LiDAR Based Quantification of Aboveground Biomass

1.0 Introduction

Several terrestrial biosphere Carbon Dioxide Removal (CDR) approaches rely on the capture and storage of carbon in living woody biomass -- both above-groundaboveground biomass (AGB) and below-groundbelowground biomass (BGB). The quantification of BGB is often dependent upon AGB¹, which is more directly observable. Thus, for these pathways, accurately and conservatively quantifying the gross storage of carbon in AGB is crucial for demonstrating net CO₂ removal.

Above-groundAboveground biomass (AGB) encompasses all living vegetation above the soil surface, including stems, branches, foliage, and bark. Woody biomass refers to plants whose structure includes lignified stems, such as bamboo, plants, shrubs, and trees.

This Module outlines requirements for quantification of AGB over a project area using Airborne Laser Scanning (ALS) measurements, also known as an area-based approach. This Module includes requirements for ALS data acquisition, ALS data processing, modeling, and verification.

ALS data can also be used to develop other types of models for forest carbon, such as through Individual Tree Crown modeling. Individual Tree Crown modeling is not in scope for this current Protocol version, but guidelines may be included in future iterations.

1.1 Background on ALS

An area-based approach for estimating [math: M_{AGB}] involves linking field measurements of forest inventory with ALS data to enable area-wide mapping of forest attributes.

Generally, the following steps are used:

1. Collection of ALS data, such as LiDAR from the air using drones, helicopters, or airplanes.
2. Processing of ALS data, including the transformation of point- clouds to structurestructured data.
3. Developing predictive models whichthat link ALS data to co-located ground measurements.
4. Using the predictive model to map forest attributes, including AGB, across a project area.

A number of decisions are made throughout the process. The requirements and recommendations for each step are detailed in subsequent sections.

The following resources are suggested for guidance on best practices:

• White, 2013: A best practices guide for generating forest inventory attributes from airborne laser scanning data using an area-based approach (Version 2.0)²
• Næsset, 2013: Area-Based Inventory in Norway–From Innovation to an Operational Reality³
• White et al., 2017: A model development and application guide for generating an enhanced forest inventory using airborne laser scanning data and an area-based approach⁴

2.0 Future Versions

This Module was developed based on the current state of the art, publicly available science regarding quantification of woody above-groundaboveground biomass and long-term monitoring of terrestrial ecosystems. This Module aims to be scientifically stringent and robust. We recognize that some requirements may exceed the status quo in the market and that there will be opportunities to improve the rigor of this Module.

Additionally, this Module will be reviewed when there is an update to published scientific literature, government policies, or legal requirements whichthat would affect net CO₂e removal quantification or the monitoring guidelines outlined in this Module, or at a minimum of every 2 years.

3.0 Applicability

This Module quantifies the mass of above-groundaboveground woody biomass across the project area ([math: M_{AGB}]) at a given time point, t, through direct vegetation sampling and parameter estimation within survey plots and scientifically validated allometric equations.

• The project area is ≥ 1 hectare;
• Reliable field measurement data can be consistently collected, reported, and verified; and
• Project activities involve direct planting, assisted natural regeneration, improved forest management, or a combination thereof.

Throughout this Module, the use of “must” indicates a requirement, whereas “should” indicates a recommendation.

4.0 Calculation of [math: MM_{AGB}][math: AGB] using ALS Data

This Module allows for two different area-based approaches for calculating [math: M_{AGB}] (tonnes) using ALS data. The recommended approach is to use wall-to-wall ALS measurements that have full coverage over the project area. In the other approach, ALS data collection can occur over representative subplots over a portion of the project area.

In the wall-to-wall approach, [math: M_{AGB}] forof the project area is calculated by summing all the values of biomass density derived from the ALS observations and modeling across the entire project area:

[math: M_{AGB}(t) = {\sum_{i=1}^{n} AGB_{pixel,i}(t) \times a_{pixel}}]

(Equation 1)

Where:

• [math: AGB_{pixel,i}(t)] is the AGB density (tonnes/hectare), in pixel [math: i].
• [math: a_{pixel}] is the pixel area, in hectares.
• [math: n] is the number of pixels.

In the alternative approach, ALS data is collected over representative subplots, and then used to estimate the mean biomass density ([math: \bar{m}_{AGB}]) over the project area.

