Clark Labs Publishes New Focus Paper on Modeling REDD Baselines using IDRISI's Land Change Modeler

Clark Labs has recently published a Focus Paper on modeling REDD baselines with the Land Change Modeler application in their IDRISI GIS and Image Processing software. Reducing Emissions from Deforestation and Forest Degradation (REDD) is a climate change mitigation strategy for the protection and maintenance of forests that has gained momentum with conservation organizations, project developers and governments in developing countries. As forests play a major role in the sequestration of carbon, especially tropical forests, REDD offers great potential for reducing greenhouse gas emissions. Carbon offset payments, based on an assigned financial value for the likely carbon storage capacity of protected or maintained forests, provide incentives for adopting such a strategy.

A REDD project is comprised of a number of significant components and requires the expertise and input of multiple parties. This paper discusses one of the required elements, the modeling of a baseline, a scenario of the likely greenhouse gas emissions if the project were not to be implemented. The modeling relies on land cover and land use change maps as its foundation. Baseline mapping includes determining historical deforestation rates and patterns, identifying the unique drivers and underlying factors of the area’s deforestation, and modeling potential scenarios.

The new Focus Paper provides a general overview, featuring a case study of a REDD project developed by the government of Madagascar, with assistance from Conservation International and the BioCarbon Fund, for the area along the Ankeniheny-Mantadia Corridor. The overall objectives for this REDD project were to reduce greenhouse gas emissions, increase protected areas, and reduce the loss of biodiversity by reconnecting forest fragments and reducing future forest fragmentation.

Download the Modeling REDD Baselines using IDRISI's Land Change Modeler Focus Paper.

Modeling Deforestation in the Bolivian Lowlands Species and Carbon Impacts - Carbon Impact

Carbon Impact

Projecting future land cover scenarios is critical also for assessing the impact on carbon mass. The traditional method has been to use a uniform carbon figure across a region (Figure 8). In our case, we have a uniform carbon figure for all forested areas per hectare and we can simply calculate the loss of carbon mass due to projected deforestation. However, using a proxy of forest land cover, we can refine this process by redistributing the gross carbon mass associated with all forested areas based on the forest’s NDVI (Figure 9) so that the intensity of the NDVI will be our spatial proxy, i.e., the higher the NDVI, the more carbon. Using the projected land cover map to 2015, we identified only the deforested areas and redistributed the gross carbon mass using the NDVI (Figure 10). Assuming business as usual, the table shows the potential additional loss of carbon mass using a non-uniform carbon mass surface (Figure 11). Using modeling tools such as GIS, we can present various scenarios with differing potential impacts for designing carbon offset projects while simultaneously managing biodiversity.

Modeling Deforestation in the Bolivian Lowlands Species and Carbon Impacts - Model Development

Model Development

Change modeling in LCM is empirically based (the empirical method chosen for this study is the multi-layer perceptron neural network), using areas of known transition from land cover maps along with user-specified explanatory variable maps that express the state of driving forces or the inherent suitability of land to undergo a specific transition. In this illustration, a total of 26 variables were used (Figure 1). The result is a transition potential map for each modeled land cover transition -- an expression of the readiness of land to undergo a transition in the next time step. These transition potential maps then serve as primary inputs to the prediction process, along with predicted quantities of change derived by Markov Chain Analysis. The specific allocation of change is then achieved by means of a multi-objective land competition process with intermediate recalculation of dynamic variables (such as proximity to disturbance) and associated transition potential maps.

In our case, forest change was modeled based on land cover maps from 1992 to 2001 (Figure 2). To assess the quality of the model, a prediction was made to 2004 for comparison with a known validation map. The resulting model attained a Pierce Skill Score of .61. Two types of maps were produced: a hard prediction map that expresses one change scenario (Figure 3) and a soft prediction map (Figure 4) which provides a comprehensive assessment of change potential, i.e., a map of vulnerability to change for the selected set of transitions. The soft prediction result therefore shows the potential that any given area may change at some future date and is very important for identifying areas of high risk: an important planning concern.

Once the model was calibrated, additional future scenarios were then produced. One such scenario was a prediction of the Bolivian Lowlands to the year 2015 (Figure 3) and the assessment of this scenario on endemic species habitat. Using collections of species range polygons, initially supplied by NatureServe and subsequently recalibrated using LCM, a species richness map was developed for all birds, mammals, and amphibians. The Bolivian Lowlands is home to 73 endemic species, 36 amphibians, 16 mammals, and 21 birds. This map was combined with the 2015 soft prediction map to derive a risk of biodiversity loss (Figure 5). We then applied the same process to all the endemic vertebrate species to derive a risk of endemic loss (Figure 6).

