multichannel linear intensity filter, which is optimal for uncorrelated channels.
For Alos PALSAR data, data pre-processing includes computation of sigma PolSAR indicator
in dB.

Indicator computation
At this stage, we use the PolSAR indicators sigma and gamma, expressed in decibels, for the
given polarizations to calculate the vegetation indices. Two vegetation indexes are used: the
ratio VH/VV or HV/HH and the RVI. Table 3 represents the different output data produced by
the SAR GRD pipeline.

© RAIL SAS 2024 Page 13/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Table 3: Different output data from SAR pre-processing

Indicator Sentinel 1 Alos Palsar Range

PolSAR indicators
- sigma HH (dB) x [-30,0]
- sigma HV (dB) x x [-30,0]
- sigma W (dB) x [-30,0]
- gamma HH (dB) [-30,0]
- gamma HV (dB) x [-30,0]
- gamma W (dB) x [-30,0]
Vegetation Index
- ratio (dB) x x [0,5]
- WI (dB) x x [0,5]

Inference
The inference is the core process of Kanop methodology to provide the primary key
performance indicators. It uses the several optimized data inputs described above to directly
produce tree height, canopy height, canopy cover and living aboveground biomass estimates.

The model is a deep neural network model with a global reach, trained to predict several
indicators concurrently and consistently over time. One of the key features of the model is its
ability to predict all indicators jointly, ensuring a higher level of accuracy for each individual
indicator. Moreover, each predicted pixel is based on an analysis of over 30 data points, the
neighboring pixels as well as local space-borne LiDAR data, resulting in comprehensive and
precise predictions. Multi-output CNN models have been shown to improve the global
performance of the model, reducing the RMSE (root mean square error) by ~ 13% (Mishra
and Passos, 2022).

The model has been trained on a vast dataset encompassing 60 million hectares of forests
with airborne LiDAR data, with a significant portion trained from high resolution airborne
LiDAR scans covering (on 10 million hectares).

Our data coverage extends from 2013 to the present, offering a resolution of 25 meters
without reliance on Sentinel satellite data. Since 2017, we've enhanced our offerings to
include data at an even finer resolution of 10 meters, ensuring our clients access the most
precise information possible.

By leveraging this cutting-edge methodology, Kanop is able to deliver unparalleled insights

into forest dynamics, enabling informed decision-making and resource management for our
clients on a global scale.

© RAIL SAS 2024 Page 14/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Post-calibration
The Standard data product's accuracy can be enhanced by recalibrating it with your own field
data, for example, to meet specific standards’ requirements such as those outlined in Verra's
VT0005 or VM0047. For more information on this process, please refer to the details
provided here.

Gaussian process with field measurements

We are capable of recalibrating our estimates based on (sampling plots) provided by the
user. Each sampling plot is combined with its corresponding estimate to create a
transformation function applicable to the entire AOI. This strategy draws inspiration from
Gaussian Processes, a powerful Bayesian machine learning technique known for its effective
method of placing a prior distribution over functions (Rasmussen and Williams, 2006).

Post-calibration improves the match between our estimates and the field measurements at
the year of the field measurements, but this is also useful for the inference of the other years
(i.e. years without field measurements) where the transformation function still makes sense
to be applied.

It is also important to mention that the Gaussian process also affects the confidence interval,
logically reducing its value when fitted with field measurements.

Height corrections
A well known issue in predicting forest structural indicators is found in the prediction of very
low values, particularly regarding tree height and canopy height indicators. This is mainly due
to data and technical limitations: models rely on data collected from field measurements that
might have a minimum detectable DBH / height, and spaceborne LiDAR might also struggle
to distinguish short vegetation (< 3 m).

We introduce thus specific correction to low height predictions:

- When the canopy cover is lower than 10%, we introduce a relative correction factor to
lower the height.
- When the aboveground biomass is lower than 50 tons DM / ha, we also introduce a
relative correction factor (relative in the sense of proportional to the distance to the
threshold value: if the aboveground biomass is estimated to be 0 tons DM / ha, the
tree height should logically set to 0m).

These threshold values (10% for canopy cover, 50 tons DM / ha) have been chosen and
validated against field measurements (see Figure 3 for experiments).