[math: \bar{m}_{AGB}(t) = \frac{\sum_{i=1}^{j} AGB_{pixel,i}(t) \times a_{pixel}}{\sum_{i=1}^{j} a_{pixel}}]

(Equation 2)

Where:

• [math: \bar{m}_{AGB}(t)] is average AGB density, in tonnes/hectare.
• [math: AGB_{pixel,i}(t)] is the AGB density (tonnes/hectare), in pixel [math: i].
• [math: a_{pixel}] is the pixel area, in hectares.
• [math: j] is the number of pixels measured across the subplots.

[math: M_{AGB}] can then be calculated using the estimate of the mean AGB density, which can then be multiplied by the project area to obtain a total AGB:

[math: M_{AGB}(t) = \bar{m}_{AGB}(t) \times A_{project} ]

(Equation 3)

Where:

• [math: A_{project}] is the total project area, in hectares.

If the project area is stratified, Equations 2 and 3 can be repeated for each project sub-area.

5.0 ALS Data Acquisition

Wall-to-wall laser scanning measurements are highly recommended. Laser scanning of representative subplots throughout the project area is also permissible with rigorous statistical analysis to constrain the uncertainty associated with subplot selection and upscaling.

Minimum requirements for data acquisition parameter values are detailed below and are based on guidance from White (2013)² and Duncanson et al. (2021)⁵. In some settings (e.g., areas with dense canopies), these parameter values may need to be more stringent than these minimum requirements in order to meet the digital elevation model resolution requirement. 

Minimum parameter requirements include:

6.0 ALS Data Processing

Project Proponents are strongly encouraged to use standardized tools and published QA/QC protocols to maximize ease of replicability and transparency, several such protocols are referenced in Duncanson et al. (2021)⁵.

7.0 Model Requirements

This section describes the requirements for developing ALS-based predictive models to link ALS structure data to AGB. Previously developed models will be considered acceptable for use when documentation is provided on the development of the model that demonstrates these procedures were followed.

7.1 Within-project Sampling Design

Overlapping ground plots with ALS plots are required to develop ALS-based predictive models.

Plot Selection:

• It is recommended to use Structurally Guided Sampling to ensure that plots representative of the study area are included for training. In Structurally Guided Sampling, ALS metrics (e.g., height) are used as the source of data for stratification. See Goodbody et al., 2023⁷ for more details on Structurally Guided Sampling.

Plot Size:

Plot Measurements:

• Extra care must be taken when conducting ground plot measurements in steep and/or complex terrain, thick canopies or areas with a dense understory.

7.2 Statistical Modeling

Predictive ALS-based AGB maps have been developed with a variety of statistical models. The selection of the best modeling approach will be site and project -specific. The following types of models are acceptable for use:

• Parametric modeling techniques, e.g., linear regression or variants such as linear mixed-effects models and non-linear mixed effects models. These models have many desirable properties when combined with a design-based sample of forest inventory plots. However, these models are restricted by assumptions (e.g., normality) and usually work best in homogeneous forest ecosystems (e.g., pine plantations).
• Regression trees, e.g., Random Forests or Gradient Boosting. These are the most common techniques used to construct AGB maps. Algorithms like Random Forests are robust to features displaying multicollinearity (most of the features computed using packages like lidrmetrics or FUSION are highly correlated) and can identify high-dimensional interactions among the predictor features. However, Random Forests are not capable of extrapolating beyond the bounds of the dataset, as the algorithms work via a recursive partitioning of the observed feature space during training. As such, when using tree-based models, it is important to ensure the plot data captures the full variability of the landscape. This again highlights the benefit of using a stratified or structurally guided sampling approach⁹.
• Deep learning, e.g., Neural Networks or Convolutional Neural Networks¹⁰. Deep learning methods are highly data dependent and require a very large number of plot acquisitions. As such, it is common to pre-train these models on large datasets and then “tune” them on a local plot dataset.

It is recommended to use the simplest possible statistical model that yields high-performance. White et al. (2017)⁴, in particular Tables 4, 5 and 6 therein can be used to support key decisions and considerations in selecting a modeling approach.

Regardless of the particular type of model selected from the above options, models used to estimate AGB must:

7.3 Model Validation

7.4 Appropriate Model Use for Map Generation

Once the model has been developed, it can be applied to ALS data to generate an AGB map. The following requirements ensure appropriate model use:

7.5 Additional Modeling Considerations

Non-parametric models commonly exhibit a phenomenaphenomenon called “regression-to-the-mean”. Machine learning models work by optimizing a loss function with the goal of obtaining the lowest possible error rate. Often, the lowest possible error will be obtained by over-estimating areas with low AGB densities and under-estimating areas with large AGB densities. This can be observed in the one-to-one regression plots of many papers (e.g., see Figure 3, Panel 1 in Pflugmacher et al., 2014)¹⁴.

