predicting future forest loss in the democratic republic of the congo's

biodiverse forest ecosystems through sustainable land ..... across the CARPE Landscapes could be important in .... The confusion matrix is designed as a way.
3MB taille 1 téléchargements 25 vues
TECHNICAL NOTE

PREDICTING FUTURE FOREST LOSS IN THE DEMOCRATIC REPUBLIC OF THE CONGO’S CARPE LANDSCAPES ELIZABETH GOLDMAN, NANCY HARRIS, AND THOMAS MASCHLER

EXECUTIVE SUMMARY The Central Africa Regional Program for the Environment (CARPE) is a long-term initiative of the U.S. Agency for International Development (USAID). CARPE’s main objectives are to help mitigate climate change and maintain the ecological integrity of Central Africa’s biodiverse forest ecosystems through sustainable land management. The Democratic Republic of the Congo (DRC) is the largest forested country in CARPE. As its economy and infrastructure investments continue to grow, it will be critical for CARPE to better understand current and likely future land cover change dynamics within these high-value forested areas. To identify key drivers of forest loss and project the locations of potential future forest loss in DRC’s six CARPE Landscapes from 2015 to 2025, we used a spatial modeling approach. Across all six landscapes, we identified the most important influences on past forest loss— biophysical conditions (rainfall and topography); accessibility (distance to rivers, roads, settlements, conflict and areas of shifting cultivation); and land management (protected area and logging concession status)—and plotted where they are most likely to occur in the landscapes. Under a business-as-usual scenario, approximately 332,200 hectares of forest area within these landscapes are projected to be lost between 2015 and 2025, with an associated 205 million metric tons of carbon dioxide (CO2) emitted to the atmosphere, or 20.5 million metric tons of CO2 per year, due to forest clearing.

CONTENTS Executive Summary ..............................................1 Background ....................................................... 2 Methods........................................................... 4 Results............................................................. 7 Discussion........................................................ 13 Conclusions ..................................................... 16 Annex 1. Model Description..................................... 17 Endnotes ......................................................... 18 References....................................................... 18 Technical notes document the research or analytical methodology underpinning a publication, interactive application, or tool.

Suggested Citation: Goldman, E., N. Harris, T. Maschler. 2015. “Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes.” Technical Note. Washington, D.C.: World Resources Institute. Available online at: www.wri.org/publication/congo-carpe-landscapes.

TECHNICAL NOTE | June 2017 | 1

Most of the loss (59 percent) is projected to occur within the Ituri-Epulu-Aru Landscape (34 percent), located in eastern Orientale Province, and the Lac Télé-Lac Tumba Landscape (25 percent), in Équateur and Bandundu Provinces. In its nationally determined contribution (NDC) under the 2015 Paris Agreement, DRC stated a goal to implement a 3-million-hectares afforestation/ reforestation program resulting in carbon sequestration estimated to reach 3 million metric tons of CO2 by 2025. This study suggests that full protection of CARPE’s forest landscapes that prevents all future deforestation could lead to climate benefits up to 68 times higher than this. Even reducing the rate of future forest loss by just 10 percent would lead to emission reductions of 20 million metric tons of CO2 by 2025. This study provides a foundation for the development of alternative scenarios that can be used to understand both the potential emission impacts of future planned activities, such as infrastructure projects, as well as the emission reduction and sequestration potential of CARPE’s forested landscapes.

BACKGROUND Tropical forests are an important ecological and economic resource that provide habitat for biological organisms and livelihoods for human populations. These forests are fundamental to Earth’s biogeochemical cycles and provide ecosystem services, such as carbon sequestration, climate regulation, water and air filtration, and habitat for two– thirds of the world’s terrestrial plant and animal species (Bradshaw et al. 2009). Many indigenous populations and others living in or near forests rely directly on the forest’s goods and services, such as food, medicine, fuel, and shelter. Despite the need for these services, rates of forest loss are high in many countries (Hansen et al. 2013). Loss of tropical forest habitat leads to the decline and extinction of species (Rolland et al. 2014) and puts at risk indirect benefits such as sediment retention and water quality. The Congo Basin is the second largest contiguous forest in the world, after the Amazon Basin. Approximately 60 million people living inside or near Central African forests rely on them for subsistence (de Wasseige et al. 2015). Certain forests are considered sacred and are of great cultural or religious value to numerous communities. Rich biodiversity can be found in the Congo Basin, including approximately 1,300 bird species, 336 amphibian

2 |

species, 400 reptile species, 20,000 plant species, and many iconic mammals, such as bonobos, chimpanzees, forest elephants, and lowland and mountain gorillas (de Wasseige et al. 2015). About half of the forest in the Congo Basin is in the Democratic Republic of the Congo (DRC), where forests constitute 62 percent of the national territory. Forest loss in DRC has increased dramatically from 0.23 percent per year in 2001 to 0.55 percent per year in 2014 (Hansen et al. 2013). This rate is relatively high compared with the 2014 rates of most other Congo Basin countries, including Cameroon (0.44 percent), Central African Republic (0.1 percent), Republic of the Congo (0.22 percent), and Gabon (0.17 percent).1

Drivers of Deforestation The drivers of deforestation in DRC include shifting cultivation (a rotational farming system in which land is cleared for agriculture and left to regenerate back to forest after a few years), woodfuel and brick production, artisanal and commercial logging, mining, conflict, infrastructure development, and bush fires (Megevand et al. 2013; Molinario, et al. 2015; Nackoney et al. 2014; Shapiro et al. 2016). Underlying influences that sustain these drivers include increasing human populations, economic growth, and consumption patterns. Drivers are frequently differentiated as direct and indirect (Geist and Lambin 2001). Direct deforestation drivers are defined as local anthropogenic actions that lead to the conversion of forest to other land uses, such as urbanization, agriculture, and mining. The impact of direct drivers can be influenced by natural processes that trigger deforestation. Biophysical characteristics can influence the likelihood of a forest being converted to other land uses. For example, the amount of precipitation and the soil quality are considered before deciding whether to convert land for agriculture. Landscape characteristics, such as slope and the presence of water bodies, can make forests more or less accessible or desirable for conversion. Indirect deforestation drivers are typically social and economic forces that occur on the local, national, or global levels. Indirect drivers such as the management of landscapes, growth of populations and consumption, and economic growth can be both the underlying causes and the sustaining factors for direct drivers (Geist and Lambin 2001).