© RAIL SAS 2024 Page 15/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Figure 3: Effect of AGB and crown cover thresholds (also called canopy cover) on tree height
correction

Forest-cover
The forest cover map is a strict application of the forest cover definition, which can be
customized in the project configuration. The default forest definition used by the platform is
from the FAO which defines a forest as a plot with a minimum of 0.5 ha surface, a minimum
tree height of 5 m and a minimum crown cover of 10%. Pixels that meet their project's
corresponding definition will be counted as Forest cover in our programming. It will be
represented as such in the maps and data we provide.

Allometric relationships
Allometric relationships are used to derive primary key performance indicators (tree height,
canopy height, canopy cover, living aboveground biomass) into secondary key performance
indicators. This involves estimating below-ground biomass, calculating total biomass,
estimating the total carbon stock of this biomass and converting it into CO2 equivalent. The
allometric equations used by our models are recommended in Verra methodologies and
modules, notably in the VMD0001 Estimation of carbon stocks in the above- and
belowground biomass in live tree and non-tree pools (CP-AB), v1.2 and can be found in the
IPCC “Good practice guidance for land use, land-use change and forestry” (IPCC, 2003).

© RAIL SAS 2024 Page 16/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

The biome classification is automatically detected for each plot of the project so that we can
use the appropriate corresponding allometric relationships.

More specific allometric relationships can also be provided in the web application or via the
API by the user. Once entered into the platform, they will automatically be applied to the
entire project area.

Belowground biomass calculation

The belowground biomass of each pixel is calculated according to its aboveground biomass.
The biomass expansion factor (BEF) and Root-to-Shoot ratios (RTS) both vary in function of
the project biome.

Table 4 presents an extract of the belowground biomass allometric equations table that we
use.

© RAIL SAS 2024 Page 17/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Table 4: Extract of the allometric equations table for belowground biomass estimation

Biome Formula Source Comment Species

IPCC,
Any exp(-1.085 + 0.9256 * ln(AGB))
2003.
Tropical and subtropical moist exp(-1.0587 + 0.8836 * IPCC,
broadleaf forests ln(AGB)) 2003.
Tropical and subtropical dry exp(-1.0587 + 0.8836 * IPCC,
broadleaf forests ln(AGB)) 2003.
Tropical and subtropical exp(-1.0587 + 0.8836 * IPCC,
coniferous forests ln(AGB)) 2003.
Temperate broadleaf and exp(-1.0587 + 0.8836 * IPCC,
mixed forests ln(AGB) + 0.2840) 2003.
exp(-1.0587 + 0.8836 * IPCC,
Temperate coniferous forests
ln(AGB) + 0.2840) 2003.
exp(-1.0587 + 0.8836 * IPCC,
Boreal forests / taiga
ln(AGB) + 0.1874) 2003.
... ... ... ... ...
Tropical and subtropical moist IPCC,
0.24 * AGB NS
broadleaf forests 2003.
Tropical and subtropical moist IPCC, AGB <125
0.42 * AGB
broadleaf forests 2003. t/ha
Tropical and subtropical dry IPCC,
0.27 * AGB NS
broadleaf forests 2003.
Temperate broadleaf and IPCC, AGB
0.35 * AGB E517
mixed forests 2003. 50-150 t/ha
Temperate broadleaf and IPCC, AGB <50
0.45 * AGB E517
mixed forests 2003. t/ha
Temperate broadleaf and IPCC, AGB >150
0.2 * AGB E517
mixed forests 2003. t/ha
Temperate broadleaf and IPCC, AGB
0.26 * AGB
mixed forests 2003. 75-150 t/ha
Temperate broadleaf and IPCC, AGB <75
0.43 * AGB
mixed forests 2003. t/ha
Temperate broadleaf and IPCC, AGB >150
0.24 * AGB
mixed forests 2003. t/ha
... ... ... ... ...
Tropical and subtropical moist Olivier, G.,
exp(-1.0587+0.8836*ln(AGB))
broadleaf forests 2017.
exp(-1.0587+0.8836*ln(AGB)+ Olivier, G.,
Boreal forests / taiga
0.1874) 2017.
Temperate broadleaf and exp(-1.0587+0.8836*ln(AGB)+ Olivier, G.,
mixed forests 0.2840) 2017.