When a model exhibits regression to the mean, it is not appropriate to simply sum the pixel values within a given area as the systematic error of the model has not been accounted for. This is a considerable problem for regrowth modeling. Non-parametric models that use satellite imagery will almost always dramatically over-estimate low-biomass densities. Similarly, there is good evidence that the “green-up” of stands to pre-disturbance levels in moderate resolution vegetation indices will predate the actual structural regeneration of the stand by many years¹⁵.

Model-assisted estimation is a framework that allows ancillary information to be incorporated into the estimation procedure. This can allow for more precise estimates of AGB density. The typical setup for a model-assisted estimation procedure is as follows:

1. A design-based sample of the landscape has been acquired.
2. A model (parametric or non-parametric) is then fit using ancillary datasets (e.g., multispectral features, LiDAR features, land cover information, etc.) and spatially applied to all pixels within the sampling domain.
3. A model -assisted estimator is then used that integrates the mapped predictions and the plots to produce a final estimate.

• It is worth noting that a model-assisted estimate of AGB will not be worse than a non-model -assisted estimator. If the model that links the ancillary variables with the plot-level data is not performant, the model-assisted estimator will produce the same result as a non-model -assisted estimator.

Details for model -assisted estimation, and an associated R package, can be found in McConville et al. (2020)¹⁶.

There are many geostatistical considerations that come into play when developing statistical models. Spatial autocorrelation can violate the assumption of independence. This is why clustered plot designs are less efficient than other sampling schemes; the clustered plots are effectively pseudo-replicates. Under a design-based model-assisted framework, this is not a concern. However, it is a concern if the pixels in a map are being summed as all pixels within the same forest patch are pseudo -replicates. Therefore, the sampling size is falsely inflated, given an optimistic estimate of the variance of the estimated mean.

8.0 Benchmarking

Field measurements must follow the requirements set forth in the Area-based Quantification of Above-groundAboveground Biomass Module.

Uses field-based measurements of vegetation species and sizes taken within sample plots, along with allometric equations, to quantify biomass.

In addition to the considerations for field sampling included in the referenced Module, field plots for the purpose of validating an external map must adhere to the following:

9.0 Uncertainty

Models and measurements of above-ground biomass inherently include uncertainty from various data sources which contribute to the quantification.

Potential sources of uncertainty to consider include, but are not limited to:

• Spatially explicit uncertainty estimates of the AGB estimates associated with ALS data collection and model estimation (e.g., ALS measurement errors, model errors)
• Uncertainty associated with any local corrections applied to the AGB product as a result of benchmarking
• Errors associated with field measurements and estimation procedures used for benchmarking (e.g., DBH measurement errors, height estimation bias, wood density variability, plot representativeness)

10.0 Reporting and Data Sharing

Project Proponents must report the following information in the PDD, to ensure transparency and with enough detail to allow for repeatability:

• Data acquisition:

  [G-6HBR-0, If Projects use the subplot method for ALS measurements, Projects must provide details of the measured plots and provide justification of the sampling design, including stratification method and how autocorrelation is accounted for.]

  • ALS plots, including justification for the stratification method used and how spatial autocorrelation is accounted for, when the subplot method is selected

  [/G-6HBR-0]
  [R-N7QP-0, Projects must provide details of the instruments and calibrations used and their alignment with manufacturer's specifications.]

  • Instruments and calibrations used, which must adhere to the manufacturer's specifications

  [/R-N7QP-0]
  [R-1JW1-0, Projects must provide details on the spatial coverage of the ALS measurements.]

  • Spatial coverage of laser scanning

  [/R-1JW1-0]
  [R-8WKB-0, Projects must provide the flight altitude and velocity of the drone/plane used for ALS measurements.]

  • Flight altitude and velocity of drone/plane

  [/R-8WKB-0]
  • Acquisition parameters, e.g., beam divergence, scan angle, scan rate, pulse density, swath overlap, environmental conditions during acquisition, dates, and altitude

• Data processing:

  [G-WEHP-0, Projects must list any software used for data processing.]

  • Processing methods and documentation, including software and specific procedures followed
• [/G-WEHP-0]