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Landscapes Identified by CARPE To better understand the remaining forests and biodiversity in the Congo Basin and maintain their ecological integrity over the long term, the U.S. Agency for International Development (USAID) established the Central Africa Regional Program for the Environment (CARPE) in 1995. The program has been implemented in phases over the past 22 years, with the landscape component established in 2000 to identify and conserve areas of biological significance and resiliency in the Republic of Congo and DRC. This study focuses on DRC’s six CARPE Landscapes: Ituri-Epulu-Aru, Lac Télé-Lac Tumba, Maiko-TaynaKahuzi-Biega, Maringa-Lopori-Wamba, Salonga-LukenieSankuru, and Virunga (Figure 1). These landscapes span 42.9 million hectares, contain 20 percent of DRC’s total forested area, and represent high-conservation-value forests for endangered flora and fauna (Hansen et al. 2013). Conflict, a suppressed economy, and limited infrastructure have slowed the development of DRC and its CARPE Landscapes, and despite recent increases in deforestation

Figure 1 |

Six CARPE Landscapes in the Democratic Republic of the Congo Central African Republic

South Sudan

Cameroon

Gabon

Ituri-Epulu-Aru

Maringa-Lopori-Wamba

Republic of Congo

Virunga

Lac Télé-Lac Tumba

Maiko-Tayna-Kahuzi-Biega Salonga-Lukenie-Sankuru Kinshasa

Tanzania

Angola

CARPE Landscape

0

250 km

Zambia

rates, much of its forest remains intact (Megevand et al. 2013). As DRC’s economy improves, the country’s forests are at an important juncture. The need for coordinated land use and conservation planning is critical.

Spatial Land Use Change Models Recent advances in remote sensing and spatial modeling to detect changes in land use and land cover are rising to meet this need. Spatial land use change models analyze historical patterns of change and allow for projections of potential future change, the exploration of alternative land use planning scenarios, and a quantitative description of the key variables that drive changes in land use and land cover. In DRC’s CARPE Landscapes, spatial models and data in high temporal and spatial resolution can assist with land management, and allow us to better understand, evaluate, and project the future role of forests in these landscapes. Many spatial modeling studies relate the location of historical land cover change with maps of variables that drive change to identify the areas with the highest susceptibility to future deforestation (Bax et al. 2016; Duong and Murayama 2010; Fuller, et al. 2011; Mas 2004; Zhang et al. 2006). Variables used to explain deforestation in these studies include biophysical landscape features, population density, and measures of distance to roads, towns, or other places. Nonspatial variables, such as sociocultural and political drivers, can be more difficult to incorporate. In this study, we used the IDRISI software Land Change Modeler (LCM), a modeling and scenario planning software available through Clark University.

Purpose of this Study The study has three objectives:

▪▪ ▪▪ ▪▪

Identify and assess the contribution of biophysical and anthropogenic driver variables in explaining historical forest loss patterns observed in DRC’s CARPE Landscapes. Estimate the likelihood of future transition from forest to nonforest. Project the potential state of CARPE Landscape forests in the year 2025.

The outputs of this study are intended to support the design and implementation of sustainable land use planning decisions by the CARPE Program.

Source: WRI Authors

TECHNICAL NOTE | June 2017 | 3

METHODS We used the IDRISI software Land Change Modeler (LCM), a spatial modeling software tool, to analyze historical forest loss, create forest loss risk maps, and model future forest loss and associated carbon emissions to the year 2025 for each CARPE Landscape. Other spatial modeling software includes Dinamica, CLUE-S, and Cellular Automata Markov (CA_MARKOV). Some researchers have attempted to compare model accuracy. However, because a single model can create various outcomes depending on model parameter selection that are greater in variation than alternative models, this type of evaluation is difficult (Pontius and Malanson 2005). Ultimately, LCM was selected for this study over other models because of its use of artificial neural networks to identify complex interactions among driver variables, its integration of scenario planning tools, and its relative ease of use. Additional information on this model is provided in Annex 1.

Study Area The study area comprises DRC’s six CARPE Landscapes situated in the western and eastern parts of the country (Figure 1). Each landscape contains areas under different types of management, primarily protected areas, community-based natural resource management zones, and extractive zones for timber and mining concessions (CBFP 2006). While the landscapes were originally identified for their high-value forested areas, a variety of land use and land cover types are present in all the landscapes. These include cities, smaller settlements, areas of shifting cultivation, and extractive zones. In the center of every landscape lies one or several protected areas managed by the national park agency (Institute Congolais pour la Conservation de la Nature; ICCN) in partnership with CARPE implementing partners such as the Wildlife Conservation Society (WCS), World Wildlife Fund (WWF), and African Wildlife Foundation (AWF). The CARPE partners’ work also extends beyond park management to engaging communities within the landscapes in natural resource management. The three landscapes in the western part of the country— Salonga-Lukenie-Sankuru, Maringa-Lopori-Wamba, and Lac Télé-Lac Tumba— are closer to the development center of Kinshasa, the country’s capital, as well as the mouth of the Congo River, a major transportation hub for people and goods. Landscapes in the mountainous, eastern part of the country— Maiko-Tayna-Kahuzi-Biega, Ituri-Epulu-Aru, and Virunga— are closer to much of the conflict that has

4 |

plagued the country over the past two decades. (Figure 1). Extractive activities differ in the two regions; most of the country’s logging concessions are in the western part of the country whereas mining is prevalent along the eastern border (Butsic et al. 2015). The western Salonga-LukenieSankuru Landscape is the most remote, with high forest cover and low population density. Because of their proximity to each other and the influence of different forest change drivers, CARPE Landscapes were grouped into two models: Western Landscapes of SalongaLukenie-Sankuru, Maringa-Lopori-Wamba, and Lac TéléLac Tumba and the Eastern Landscapes of Maiko-TaynaKahuzi-Biega, Ituri-Epulu-Aru, and Virunga. The Western Landscapes cover 26.3 million hectares and the Eastern Landscapes cover 16.8 million hectares.

Historical Forest Loss and Emissions Reference maps used to analyze historical forest cover loss in the CARPE Landscapes were created using the Hansen et al. (2013) version 1.2 of forest cover and forest cover loss products available on the Global Forest Watch website.2 For this analysis, we defined forest as any area with vegetation over 5 meters high and with a greater than 30 percent tree canopy cover. Forest loss is defined as complete removal of the tree canopy at the Landsat pixel scale by natural or human causes. Both forest extent and forest loss products are derived from 30-meter Landsat satellite data. Forest extent is representative of the year 2000, and annual forest loss data are available from 2001 to 2014. We used these data products to create reference land cover maps for the years 2000, 2007, and 2014 for two land categories: “forest” and “nonforest,” as well as to determine the historical rates and trends of annual forest loss.

Forest Loss Risk Maps After quantifying historical forest loss within DRC’s CARPE Landscapes, we selected variables that were thought to influence or contribute to these changes. These “influence variables” were incorporated into the model to create a map of forest loss risk, or the potential of any given pixel on the map to transition from forest to nonforest. We refer to these as forest loss risk maps. The influence variables were selected on the basis of a literature review of deforestation influence variables in DRC, as well as data availability. Data for the spatially explicit influence variables used in this analysis are

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

summarized in Table 1 under three categories: biophysical variables, accessibility variables, and land management variables. Biophysical variables included elevation, slope, and average precipitation. Spatial modeling studies in the Peruvian Amazon, Mexico, and Myanmar found elevation to be negatively correlated with forest loss (Bax et al. 2016; Mas 2004; Mon et al. 2012). Large differences in elevation across the CARPE Landscapes could be important in explaining forest loss patterns because lower elevations with gentler slopes are likely to be both more accessible and more suitable for alternative land uses such as agriculture due to their more suitable climate conditions.