© RAIL SAS 2024 Page 18/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Total biomass calculation

Total biomass is the sum of the aboveground biomass and the belowground biomass.

𝑇𝑜𝑡𝑎𝑙 𝑏𝑖𝑜𝑚𝑎𝑠𝑠 = 𝑎𝑏𝑜𝑣𝑒𝑔𝑟𝑜𝑢𝑛𝑑 𝑏𝑖𝑜𝑚𝑎𝑠𝑠 + 𝑏𝑒𝑙𝑜𝑤𝑔𝑟𝑜𝑢𝑛𝑑 𝑏𝑖𝑜𝑚𝑎𝑠𝑠

Carbon stock calculation

The stock of carbon contained in each pixel is calculated from the average carbon content of
the biomass. This carbon content depends on the biome (we use a similar table as for
belowground biomass estimation, cf. Table 4), and the allometric relationship can also be
configured by the user.

CO2 equivalent calculation

The amount of CO2 equivalent in each pixel is calculated according to its carbon stock.

𝑀(𝐶𝑂2)
𝐶𝑂2 𝑒𝑞 𝑠𝑡𝑜𝑐𝑘 = 𝑀(𝐶)
× 𝐶 𝑠𝑡𝑜𝑐𝑘

M being the molar mass of the corresponding element.

Biodiversity indicators
Rao’s Q diversity index
Rao’s Q diversity index is a biodiversity index to measure the variability among living
organisms in forest ecosystems. Rao’s Q diversity index measures variation of the spectral
data of a pixel and compares it with its neighboring pixels (the closer the pixel the more
weight it carries in our model's calculations). This means that it is a dimensionless value, and
it is relative to its proximal pixels. Rao’s Q is best suited for identifying localized areas of high
diversity within a project area.

It appears to be one of the most adapted biodiversity index to the Spectral Variability
Hypothesis (SVH) which essentially states that "the more variable the spectral data, the more
likely we are to have high biodiversity”.

Rao’s Q diversity index is a strict application of Rao (1982). As stated by Rocchini et al.
(2017) “Given an image of N pixels, the Rao’s Q is related to the sum of all the pixel values
pairwise distances, each of which is multiplied by the relative abundance of each pair of
pixels in the analyzed image. Rao’s Q is the expected difference in reflectance values
between two pixels drawn randomly with replacement from the considered evaluated pixels
set.” (Rocchini et al., 2017; Torresani et al., 2019).

Confidence interval
For each output of our DNN models (i.e. for tree height, canopy height, above-ground
biomass and, indirectly, forest cover), we estimate the outputs ten times by performing a

© RAIL SAS 2024 Page 19/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

'Test Time Augmentation'. Test Time Augmentation (TTA) allows the results to vary slightly
through small variations in the selection of images (which are seeded, i.e. randomized but
still reproducible). Instead of having a single direct output for each model, we have one main
estimate and ten other estimates for each key variable. This allows us to estimate the
standard deviation at the pixel level:

This standard deviation is then rescaled to fit our real uncertainty. This rescaling operation is
mainly based on the results of our accuracy assessment.

Finally, each key variable is calculated for each rescaled TTA prediction in order to have
distribution estimates for belowground biomass, total living biomass, carbon stock and CO2
equivalent stock.

The confidence interval is calculated to fit a 95% confidence level based on the following
relationship:
𝐶𝐼 = μ ± 1. 96 × σ
As evoked before, CI will be affected by Gaussian process post-treatment to logically lower it
when field data has been measured.

© RAIL SAS 2024 Page 20/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

V. Validation
Kanop has carried out two main validation processes to assess the accuracy of its model.
These two reports provide comprehensive methodology and results used to estimate the
accuracy of our model.

The first report relates accuracy results at the global scale, comparing our estimates on 300
areas of interest of 1 km x 1 km with global, independent or dependent reference dataset
such as: GEDI L2A (tree height sparse and dependent dataset) ; GEDI L4A (aboveground
biomass sparse and dependent dataset) ; GEDI L4B (aboveground biomass global and
independent dataset) ; and the Global Forest Canopy Height dataset (tree height global and
independent dataset).