Assuming business as usual, the potential loss could be devastating to biodiversity in Bolivia. For the period we do know, from 2001 to 2004, 4% of current mammal habitat was lost (Figure 7). Tools such as these provide an important means to assist those engaged in important conservation practices and to prioritize threatened environments.

Modeling Deforestation in the Bolivian Lowlands Species and Carbon Impacts - Introduction

Introduction

With the importance of tropical forests for climate system function and biodiversity, land cover change in the Amazon has been the focus of intensive research. In Bolivia, the rate of deforestation is second only to that of Brazil. Due in part to a tremendous amount of economic growth in recent decades, the Bolivian Lowlands have seen forest loss of almost 3 million hectares between the years of 1992 and 2004, with two-thirds of this loss coming between 2001 and 2004 alone. With forests dominating over 50% of the land cover in lowland Bolivia, special planning tools are essential to manage, monitor and prioritize the risks associated with economic development and to assist the conservation community better understand the changes and impacts that are taking place. This poster demonstrates a set of techniques to assess land cover change, predict future scenarios, and assess the potential impacts of these scenarios on biodiversity and carbon loss. These techniques will be demonstrated using the tool Land Change Modeler (LCM) as found in the IDRISI GIS and Image Processing system.

Comparison of Two GIS-based Land Change Modules for Constructing REDD Baselines in Bolivia

AAG - 2010 Annual Meeting, Washington, DC

Comparison of Two GIS-based Land Change Modules for Constructing REDD Baselines in Bolivia

is part of the Paper Session:
Land Use and Spatial Analysis

scheduled on Sunday, 4/18/10 at 8:00 AM.

Author(s):
Oh Seok Kim* - Department of Geography, University of Southern California

Abstract:
With the increasing concerns in developing methodologies for Reducing Emissions from Deforestation and Degradation (REDD) projects, there is a need to understand the characteristics of existing Land-Use/Cover Change (LUCC) modules. This research presents a comparison of two existing approaches: GEOMOD Modeling (GM) and Land Change Modeler (LCM). The comparison uses data from a case study in Chiquitanía, Bolivia. Data from 1986 and 1994 are used to simulate land-cover of 2000; the resulting maps are compared to an observed land-cover map of 2000. GM and LCM simulate baseline deforestation at the pixel-level. The model structures of linear extrapolation and Markov chain are compared to review quantity of LUCC; and the model structures of empirical frequency, logistic regression and multilayer perceptron are compared to review allocation of LUCC. Relative operating characteristics, figure of merit and multiple resolution analysis are employed to assess predictive accuracy of multiple transition modeling. By design, GM lacks the potential to model multiple transitions, while LCM may produce different results for each simulation. Based on the model structure and predictive accuracy comparisons, the LCM's logistic regression seems the most suitable LUCC module to construct a REDD baseline in this case. However, it is crucial to emphasize that these results are strictly limited to this particular case. Thus, there is no obvious method that is most accurate, so if a REDD project employs predictive GIS-based LUCC modeling for its spatially-explicit baseline construction, it should include the framework employed in this research to establish the baseline in a scientific manner.

Workshop at the AAG Annual Meeting

Modeling REDD-Baselines Using IDRISI's Land Change Modeler

Tuesday, April 13, 1:00pm – 4:00pm
http://www.aag.org/annualmeetings/2010/workshops.htm

Instructors:
Stefano Crema, Clark Labs, Clark University
Florencia Sangermano, Clark Labs, Clark University
Marc Steininger, Conservation International

Workshop Capacity: 30
Room: Lincoln Room 3

The implementation of REDD projects requires a best estimate of the deforestation and/or degradation baseline. The baseline is derived from the spatial modeling of the expected future trends of deforestation using maps of historical deforestation and biomass estimates.

Land Change Modeler (LCM) is a robust modeling tool that allows calibration, validation and the creation of maps of expected future deforestation trends. Clark Labs, in conjunction with Conservation International, has designed a REDD tool within LCM which guides the user through the steps of baseline development and directly produce tables and graphics necessary in the reporting process.

This workshop will cover the complexity of REDD baseline creation using real case studies to illustrate the process.

LCM on Google's Earth Engine interface.

