Overview
This SOP describes the methods used for calculating changes in forest carbon storage. This document provides information about the goal of carbon calculations and step by step directions for completing the carbon calculations. Additional information needed for these methods include the LandTrendr outputs.
The goal of this protocol is quantify increases, maintenance, or decreases in carbon stored as biomass over time for the ABF deals. Changes in carbon stored in vegetation are expected in TerraBio due to reforestation activities and growth of trees (agroforestry and/or natural forests). Note, the changes that we capture in carbon storage over time is from changes in tree canopy cover. A different set of methods will be needed to monitor changes in carbon over time with changes to other land covers, such as grasses and shrubs.
The basic idea is to use remote sensing data to estimate forest biomass (through stand age and growth curves) and then convert biomass to carbon storage.
Approach
Methodology: Changes in carbon stored in above and below-ground biomass are estimated using four steps:
- Map forest cover and stand age at annual increments.
- LandTrendr: The Kennedy et al. (2010) paper describes LandTrendr, an algorithm that detects changes (decrease or increase) in forest canopy and can be used to date forests that have regrown
- Generate a probabilistic sample (e.g., simple random or stratified random). At each sample, we will interpret stand age by looking at time series plots and associated remote sensing imagery (e.g., Landsat).
- Oloffson et al. (2014): map and sample based approaches to estimate area of stand ages. We conduct image interpretation in CEO Saah et al. 2019 to correct for systematic map errors of stand age.
- Apply growth curve functions to forest stands using growth curve functions developed for our region of interest
- Bernal et al. (2018): The Bernal et al. (2018) paper provides biomass accumulation curves for different areas of the world. In particular, “…we developed biomass accumulation rates for a set of FLR activities (natural regeneration, planted forests and woodlots, agroforestry, and mangrove restoration) across the globe and global CO2 removal rates with corresponding confidence intervals, grouped by FLR activity and region/climate”.
- Assess the uncertainty of the estimates using a sample based approach and propagation of confidence intervals.
Input Data:
- Landsat time series (NBR),
- Potential forest layer (from supervised classification of NICFI and Map Biomass) Computing Environment:
- GEE: to map stand age,
- CEO: image interpretation
- R or GEE: generate statistics Output: Carbon sequestration per year per land use
Instructions
Summary notes on discussion regarding suggested improvements - 10/4/2023
** these need updating **
1. Map forest cover and stand age at annual increments.
We use LandTrendr, implemented in GEE, to map stand age. The Kennedy et al. (2010) paper describes LandTrendr, an algorithm that detects changes (decrease or increase) in forest canopy. This is the foundation that we use to date when forests or trees have started to regrow. We used NBR OR NDFI note to specify which one we use.
Setting the Time frame
We will run the analysis going back as far in time as possible, since we want to be able to assign a stand age to all stands – even those that are old. So our mapping time frame goes back to 1984, when Landsat images were first collected. We use Landsat 5-8 Surface Reflectance for 1985-2021.
Some trees, or stands of trees, were established prior to the collection of Landsat data in 1984. In these cases, we need to assign an age of establishment In these cases, we have set the age of these stands to 37.5 year because in the Brazilian Amazon ** this is the reasons**.
Potential forest –
Making this operational
Often, multiple runs might be needed in order to fully customize the LandTrendr algorithm to the local area, and recording these tuned parameters in the data dictionary.
Postprocessing
- Postprocess the LandTrendr outputs to identify areas of growth/loss. Filter the vertex information to identify the greatest change segments and filter for relevant years (e.g. 1990-2021) and other variables (e.g. NBR index greater than 300).
- Classify regrowth age from areas of identified Landtrendr forest gain for each time step.
- Classify loss events, secondary forest, and secondary forest stand age as locations where current land cover is forest and a loss event occurred.
- Year of loss from LandTrendr (and/or year of loss from Hansen global forest loss), then compile to determine the most recent year of loss.
- Determine loss area age by subtracting year of loss from current year.