The second report relates accuracy results on three areas in the region of Paragominas, in
Brazil. The reference dataset is a published dataset that contains field measurements of the
aboveground biomass and its generalization to the three local areas with airborne LiDAR- and
PALSAR-derived AGB (KELLER et al., 2019).

Hereafter, we provide a summary for the comparison of our estimates with GEDI L2A, Global
Forest Canopy Height, GEDI L4B and the airborne LiDAR- and PALSAR-derived AGB reference
datasets.

© RAIL SAS 2024 Page 21/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

GEDI L2A comparison

Figure 4: Comparison between GEDI L2A data and Kanop estimates for the studied AOIs (in
meters, black line represents the 1:1 line)

The GEDI L2A dataset has been presented in the Data Sources section. We compared our
estimates with the average of the sparse GEDI L2A data points within each of the 300 areas
of interest. The GEDI L2A data used for validation purposes has not been used in the training
of our models.

We observed a Mean Absolute Error (MAE) of 2.9 m, or a standard deviation (SD) of errors of
± 13.6 % when compared to the GEDI L2A reference dataset.

This represents slightly lower accuracy than other tree height estimation models such as
Staben et al. (2018) or Li et al. (2020) (RMSE = 3.8 m for Kanop, 2.8 m for Staben et al., 2.64
m for Li et al.). Our accuracy is also close to the accuracy obtained with LiDAR-derived
canopy height model (CHM), which highlight the high performance of our height model
(RMSE = 1.42 m with optimized LiDAR-derived CHM, which requires direct measurements
with aerial LiDAR, Mielcarek et al., 2018). This is important to note given that our model
represents a global model unlike the model of Staben et al., Li et al. or LiDAR-derived models.

© RAIL SAS 2024 Page 22/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Global Forest Canopy Height comparison

The Global Forest Canopy Height dataset represents global tree heights based on a fusion of
spaceborne-lidar data (2005) from the Geoscience Laser Altimeter System (GLAS) and
ancillary geospatial data (Simard et al., 2011). We observed a MAE of around 4 m. The MAE
ranges from 2.3 m for boreal forests and taigas, to 5.8 m for mediterranean forests,
woodlands, and scrub or sclerophyll forests, which is probably among the biomes where tree
height varies the most.

GEDI L4B comparison

The GEDI L4B dataset is a published dataset: “GEDI L4B Gridded Aboveground Biomass
Density, Version 2.1” (published at the end of October 2023). This dataset contains reference
data that has been produced by NASA and recently updated for the years 2019 to 2023
(Dubayah et al., 2023). It consists of a compilation of GEDI L4A data to ensure better
consistency. The GEDI L4A data used for validation has not been used in the training of our
models.

Figure 5: Comparison between GEDI L4B data and Kanop estimates for the studied AOIs (in
tons DM / ha, black line represents the 1:1 line)

We observed a Mean Absolute Error (MAE) of 15.9 tons DM / ha or a standard error of ± 14.1
%. We observed a global trend where Kanop estimates seem to be lower than GEDI L4B data.
MAE varies slightly between different biomes. From a MAE of 11.1 tons DM / ha for
temperate broadleaf and mixed forests to a MAE of 20.8 tons DM / ha for tropical and
subtropical coniferous forests.

© RAIL SAS 2024 Page 23/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

Airborne LiDAR- PALSAR-derived AGB comparison

We compared our estimates with airborne LiDAR- and PALSAR-derived AGB from three study
areas in Brazil.

Figure 6: From left to right, Kanop estimates, Kanop resampled estimates with the first
quartile method, and the reference dataset for the “And” study area

We observed a MAE of 25.6 tons DM / ha for the “And” study area which has an average AGB
of 81.9 tons DM / ha. For low aboveground biomass regimes (lower than 30 tons DM / ha),
the MAE drops down to around 5 tons DM / ha.

The MAE for the “Cau” study area is lower than for the “And” study area, it is 19.9 tons DM /
ha. “Cau” study area has an average AGB of 181.1 tons DM / ha.