Clark Labs Receives Support from Google.org to Develop an On-line Prototype of its Land Change Modeler Application to be run on Google’s Earth Engine

Worcester, MA -- Google has contracted Clark Labs to develop an on-line prototype of the Land Change Modeler application to run on Google’s Earth Engine interface. This development complements the on-line forest monitoring applications of Carnegie Institution for Science (CLASlite) and IMAZON (SAD), demonstrated on Earth Engine at the COP 15 meeting in Copenhagen last month.

The on-line prototype will contain much of the functionality of the land change modeling component of Clark Labs’ Land Change Modeler application, included within the IDRISI GIS and Image Processing software, as well as within its extension to ArcGIS. The system will allow users to analyze, model and predict deforestation and then to evaluate the amount of carbon involved.

The hope is that this on-line system will help support the implementation and evaluation of REDD (Reducing Emissions from Deforestation and Degradation) projects. The intent of REDD is to assign financial value to the carbon stored in forests such that sequestration of that carbon becomes a viable alternative to other uses such as conversion to pasture and farmland. The Land Change Modeler provides a critical ingredient for REDD projects in that it can be used to establish the business-as-usual (BAU) baseline that projects the future deforestation in the absence of a REDD project. Quantification of this projected deforestation and comparison with various governance scenarios provides the basis for evaluation of the carbon stocks involved and potential compensation.

It is hoped that the REDD program will lead to a significant reduction in CO2 emissions, enhance ecosystem services and biodiversity conservation, and provide opportunities for poverty reduction.

Implementing REDD projects will depend on modeling which requires significant data and computing resources. Google’s new Earth Engine platform will ameliorate the financial burden of implementing REDD projects.

“The objective of Earth Engine is to enable organizations such as Clark Labs to run their algorithms on-line, powered by Google’s computational capacity, with easy access to massive earth observation data sets,” indicated Dr. Amy Luers, Senior Environment Program Manager for Google.org.

At this time, many governmental and non-governmental organizations are using the Land Change Modeler as their primary analytical tool for REDD project development. Clark Labs, in conjunction with Conservation International, is currently augmenting its Land Change Modeler suite to ensure compatibility with the World Banks’ BioCarbon Fund (BioCF) methodological framework.

Modeling REDD-baselines using IDRISI’s LCM - Step 4

Modeling Future Emissions

Clark Labs is designing a REDD tool within the Land Change Modeler module of IDRISI. This tool will guide the user through the steps of the baseline development and directly produce tables and graphics necessary for reporting. This will result in a tool that can greatly reduce the time and cost in the development of a deforestation baseline, which currently is delaying the finalization of project development for many prospective REDD projects. The project will also include case studies, testing of different REDD scenarios and production of user guidelines. The tool will be co-designed by the Center for Applied Biodiversity Science (CABS) at Conservation International and Clark Labs. CABS will provide all case study data for testing the tool.

Figure 4. The REDD tab facility for the extraction and calculation of future carbon emission for the project area.

The following parameters are required; the reference area, the project area and the leakage area, the starting and ending date of the project and the number of reporting intervals. The carbon pools as well as the sources of green house gases to be included in the project also must be specified. Average carbon density can be entered either as a constant for each landuse or as a continuous surface image of carbon.

This information combined with the predicted landuse for the REDD project generates a series of tables of future emissions. These emissions estimates must be included in the final project to be submitted to VCS for approval.

Modeling REDD-baselines using IDRISI’s LCM - Step 3

Predicting Change

The Change Prediction tab provides the controls for a dynamic landcover change prediction process. Using the transition potential maps created in the Transition Potentials tab and after specifying the end date, the quantity of change in each transition can be modeled. Both hard and soft outputs are provided.

Figure 3. Hard and soft prediction images output from the model with the transition probability matrix of change.

This step allows one to determine the amount of change that will occur to some point in the future using the Markov Chain prediction process or a user-specified model. The amount of change can be determined by the default procedure: Markov Chain. Using the earlier and later landcover maps along with the date specified, it determines exactly how much land would be expected to transition from the later date to the prediction date based on a projection of the transition potentials into the future and creates a transition probabilities file. The transition probabilities file is a matrix that records the probability that each landcover category will change to every other category. Alternately, you can specify a transition probability file from an external model.

The final stage is the allocation process in which the parameters for the prediction are set and run the process. Both hard and soft prediction maps can be produced. The hard prediction is based on a multi-objective land competition model. The soft prediction output is a continuous map of vulnerability to change for a selected set of transitions. The soft prediction model is generally preferred for habitat and biodiversity assessment since it provides a comprehensive assessment of change potential.