Table 1 |

The variation in annual precipitation rates across the CARPE Landscapes, which is related partially to elevation, may be an important forest loss predictor because drier forests burn more easily (Aragao et al. 2008). Accessibility variables include navigable rivers, roads, settlements, shifting cultivation, and conflict areas. Navigable rivers and roads are likely to influence the landscape by providing access to forested areas (Ali et al. 2015; Li et al. 2015; Megevand et al. 2013). Settlement data indicate an increased presence of humans who are

Variables and Data Sources Used in the Model

VARIABLE

DATE

SOURCE

TYPE/RESOLUTION

DATA PREPARATION

Elevation

2000

SRTMa

Raster, 30 meter

Resample

Slope

2000

SRTM

Raster, 30 meter

Slope tool, from elevation

Average precipitation

1960–1990

WorldClimb

Raster, 300 meter

Resample

Distance to navigable rivers

2016

CICOSc

Vector

Euclidian distance to navigable rivers

Distance to roads

2017

OSM logging roadsd

Vector

Euclidian distance to roads

Distance to settlements

2016

DIAF and MEDDe

Vector

Euclidian distance to settlements

Rural complex (shifting cultivation)

2000

University of Marylandf

Raster, 60 meter

Resample

Distance to armed conflict

2016

ACLEDg

Vector

Euclidian distance to events

Protected areas

2016

WDPAh

Vector

Polygon to raster conversion

Logging concessions

2016

DIAF and DGFi

Vector

Polygon to raster conversion

Biophysical

Accessibility

Land Management

Note: For explanation of data preparation, see text. Sources: a. Shuttle Radar Topography Mission. 2000. U.S. Geological Survey. https://earthexplorer.usgs.gov/. b. Hijmans, R., S. Cameron, J. Parra, P. Jones, and A. Jarvis. 2005. “Very High Resolution Interpolated Climate Surfaces for Global Land Areas.” International Journal of Climatology 25: 1965–78. c. Commission Internationale de Bassin Congo-Oubangui-Sangha. 2016. “Atlas du Bassin du Congo.” d. Open Street Map Logging Road initiative. 2017. https://loggingroads.org/. e. Direction Inventaire et Aménagement Forestiers and Ministère de l’Environnement et Développement Durable. 2016. http://cod.forest-atlas.org. f. M  olinario, G., M. Hansen, and P. Potapov. 2015. “Forest Cover Dynamics of Shifting Cultivation in the Democratic Republic of Congo: A Remote Sensing–Based Assessment for 2000–2010.” Environmental Research Letters 10 (9): 94009. g. Raleigh, C., A. Linke, H. Hegre, and J. Karlsen. 2010. “Introducing ACLED – Armed Conflict Location and Event Data Project.” Journal of Peace Research 47 (5) 651–60. h. IUCN and UNEP-WCMC. 2016. The World Database of Protected Areas (WDPA). Cambridge, UK: UNEP-WCMC. www.protectedplanet.net. i. Direction Inventaire et Aménagement Forestiers and Direction de Gestion Forestière. 2016. http://cod.forest-atlas.org.

TECHNICAL NOTE | June 2017 | 5

likely to increase forest loss (Mon et al. 2012; Moone et al. 2016). Conflict locations were included because human displacement caused by conflict often results in forest loss. The rural complex variable represents areas of shifting cultivation— a mosaic of active and fallow agriculture and secondary forest— which is a common driver of land use change in the region (Molinario et al. 2015).

represents perfect spatial agreement between the model and observed outcome, whereas a value of 0.5 indicates that the model performed no better than random chance. The forest loss risk map with the highest ROC score was selected and used in the forest loss prediction process.

Land management variables included land managed as protected areas and as logging concessions.

Modeling forest loss to a future date relies on both the projected quantity of future forest loss and the forest loss risk map. For this analysis, maps of predicted forest loss to the year 2025 were created for two future scenarios. Scenario 1, the more conservative scenario, projected annual future forest loss as a continuation of the historical average annual forest loss rate (2000– 2007) within the western and eastern CARPE landscape groups. Scenario 2, the more aggressive scenario, projected that annual future forest loss would increase linearly over time from the observed historical trend (Figure 2). The historical rates and trends of annual forest loss in these two scenarios served as the basis for converting pixels in the forest loss risk map.

Some variables were excluded because spatial data were either nonexistent or not readily accessible at the time of analysis. These variables included soil fertility, agricultural productivity, mineral deposits, GDP, and a variable for energy consumption, such as night lights, electrification rate, or charcoal consumption/production rate. Acquisition of additional data layers could improve prediction maps as noted in the Discussion section, below. Because the modeling software used in this study required consistency in cell size and projection across all input variables, all layers were reprojected into WGS 1984, resampled using the nearest neighbor algorithm to match the forest loss data (a cell size of 0.00025 degrees, or approximately 30 meters at the equator), and extracted (clipped) with the CARPE Landscape borders. Euclidian distance was calculated to roads, rivers, settlements, and conflict areas to create layers with pixel values representing the distance to the closest feature of interest. Once all influence variables were prepared, they were incorporated into the modeling software and combined with historical forest loss data over multiple iterations to test the extent to which different variable combinations held explanatory power in defining where past forest loss had occurred. Candidate forest loss risk maps were assessed using a receiver operating characteristic (ROC) statistic (Pontius and Schneider 2001). ROC evaluates whether observed forest loss was concentrated within locations of high risk. ROC uses an “area under the curve” (AUC) value to measure the strength of the model. An AUC value of 1

6 |

Forest Loss Prediction Maps

As the model projected forest loss through time, pixels in the forest loss risk map with the highest change potential values that had not yet been converted were filled in with forest loss, starting with areas of highest forest loss risk and continuing to areas of lower forest loss risk until the specified quantity of forest loss was reached for a given year. The resulting forest loss prediction maps were then overlaid with a 30-meter resolution map of aboveground biomass density for the year 2000 to produce estimates of carbon emissions resulting from the predicted loss of this aboveground forest biomass (Zarin et al. 2016), assuming that a forest loss pixel results in the loss of all associated aboveground biomass from that pixel and all emissions occur within the year of clearing. Carbon present in other forest pools (roots, dead wood, litter, soil organic carbon) was excluded from our analysis.