© RAIL SAS 2024 Page 24/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

VI. References
Baetens, L., Desjardins, C., Hagolle, O., 2019. Validation of Copernicus Sentinel-2 Cloud
Masks Obtained from MAJA, Sen2Cor, and FMask Processors Using Reference Cloud
Masks Generated with a Supervised Active Learning Procedure. Remote Sens. 11,
433. [Link]
Canadell, J.G., Raupach, M.R., 2008. Managing Forests for Climate Change Mitigation.
Science 320, 1456–1457. [Link]
Dubayah, R.O., Armston, J., Healey, S.P., Yang, Z., Patterson, P.L., Saarela, S., Stahl, G.,
Duncanson, L., Kellner, J.R., Bruening, J., Pascual, A., 2023. Global Ecosystem
Dynamics Investigation (GEDI)GEDI L4B Gridded Aboveground Biomass Density,
Version 2.1 0 MB. [Link]
Harris, I., Osborn, T.J., Jones, P., Lister, D., 2020. Version 4 of the CRU TS monthly
high-resolution gridded multivariate climate dataset. Sci. Data 7, 109.
[Link]
IPCC, 2003. Good practice guidance for land use, land-use change and forestry /The
Intergovernmental Panel on Climate Change. Ed. by Jim Penman. Hayama,
Kanagawa.
Keenan, R.J., 2015. Climate change impacts and adaptation in forest management: a review.
Ann. For. Sci. 72, 145–167. [Link]
KELLER, M.M., DUFFY, P., BARNETT, W., 2019. LiDAR and PALSAR-Derived Forest
Aboveground Biomass, Paragominas, Para, Brazil, 2012 23.291921 MB.
[Link]
Li, W., Niu, Z., Shang, R., Qin, Y., Wang, L., Chen, H., 2020. High-resolution mapping of forest
canopy height using machine learning by coupling ICESat-2 LiDAR with Sentinel-1,
Sentinel-2 and Landsat-8 data. Int. J. Appl. Earth Obs. Geoinformation 92, 102163.
[Link]
Liesenberg, V., De Souza Filho, C.R., Gloaguen, R., 2016. Evaluating Moisture and Geometry
Effects on L-Band SAR Classification Performance Over a Tropical Rain Forest
Environment. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9, 5357–5368.
[Link]
Mielcarek, M., Stereńczak, K., Khosravipour, A., 2018. Testing and evaluating different
LiDAR-derived canopy height model generation methods for tree height estimation.
Int. J. Appl. Earth Obs. Geoinformation 71, 132–143.
[Link]
Mishra, P., Passos, D., 2022. Multi-output 1-dimensional convolutional neural networks for
simultaneous prediction of different traits of fruit based on near-infrared
spectroscopy. Postharvest Biol. Technol. 183, 111741.
[Link]
Mondal, S., Palit, D., 2022. Challenges in natural resource management for ecological
sustainability, in: Natural Resources Conservation and Advances for Sustainability.
Elsevier, pp. 29–59. [Link]
Rao, C.R., 1982. Diversity and dissimilarity coefficients: A unified approach. Theor. Popul.
Biol. 21, 24–43. [Link]
Rasmussen, C.E., Williams, C.K.I., 2006. Gaussian processes for machine learning, Adaptive
computation and machine learning. MIT Press, Cambridge, Mass.
Rocchini, D., Marcantonio, M., Ricotta, C., 2017. Measuring Rao’s Q diversity index from
remote sensing: An open source solution. Ecol. Indic. 72, 234–238.
[Link]
Simard, M., Pinto, N., Fisher, J.B., Baccini, A., 2011. Mapping forest canopy height globally
with spaceborne lidar. J. Geophys. Res. 116, G04021.

© RAIL SAS 2024 Page 25/26

Kanop - Data Products Methodology - Version 2.0.2 - March 2024

[Link]
Staben, G., Lucieer, A., Scarth, P., 2018. Modelling LiDAR derived tree canopy height from
Landsat TM, ETM+ and OLI satellite imagery—A machine learning approach. Int. J.
Appl. Earth Obs. Geoinformation 73, 666–681.
[Link]
Torresani, M., Rocchini, D., Sonnenschein, R., Zebisch, M., Marcantonio, M., Ricotta, C., Tonon,
G., 2019. Estimating tree species diversity from space in an alpine conifer forest: The
Rao’s Q diversity index meets the spectral variation hypothesis. Ecol. Inform. 52,
26–34. [Link]

© RAIL SAS 2024 Page 26/26