Modeling REDD-baselines using IDRISI’s LCM - Step 2

Modeling the Potential for Change

In the Land Change Modeler, change analysis and prediction are organized around a series of empirically evaluated transition sub-models. A transition sub-model can consist of a single landcover transition or a group of transitions that are thought to have the same underlying driver variables. All selected transition sub-models must be modeled before change prediction can be undertaken. The Transition Potentials tab allows one to group transitions between two landcover maps into a set of sub-models resulting in a transition potential map for each transition—an expression of time-specific potential for change. Transitions are modeled using either Logistic Regression or a Multi-Layer Perceptron (MLP) neural network.

Figure 2. Transition potential image derived from a set of driver variables such as slope, distance from roads, distance from towns, elevation, etc...

Driver variables can be added to the model either as static or dynamic components. Static variables express aspects of basic suitability for the transition under consideration, and are unchanging over time. Dynamic variables are time-dependent drivers such as proximity to existing development or infrastructure (roads) and are recalculated over time during the course of a prediction.

Once model variables have been selected here, each transition is modeled in the Run Transition Sub-Model panel using either Logistic Regression or the Multi-Layer Perceptron (MLP) neural network. The MLP neural network offers an automatic mode that requires no user intervention. At this stage samples are extracted from the two landcover maps of areas that underwent the transitions being modeled as well as the areas that were eligible to change, but did not. The result in either case is a transition potential map for each transition.

Modeling REDD-baselines using IDRISI’s LCM- Step 1

Analyzing Change

The first stage of a REDD project is the understanding of the historical trends in landuse change. The Change Analysis tab provides a set of tools for the rapid assessment of change, allowing one to generate a series of evaluations of gains and losses, net change, persistence and specific transitions both in map and graphical form. For the change and prediction analyses, a minimum requirement is the specification of two landcover maps that can be used as the basis of understanding the nature of change in the study region and the means of establishing samples of transitions that should be modeled. Other essential files associated with the study area are also specified: reference area, basis roads, and elevation.

Figure 1. Illustrates the historical pattern of change between three dates in time of the Ankeniheny-Mantadia Corridor in Madagascar.

The Change Analysis panel provides a series of very useful graphs of change between the two maps specified in the Project Parameters panel. One can view these graphs in a variety of units (cells, hectares, square kilometers, acres, square miles, % of change and % of area). This rapid assessment allows for an understanding of the dynamic of change among the different landuse categories of the two time periods.

With a greater understanding of the historical changes, the Change Maps panel provides the ability to create a spatial representation of the historical change, including maps of persistence, gains and losses, transitions and exchanges.

Modeling REDD-baselines using IDRISI's LCM - Introduction

Introduction

Initiatives to implement REDD at the site level require the development of a reference scenario of greenhouse gas emissions. This is an estimate of the probable emissions rate throughout a REDD project period if the project were not to be implemented. This must be calculated for the project area as well as the surrounding leakage area. During project implementation, the actual emissions are monitored in both areas and compared to the reference scenario emissions to calculate the creditable emissions reductions.

Several methods for estimating reference emissions levels have been proposed to the Voluntary Carbon Standards (VCS) group, and all require modeling future emissions based on historical trends in rates and relationships between deforestation patterns and drivers of deforestation. One of the first methodologies submitted to VCS for review is the Mosaic methodology, submitted by the World Bank BioCarbon Fund. As part of this submission, the application of this methodology for the Mantadia REDD site in Madagascar was provided by Conservation International (CI) and Clark Labs at Clark University.

The steps include:

1) Definition of the spatial and temporal boundaries of the project, including the project site and leakage area.

2) Estimation of historical deforestation rates and patterns over at least three time periods covering at least 10 years, the first period used to calibrate the spatial model and the second period to validate it.

3) Projection of the future deforestation rate for the study area, which may be different from the historical rate, given assumptions of population change and infrastructure development.

4) Spatial modeling to determine the patterns of the potential for deforestation, based on the relationships between historical deforestation patterns and spatial variables that characterize land access and suitability.

5) Spatial modeling of deforestation patterns for the validation period, based on (3) and (4).

6) A comparison of the estimated and actual deforestation rates inside vs outside of the project area, in order to validate the model’s ability to estimate the amount of deforestation that will occur inside the project site.

7) Spatial modeling of deforestation patterns for each reporting period throughout the project duration.

8) Combination of the deforestation projection and data on biomass to estimate GHG emissions over each reporting period, for comparison with actual emissions estimates to be produced during the project-monitoring phase of the project.