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Figure 2 |

Rates and Trends of Forest Loss for Western and Eastern Carpe Landscapes, 2001–2007

Eastern Landscapes

Western Landscapes Trend = 0.68x + 25.6

Trend = 1.02x + 28.29 Average = 32.38

Average = 28.40

35

Forest Loss (1000 hectares)

Forest Loss (1000 hectares)

40 30 25 20 15 10 5 0 2001

2002

2003

2004

2005

2006

2007

Year

50 45 40 35 30 25 20 15 10 5 0 2001

2002

2003

2004

2005

2006

2007

Year

Source: WRI Authors

To validate a prediction map, a modeled outcome must be compared with an observed outcome. We used out-of-sample testing to re-create observed historical change in the model and compared the results to actual observations from the same time period. We used reference data from 2000 to 2007 to calibrate the forest loss risk map and used data from 2008 to 2014 as our validation period to evaluate model performance. These periods were chosen so that the same number of years could be used in both model calibration and model validation periods. Validation statistics of overall accuracy, kappa value, user’s accuracy, and producer’s accuracy were generated from a confusion matrix to assess overall model performance. A confusion matrix, also known as an error matrix, is a table that allows evaluation of model performance where each column of the matrix represents results from the prediction model and each row represents results from actual observations (or vice versa). The confusion matrix is designed as a way to see whether the model is confusing two classes (i.e., forest with nonforest).

RESULTS Results from this study include an analysis of historical forest loss in primary and secondary forests, both inside and outside each CARPE Landscape. Forest loss risk maps were produced for both groups of landscapes, and the influence variables were ranked in order of their contribution to predicting forest loss. The forest loss prediction maps display likely land cover maps in the year 2025 under the two scenarios. Results were validated using a ROC statistic and a confusion matrix.

Historical Forest Loss and Emissions During the years 2001 through 2014, DRC’s CARPE Landscapes lost 2.8 percent (1.1 million hectares) of their 2000 forested area, corresponding to an annual forest loss rate of 0.20 percent per year. The rate of forest loss increased from 0.15 percent per year in the early 2000s (2001–2007) to 0.25 percent per year from 2008 to 2014. Over this 14-year period, these landscapes lost 2.5 percent of their carbon stocks in above-ground biomass, corresponding to 525 million metric tons of carbon emissions from biomass loss, or 38 million metric tons per year. The average rate of forest loss within the CARPE Landscapes was 35 percent lower than for the rest of DRC’s forests, or a difference of a tenth of one percent (Table 2).

TECHNICAL NOTE | June 2017 | 7

Table 2 |

Historical Forest Extent, Loss, Biomass, and Emissions Inside and Outside DRC’s CARPE Landscapes WITHIN CARPE LANDSCAPES V

ALL

OUTSIDE CARPE LANDSCAPES

9.95

0.39

35.36

69.22

104.58

0.59

0.38

0.98

5.18

89.47

94.64

7.21

10.33

1.37

40.54

158.69

199.22

5.2

3.6

5.2

0.7

20.3

79.7



145,495

164,646

62,851

46,555

30,330

528,767

2,080,097

2,608,864

2.1

2.5

1.9

0.9

0.5

7.9

1.5

3.0

2.5

ha

33,364

111,436

212,328

113,269

64,130

56,654

591,182

4,776,964

5,368,146

%

13.5

8.2

13.2

19.3

16.7

5.8

11.4

5.3

5.7

ha

112,253

256,932

376,973

176,120

110,685

86,985

1,119,948

6,857,061

7,977,010

%

2.8

3.6

3.6

2.4

1.1

6.4

2.8

4.3

4.0

%

1.4

3.2

4.7

2.2

1.4

1.1

14.0

86.0



ha/yr

5,635

10,393

11,760

4,489

3,325

2,166

37,769

148,578

186,347

%/yr

0.1

0.2

0.1

0.1

0.03

0.6

0.1

0.2

0.2

ha/yr

2,383

7,960

15,166

8,091

4,581

4,047

42,227

341,212

383,439

%/yr

1

0.6

0.9

1.4

1.2

0.4

0.8

0.4

0.41

ha/yr

8,018

18,352

26,927

12,580

7,906

6,213

79,996

489,790

569,786

UNIT

IEA

LT-LT

MTKB

MLW

Primary forest

Mha

3.83

5.79

8.79

6.62

Secondary forest

Mha

0.25

1.36

1.61

Total

Mha

4.07

7.15

10.41

% of DRC total

%

2

3.6

ha

78,889

%

SLS

NATIONAL

Area Tree cover extent (2000)

Primary forest

Total tree cover loss (2001–2014)

Secondary forest Total % of DRC total Primary forest

Average rate of tree cover loss (2001–2014)

Secondary forest Total

%/yr

0.2

0.3

0.3

0.2

0.1

0.5

0.2

0.3

0.3

% of DRC total

%/yr

1.4

3.2

4.7

2.2

1.4

1.1

14.0

86.0



Primary forest

Tg C

653

673

1,340

987

1,488

53

5,193

10,277

15,470

Secondary forest

Tg C

29

80

167

69

38

74

458

5,923

6,381

Total

Tg C

682

752

1,506

1,056

1,527

127

5,650

16,200

21,850

% of DRC total

%

3.1

3.4

6.9

4.8

7

0.6

25.9

74.1



Primary forest

Mg C/ha

171

116

152

149

150

137

147

148

148

Secondary forest

Mg C/ha

120

59

103

118

100

75

88

66

67

Total

Mg C/ha

168

105

145

146

148

93

139

102

110

Primary forest

Mt CO2

43

66

93

34

25

16

276

1,122

1,398

Secondary forest

Mt CO2

16

37

96

50

27

22

248

1,763

2,011

Total

Mt CO2

59

103

189

84

52

38

524

2,885

3,409

% of DRC total

%

1.7

3

5.5

2.5

1.5

1.1

15.4

84.6



Primary forest

Mt CO2/yr

3

5

7

2

2

1

20

80

100

Secondary forest

Mt CO2/yr

1

3

7

4

2

2

18

126

144

Total

Mt CO2/yr

4

7

14

6

4

3

37

206

244

% of DRC total

%/yr

1.7

3

5.5

2.5

1.5

1.1

15.4

84.6



Carbon Total carbon in aboveground biomass (2000)

Average aboveground carbon density (2000)

Total CO2 emissions (2001-14)

Average CO2 emissions (2001-14)

Notes: DRC = Democratic Republic of the Congo, CARPE = Central Africa Regional Program for the Environment, IEA = Ituri-Epulu-Aru, LT-LT = Lac Télé-Lac Tumba, MTKB = Maiko-Tayna-KahuziBiega, MLW = Maringa-Lopori-Wamba, SLS = Salonga-Lukenie-Sankuru, and V = Virunga Source: WRI Authors

8 |

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Forest Loss Risk Maps The forest loss risk maps produced by this study can serve as the basis for generating prediction maps of future forest loss. Figure 3 shows the likelihood of forest loss within the CARPE Landscapes, with clustering of highforest-loss risk areas near the rural complex (shifting cultivation), settlements, and roads, indicating the strong influence of human presence on the landscape. The influence of each variable used in the model to explain past forest loss patterns in the CARPE Landscapes is summarized in Table 3. The rank order shows the variable with the most influence (1) to the least influence (10). The measures of Cramer’s V, Accuracy, and Skill Measures in the last three columns, which help determine the rank of the variables, are explained in the table footnotes. Both Figure 3 and Table 3 demonstrate that human presence had a significant influence on forest loss in both models, with the rural complex (shifting cultivation)