- In forests, identify locations that are currently forest but a loss event has occurred in the past. Classify the stand age of these locations as the age since lost.
- We repeat this process for each time step (e.g., we are interested in calculating carbon over a 5 year baseline).
- Classify secondary forest and secondary forest stand age as locations where current land cover is forest and a gain event occurred.
- Take the stand age determined, the forest mask, and annual LandTrendr regrowth age. In forests, identify locations that are currently non forest, but a gain event occurred in the past.
- The stand age of these locations is equal to the regrowth age. Repeat for each time step.
- Classify mature forest and mature forest stand age as locations where current land over is forest and no disturbance events have occurred.
- In forests, identify locations that are currently forest and have an unknown stand age. You may need to make assumptions about stand age (e.g. assume forests that have not been disturbed are 40 years old at the start date).
- Classify the stand age of these locations accordingly.
- Compile the information from steps 5, 6, and 7 sequentially into a final stand age of forest layer.
2. Image Intepretation of Sample
We correct for systematic bias in our stand age map using sample based approaches (Oloffson et al. (2014)). Samples are interpreted in Collect Earth Online (CEO, Saah et al. 2019).
2.1 Sample Design
All maps have systematic errors, and with pixel-counting alone we are likely either overestimating or underestimating stand age and the associated carbon removals. We apply a sample based approach on top of the map in order to characterize and correct for systematic map errors.
We place a random sample of plots across the study site. Note, we do not exclude places that are mapped as non-tree cover, since these could be errors as well as lands mapped as having tree cover (with associated stand ages).
We will upload this probabilistic sample into a project in Collect Earth Online (CEO).
2.2. Image interpretation
Once the sample has been generated, we create a project in CEO to do image interpretation of the forest dynamics at each sample plot using time series information and imagery. The questions we answer in CEO include questions about forest loss/gain events (and year of the event) along with the land cover in the final year (e.g. 2021).
Once the project is set up, interpreters work with high resolution imagery and time series graphs of Landsat images to answer these forest dynamic question at each sample.
2.3 Post-processing of sample answers
For young forests, where interpreters have indicated there were loss or regrowth activities, we need to convert the CEO questions into an establishment age. We use the following approach to determine the establishment age. After forest loss, if there was a subsequent indication of regrowth (or the interpreter indicated that the land cover in the final year was forest) we set the stand age at 0 when the loss event occured. Then we increase the stand age in annual increments. Since forest regrowth is only detectable after trees begin maturing, we cannot count the year of regrowth as stand age of 1 (it is likely older than 1 year). Therefore we begin aging the stand immediately after loss, starting at age 1. If the earliest recorded forest change event is gain (e.g. 1990), we assume it started growing four years earlier (in 1986) (e.g., we can see trees approximately 4 years old using image interpretation). If no land change event was detected, that sample is a stable forest or non-forest based on CEO answer for final year land cover (e.g. 2021). For stable forest samples, we set the stand age to be increasing from 37.5 years starting from 2017 if stand age in the initial years cannot be determined through previous logic.
3. Apply growth curve functions
We apply growth curve functions developed for our region of interest based on published studies Bernal et al. (2018). The biomass accumulation curves were developed for a set of FLR activities (natural regeneration, planted forests and woodlots, agroforestry, and mangrove restoration) with corresponding confidence intervals, grouped by FLR activity and region/climate.
- Identify the correct growth curve function(s) for use in the study area.
- Calculate carbon in tonnes per hectare based on the stand age using the growth curve function(s).
- Use the Richard Chapman function: 𝛽0(1-EXP(-𝛽1Age))^𝛽2
- This creates a smooth change of carbon stock with increasing age of forest, including for young forests. Note that the confidence intervals change with age, and the curves differ depending on forest type.
4. Estimate carbon removals or emissions and assess the uncertainty of the estimates
Calculate unbiased estimates and uncertainties using the carbon values resulting from the interpreted samples and their associated sampling weights. Plot a time series of the total carbon estimates and uncertainties for the study farms. Repeat for each land use type. This is currently done in R.