Figure 3 |

areas ranking first in both models, and navigable rivers (Western Landscape Model) and roads (Eastern Landscape Model) ranking in the top three. Risk maps show that the highest potential for transition from forest to nonforest is in rural complex areas and along roads. The largest difference in variable influence between the two models is elevation, which was the lowest ranking variable in the flatter Western Landscapes but the fourth highest in the mountainous Eastern Landscapes. Distance from navigable rivers also influenced forest loss patterns in the two landscapes differently, ranking third in the more populated Western Landscapes, but last in the Eastern Landscapes. Both forest loss risk maps have an ROC score above 0.90, with Western Landscapes at 0.923, and Eastern Landscapes at 0.903. This indicates that the observed change from 2008 to 2014 was strongly clustered in the areas that were mapped as having high risk for converting from forest to nonforest.

Forest Loss Risk Map for Western and Eastern CARPE Landscapes

Western Landscapes

Forest Loss Risk

Eastern Landscapes

High risk Low risk Protected area

Republic of Congo Uganda

Rwanda

Burundi

0

50 Km

0

50 Km

Tanzania

Notes: Red and orange pixels represent areas with a higher likelihood of forest loss. Source: WRI Authors.

TECHNICAL NOTE | June 2017 | 9

Table 3 |

Ranking of the Variables Used in the Forest Loss Risk Maps for Western and Eastern CARPE Landscapes According to Their Influence on Forest Loss WESTERN LANDSCAPES

RANKa

VARIABLE

CRAMER’S Vb

ACCURACY (%)c

SKILL MEASUREd

1

Rural complex

.1364

73.98

.4796

2

Average precipitation

.2112

83.02

.6603

3

Distance from navigable rivers

.2384

83.87

.6774

4

Distance from roads

.0804

84.25

.6850

5

Distance from settlements

.1208

84.57

.6904

6

Logging concession

.0328

84.77

.6955

7

Distance from conflict

.0559

84.78

.6957

8

Protected area

.0666

84.91

.6983

9

Slope

.2307

85.05

.7011

10

Elevation

.4006

85.12

.7025

EASTERN LANDSCAPES RANK

VARIABLE

CRAMER’S V

ACCURACY (%)

SKILL MEASURE

1

Rural complex

.2440

72.87

.4574

2

Distance from roads

.1589

78.58

.5716

3

Average precipitation

.4167

80.29

.6058

4

Elevation

.1708

80.32

.6064

5

Protected areas

.0932

80.58

.6117

6

Distance from conflict

.1841

80.95

.6189

7

Slope

.2297

80.98

.6195

8

Distance from settlements

.0989

80.98

.6195

9

Logging concessions

.0547

81.04

.6207

10

Distance from navigable rivers

.3090

81.06

.6211

Notes: a. Rank represents the influence of the variable on the transition from forest to nonforest. b. Cramer’s V is a correlation coefficient that represents a test of the potential explanatory power of each variable, with values ranging from 0.0 for no correlation, to 1.0 for perfect correlation. c. A ccuracy (percent) is calculated using randomly generated validation points to show how much the addition of each variable in the model improves the agreement between observed persistence and change with the model output. d. Skill measure is calculated by subtracting the accuracy of the model by the accuracy expected by chance. Values range from -1 to 1, with 1 indicating perfect prediction and 0 indicating a value close to random chance. Source: WRI Authors.

10 |

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Forest Loss Prediction Maps Two scenarios were considered to generate prediction maps for the years 2015 through 2025. Scenario 1 was based on the assumption that the historical average rate of forest loss from 2000 to 2007 would continue into the future (Figure 4). Prediction maps generated from Scenario 1 show that Figure 4 |

332,200 hectares of forest could be converted to other uses between 2015 and 2025, potentially resulting in emissions of 205 million metric tons of CO2. The majority of this forest loss is expected to occur in Ituri-EpuluAru (34 percent) and Lac Télé-Lac Tumba (25 percent), with loss concentrating near the rural complex areas,

Predicted Future Forest Loss under Two Scenarios, 2015–2025

Scenario 1: Predicted Forest Loss Using Historical Average Forest Loss Western Landscapes

Future loss 2015–2025

Eastern Landscapes

Forest loss 2001–2014 Republic of Congo

Forest Non-forest Uganda

Protected area

Rwanda

Burundi 0

50 Km

0

Tanzania

50 Km

Scenario 2: Predicted Forest Loss Using the Historical Trend of Forest Loss Additional Loss Future loss 2015–2025 Forest loss 2001–2014

Republic of Congo

Forest

Uganda

Non-forest Protected area

Rwanda

Burundi 0

50 Km

0

50 Km

Tanzania

Source: WRI Authors.

TECHNICAL NOTE | June 2017 | 11

along transportation corridors, and near settlements in all landscapes (Figure 3). Under this scenario, 22,300 hectares of forest loss is projected to take place in protected areas, 85,400 hectares in mining concessions, and 147,300 hectares in logging concessions. Protected areas in Salonga-Lukenie-Sankuru and Lac Télé-Lac Tumba Landscapes are particularly vulnerable and are projected to lose 11,700 hectares (0.3 percent of total forests in the landscapes) and 5,000 hectares of forest (0.4 percent), respectively. All forest loss in SalongaLukenie-Sankuru is expected to take place in Oshwe Game Reserve, and in Lac Télé-Lac Tumba, most forest loss (4,700 hectares) is expected to take place in TumbaLediima Nature Reserve.

Table 4 |

Scenario 2 assumed that the increasing trend in forest loss rate observed from 2000 through 2007 would continue into the future for each landscape region (Figure 4). This scenario resulted in an additional 578,800 hectares of forest loss by the year 2025, totaling 911,000 hectares and 401 million metric tons of CO2 emissions. Forest loss continued to concentrate near the loss areas in Figure 3, particularly around National Road No. 3 in MaikoTayna-Kahuzi-Biega Landscape; northwest of Beni in Salonga-Lukenie-Sankuru where conflict has occurred; and in the southwest areas of Lac Télé-Lac Tumba and Salonga-Lukenie-Sankuru Landscapes where settlements are found near roads and navigable rivers. Under this scenario, 41,900 hectares of forest loss is likely to take place in protected areas, 139,800 hectares in mining concessions, and 182,900 hectares in logging concessions.

C onfusion Matrix for Forest Loss Prediction Results in Scenario 1 Western CARPE Landscapes FOREST LOSS PREDICTION

Forest Loss Reference

NONFOREST

FOREST

TOTAL

PRODUCER’S ACCURACY (%)

Nonforest

1,621,749

325,656

1,947,405

83.3

Forest

177,773

24,131,739

24,309,512

99.3

Total

1,799,522

24,457,394

User’s Accuracy (%)

90.1

98.7

OVERALL (%)

98.1

KAPPA

0.855 Eastern CARPE Landscapes FOREST LOSS PREDICTION

Forest Loss Reference

NONFOREST

FOREST

TOTAL

PRODUCER’S ACCURACY (%)

Nonforest

901,689

316,069

1,217,758

74.0

Forest

191,011

15,180,409

15,371,420

98.8

Total

1,092,700

15,496,478

User’s Accuracy (%)

82.5

98.0

OVERALL (%)

96.9

KAPPA

0.764

Notes: Hectares unless otherwise noted Source: WRI Authors

12 |

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Table 5 |

C onfusion Matrix for Forest Loss Prediction Results in Scenario 2 Western CARPE Landscapes FOREST LOSS PREDICTION

Forest Loss Reference

NONFOREST

FOREST

TOTAL

PRODUCER’S ACCURACY (%)

Nonforest

1,624,120

323,285

1,947,405

83.4

Forest

196,195

24,113,316

24,309,512

99.2

Total

1,820,315

24,436,601

User’s Accuracy (%)

89.2

98.7

OVERALL (%)

98.0

KAPPA

0.852 Eastern CARPE Landscapes FOREST LOSS PREDICTION

Forest Loss Reference

NONFOREST

FOREST

TOTAL

PRODUCER’S ACCURACY (%)

Nonforest

906,402

311,357

1,217,758

74.4

Forest

214,740

15,156,680

15,371,420

98.6

Total

1,121,142

15,468,036

User’s Accuracy (%)

80.8

98.0

OVERALL (%)

96.8

KAPPA

0.758

Notes: Hectares unless otherwise noted Source: WRI Authors

Model Validation Based on results from the confusion matrix, both scenarios performed well in predicting future land cover between 2008 and 2014 based on training the model with data from the 2000–2007 period. When compared to the 2014 reference map, Scenario 1 prediction maps have an overall accuracy of 98.1 percent for Western Landscapes and 96.9 percent for Eastern Landscapes (Table 4). Scenario 2 prediction maps have slightly lower overall accuracy, with 98.0 percent for Western and 96.8 percent for Eastern Landscapes (Table 5). Eastern Landscape models in both scenarios have lower producer’s accuracies and kappa values, meaning there are more areas of nonforest called forest (false positives) in these landscapes and that the models did not predict future land cover as well.

DISCUSSION

Data and Model Performance Using historical data as a basis to predict forest loss to 2025, 0.33 million hectares (Scenario 1) to 0.91 million hectares (Scenario 2) of forest are predicted to be lost in DRC’s CARPE Landscapes, with an associated 205 million to 401 million metric tons of CO2 emitted to the atmosphere, respectively. CARPE Landscapes contain high value tropical forests that are under pressure from an increasing population and improving economy in DRC, putting the ecosystem services they provide and the

TECHNICAL NOTE | June 2017 | 13

populations who rely on them at risk. This study provides the methodology and results to project the effects of future activities, such as infrastructure projects and conservation efforts, through scenario planning.

Limitations to This Study While prediction models can lead to a better understanding of land change dynamics in a region, a number of assumptions and limitations must be considered. Predictions are based on a model calibration period that incorporates historical information, and a major model assumption is that what occurred in the recent past is a good predictor of what is most likely to occur in the near future, and that the variables that influenced the spatial patterns of past forest loss are the same variables that are likely to influence the spatial patterns of future forest loss. Thus, if no historical precedent existed for a specific policy or infrastructure change that occurs in the future, then model predictions will be inaccurate. There are some limitations to the forest area change data used in this analysis. While the Hansen et al. gross forest loss dataset is groundbreaking due to its global coverage, high (30 meters) spatial resolution, and high accuracy at the global scale, the accuracy of the data within a given region or country varies considerably. Based on samples from multi-temporal Landsat and high resolution imagery, Tyukavina et al. (2015) calculated that version 1.0 of the Hansen et al. product (2001–2012) underestimated forest loss in DRC by between 11 and 54 percent, although missed loss was located near detected loss. The updated loss detection methodology used for versions 1.1 (2001–2013) and 1.2 (2001–2014) have not yet been fully validated. These newer versions (version 1.2 used in this analysis) also contain some temporal inconsistencies due to an improved loss detection methodology for years since 2013. Data from the new Landsat 8 sensor became available starting in 2013, and an updated Hansen et al. forest loss algorithm was applied starting with the 2014 annual update. The incorporation of Landsat 8 data led to improved detection of loss in boreal forests, smallholder agricultural clearing, selective logging, and short-cycle plantation clearing. The entire loss time series (2001–2014) is being updated using the new algorithm. In the meantime, a moving window has been applied whereby with each new year of loss data, the previous two years are also updated to reflect the new

14 |

algorithm. A validation study is planned and may find a more sensitive method for detecting forest disturbance with the newly incorporated Landsat 8 data. Once the forest loss product is available with algorithm updates, our models should be rerun to incorporate a forest loss product that is more consistent across the time series. It is also possible to run our models using alternate forest change data, if a different product were shown to be more accurate than the Hansen et al. product for DRC. The reference land cover maps used in this analysis are a simplified representation of the actual land cover in CARPE Landscapes that account for only two land cover types: forest and nonforest. Other land cover types, such as agriculture and urban development would ideally be incorporated into future models. The forest maps also represent gross loss across years rather than net changes; therefore, highly dynamic landscapes such as shifting cultivation areas may not be appropriately captured by including only the loss component. From 2001 through 2012, DRC gained 1.39 million hectares of forest that is not accounted for in this study. If annual gain data become available, it will be possible to incorporate information on net forest loss into the model. Finally, forest degradation, while common in DRC, is difficult to capture on a large scale or in automated datasets and is generally not captured well by the forest loss data. In addition to the limitations of the reference maps, the absence of certain influence variables may have impacted the results of this study. Ideally, additional variables that represent pressure from intensive agriculture, mineral deposits, GDP, and energy consumption will be included in future iterations of the model to test whether these variables more accurately predict future forest loss. Some existing variables could also be improved. In particular, population is currently represented by distance from settlement locations. Many low-resolution population density datasets exist, but a higher resolution population density and change in population density dataset would likely lead to better model performance. The distance from roads variable is crowdsourced and while checked and updated periodically with missing road data, it is still incomplete. The dataset has incomplete attribute information that could be helpful in future model iterations, including road surface, creation date, and active dates. Despite model limitations, accuracy results indicate

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Table 6 |

Emission Reduction Scenarios for 2025 for CARPE Landscapes, Based on Average Rate of Forest Loss, 2001–2014 CARBON EMISSIONS SAVINGS (MT CO2)

Rates of Forest Loss Reduction (%)

IEA

LT-LT

MTKB

MLW

SLS

V

ALL

10

3

7

2

4

2

2

20

20

7

14

4

7

4

4

40

50

28

38

38

1

12

10

108

100

53

83

83

12

25

16

205

Notes: IEA = Ituri-Epulu-Aru, LT-LT = Lac Télé-Lac Tumba, MTKB = Maiko-Tayna-Kahuzi-Biega, MLW = Maringa-Lopori-Wamba, SLS = Salonga-Lukenie-Sankuru, and V = Virunga Source: WRI Authors

good overall model performance. ROC statistic scores in both landscape groupings indicate strong clustering of observed loss in high-risk areas and are high for the range of scores found in other studies (Pontius and Schneider 2001). Prediction maps also had good overall accuracy and kappa scores for prediction maps and are comparable to other spatial modeling studies (Calijuri et al. 2015; Mas 2004); however, some false positive and negatives occurred in both landscape groupings. In Western Landscapes, false positives occurred in the rural complex along the southwestern boarder of Salonga-Lukenie-Sankuru Landscape and in the rural complex south of Bolobo in Lac Télé-Lac Tumba Landscape. False negatives occurred in eastern rural complex areas of Maringa-Lopori-Wamba Landscape, and rural complex areas between and south of Salonga National Park in Salonga-Lukenie-Sankuru Landscape. In Eastern Landscapes, areas of false positives occurred along National Road 4 and the connector road between Mambasa and Mungere in Ituri-Epulu-Aru Landscape. False positives also occurred in the rural complex around the settlements of Lubutu, Osokari, and Kalole in Maiko-Tayna-Kahuzi-Biega Landscape. False negatives occurred in the rural complex southwest of Kahuzi-Biega National Park and along National Road 3 in Maiko-Tayna-Kahuzi-Biega Landscape.

Policy Implications The predictions of future forest loss patterns provided here can assist policy and decision-making agencies within the CARPE program to identify forest conservation targets and demonstrate the need for further incentives to reduce greenhouse gas emissions from forest loss. Under the Paris Agreement, DRC’s nationally determined contribution calls for a 17 percent reduction in greenhouse gas emissions by 2030 as compared to a business-as-usual scenario, or a reduction of just over 70 million metric tons of CO2 emissions. To meet its emission reduction goals, DRC outlines a target to reforest approximately 3 million hectares of forest by 2025 as part of afforestation/ reforestation programs that are estimated to sequester approximately 3 million metric tons of CO2. Emissions from land-use change, primarily from tropical forest loss, account for approximately 12 percent of greenhouse gas emissions globally (Tyukavina et al. 2015) and about 80 percent of DRC’s total emissions (CAIT). Results from this study indicate that increased conservation efforts, particularly in Ituri-Epulu-Aru and Lac Télé-Lac Tumba Landscapes, can help mitigate future greenhouse gas emissions, because almost 60 percent of future loss from CARPE Landscapes is

TECHNICAL NOTE | June 2017 | 15

expected to occur in these landscapes. According to model predictions, a reduction in the forest loss rate of just 10 percent would lead to emission reductions of 20 million metric tons of CO2 by 2025 (Table 6). This represents emission reductions that are up to six times higher than the expected sequestration resulting from reforestation activities as outlined in DRC’s NDC. Future forest loss and associated carbon emissions from new road and other infrastructure projects could be calculated to enact counteracting measures and ensure DRC’s nationally determined contribution goals are met. Our model results provide the basis for scenario analysis, in which impacts of various land use planning activities on forest loss and carbon emissions can be evaluated against the business-as-usual scenario outlined here. The impact of new roads, settlements, or other infrastructure changes can be assessed to mitigate and reduce forest loss, particularly in protected or other high-value areas. The implications of increased law enforcement or new protected areas can be better understood by seeing where forest loss pressure is likely to shift in the landscapes due to forest protection.

Recommendations for Future Work Future work should be undertaken to incorporate additional datasets and improved reference maps. In addition to nonforest and forest, other land use types should be incorporated to understand the more complex dynamics of land use change. Additional work can also be done to better comprehend and model regional drivers of forest loss. CARPE partners and other local actors should be involved in future iterations of this analysis to incorporate their feedback on additional datasets and to create practical planning scenarios that are based in on-the-ground realities.

16 |

CONCLUSIONS The use of spatial modeling can help to estimate future land cover changes based on the relationship between past land cover change trends and selected variables. In DRC’s CARPE Landscapes, 332,200 hectares of forest are estimated to be lost and with an associated 205 million metric tons of CO2 emitted to the atmosphere by 2025 under a business-as-usual scenario. These predictions can assist policy and decision-making agencies involved with the CARPE program to identify forest conservation targets and demonstrate the need for incentives to reduce greenhouse gas emissions from forest loss. Future work should be undertaken to incorporate additional datasets and improved reference maps, as well as feedback from CARPE partners and other actors in the forest sector.

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

ANNEX 1. MODEL DESCRIPTION IDRISI’s Land Change Modeler (LCM) uses three steps to create forest loss risk and prediction maps:

▪▪ ▪▪ ▪▪

Analyze historical change using two land cover maps for different years. Relate past land cover change information to maps of environmental, biophysical, and socioeconomic variables that might drive or explain such changes and produce a map of forest loss risk. Model change to a future date using this historical change and transition information.

Within the LCM software package, the multi-layer perceptrons (MLP) option was selected to create a spatial model of forest to nonforest risk, expressing for each pixel the probability to change from forest to nonforest. MLP uses artificial neural network (ANN) architecture to identify complex interactions among driver variables. ANNs are nonlinear mapping structures based on the functions of a human brain. Unlike statistical methods, such as logistic regression, which become invalid with the use of correlated variables, ANNs can handle correlated driver variables, data redundancy, and even poor or noisy data well (Haykin 1994; Li and Yeh 2002; Mas 2004). This capability is a major advantage for assessing relationships between deforestation drivers, which are frequently nonlinear and complex.

is repeated until the model can produce a predicted output that closely resembles the training data. Once constructed, the model is used to produce a forest loss risk map, with each pixel representing a combination of driver variables that are more like places where the transition occurred during the calibration period compared to places where the transition did not occur, or the probability that it will be converted. Details on MLP, and artificial neural networks more generally, can be found in Mas (2004) and Ingram et al. (2005). The default parameter values within the LCM software were left unchanged, including 10,000 model iterations and 5,000 training and 5,000 testing pixels (30-meter resolution). Automatic training and dynamic learning parameters were selected. The forest loss risk maps were calibrated using the Hansen et al. (2013)–based forest cover reference maps for 2000 and 2007, leaving the 2008–2014 period for an independent data set used for model validation. Outputs of the modeling process include a soft prediction map, depicting probability of change, as well as a hard prediction map of the expected location of forecasted land use change. While forest loss risk maps are ideal for identifying areas of high risk, prediction maps show the potential consequences of following a businessas-usual approach to land use decisions, including expected forest loss and associated carbon emissions.

MLP is a layered feed-forward ANN, which arranges nonlinear elements in successive layers and flows information unidirectionally from the input layer, through hidden layers, to the output layer. Using a Backpropagation algorithm (Bishop 1995), the system “learns” by predicting output data based on patterns observed from a set of training data. Randomly selected training data from two dates of land cover maps are used to evaluate predictive output layers and adjust weights within the network. This adaptive learning process

TECHNICAL NOTE | June 2017 | 17

ENDNOTES

REFERENCES

1. Rates calculated at a 30 percent forest canopy threshold. 2. Globalforestwatch.org

Ali, R., J. Nash, J. Russ, R. Damania, A. Barra, C. Berg, and Agence française de développement. 2015. “Highways to Success or Byways to Waste: Estimating the Economic Benefits of Roads in Africa.” Africa Development Forum. Washington, DC: World Bank Publications. Aragao, L., Y. Malhi, N. Barbier, A. Lima, Y. Shimabukuro, L. Anderson, and S. Saatchi. 2008. “Interactions between Rainfall, Deforestation and Fires during Recent Years in the Brazilian Amazonia.” Philosophical Transactions of the Royal Society B: Biological Sciences 363 (1498): 1779–85. Bax, V., W. Francesconi, and M. Quintero. 2016. “Spatial Modeling of Deforestation Processes in the Central Peruvian Amazon.” Journal for Nature Conservation 29 (February): 79–88. Bishop, C.M. 1995. Neural Networks for Pattern Recognition. Oxford: Oxford University Press, 482. Bradshaw, C.J., N.S. Sodhi, B.W. Brook. 2009. “Tropical Turmoil: A Biodiversity Tragedy in Progress.” Frontiers in Ecology and the Environment 7 (2): 79-87. Butsic, V., M. Baumann, A. Shortland, S. Walker, and T. Kuemmerle. 2015. “Conservation and Conflict in the Democratic Republic of Congo: The Impacts of Warfare, Mining, and Protected Areas on Deforestation.” Biological Conservation 191 (November): 266–73. CAIT (Climate Data Explorer). 2017. Washington, DC: World Resources Institute. Available at: http://cait.wri.org/. Calijuri, M., J. Castro, L. Costa, P. Assemany, J. Alves. 2015. “Impact of Land Use/Land Cover Changes on Water Quality and Hydrologcal Behavior of an Agricultural Subwatershed.” Environmental Earth Sciences 74 (6): 5373-82. CBFP (Congo Basin Forest Partnership). 2006. “The Forests of the Congo Basin: State of the Forest.” http://carpe.umd.edu/Documents/2006/THE_ FORESTS_OF_THE_CONGO_BASIN_State_of_the_Forest_2006.pdf. de Wasseige C., M. Tadoum, R. Eba’a Atyi, and C. Doumenge. 2015. The Forests of the Congo Basin—Forests and Climage Change. Belgium: Weyrich, 128. Duong, D. and Y. Murayama. 2010. “Forecasting Areas Vulnerable to Forest Conversion in the Tam Dao National Park Region, Vietnam.” Remote Sensing 2 (5): 1249–72. Fuller, D., M. Hardiono, and E. Meijaard. 2011. “Deforestation Projections for Carbon-Rich Peat Swamp Forests of Central Kalimantan, Indonesia.” Environmental Management 48 (3): 436–47. Geist, H. and E. Lambin. 2001. “What Drives Tropical Deforestation? A Meta-analysis of Proximate and Underlying Causes of Deforestation Based on Subnational Case Study Evidence.” AGRIS, EBSCOhost. Haykin, S. 1994. Neural Networks. New York: Macmillan, 936.

18 |

Predicting Future Forest Loss in the Democratic Republic of the Congo’s CARPE Landscapes

Hansen, M, P.V. Potapov, R. Moore, M. Hancher, S.A. Turubanova, A. Tyukavina, D. Thau, S.V. Stehman, S.J. Goetz, et al. 2013. “High-Resolution Global Maps of 21st-Century Forest Cover Change.” Science 342 (6160): 850–53.

Rolland, J., F.L. Condamine, F. Jiguet, and H. Morton. 2014. “Faster Speciation and Reduced Extinction in the Tropics Contribute to the Mammalian Latitudinal Diversity Gradient.” PloS Biology 12 (1): e1001775.

Ingram, J., T. Dawson, R. Whittaker. 2005. “Mapping Tropical Forest Structure in Southeastern Madagascar Using Remote Sensing and Artificial Neural Networks.” Remote Sensing of the Environment 94 (4): 491-507.

Shapiro, A., N. Aguilar-Amuchastegui, P. Hostert, and J.F. Bastin. 2016. “Using Fragmentation to Assess Degradation of Forest Edges in Democratic Republic of Congo.” Carbon Balance and Management 11 (1): 1–15.

IUCN and UNEP-WCMC. 2016. The World Database on Protected Areas (WDPA) [On-line], Cambridge, UK: UNEP-WCMC. www.protectedplanet.net.

Silva, R. da, M. Batistella, and E. Moran. 2016. “Drivers of Land Change: Human-Environment Interactions and the Atlantic Forest Transition in the Paraíba Valley, Brazil.” Land Use Policy 58 (December): 133–44.

Li, M., A. De Pinto, J. Ulimwengu, L. You, and R. Robertson. 2015. “Impacts of Road Expansion on Deforestation and Biological Carbon Loss in the Democratic Republic of Congo.” Environmental and Resource Economics 60 (3): 433–69. Li, X., and A. Yeh. 2002. “Neural-Network-Based Cellular Automata for Simulating Multiple Land Use Changes Using GIS.” International Journal of Geographical Information Systems 4: 323. Mas, J. 2004. “Modelling Deforestation Using GIS and Artificial Neural Networks.” Environmental Modelling & Software 19 (5): 461–71. Megevand, C., A. Mosnier, J. Hourticq, K. Sanders, N. Doetinchem, and C. Streck. 2013. Deforestation Trends in the Congo Basin: Reconciling Economic Growth and Forest Protection. The World Bank. http://elibrary.worldbank.org/ doi/book/10.1596/978-0-8213-9742-8.

Tyukavina A., A. Baccini, M. Hansen, P. Potapov, S. Stehman, R. Houghton, A. Krylov, S. Turubanova, S. Goetz. 2015. “Aboveground Carbon Loss in Natural and Managed Tropical Forests from 2000 to 2012.” Environmental Research Letters 10 (7): 074002. Zarin, D., N. Harris, A. Baccini, D. Aksenov, M. Hansen, C. Azevedo-Ramos, T. Azevedo, et al. 2016. “Can Carbon Emissions from Tropical Deforestation Drop by 50 percent in 5 Years?” Global Change Biology 22 (4): 1336–47. Zhang, Q., C. Justice, M. Jiang, J. Brunner, and D. Wilkie. 2006. “A GIS-Based Assessment on the Vulnerability and Future Extent of the Tropical Forests of the Congo Basin.” Environmental Monitoring and Assessment 114 (1–3): 107–21.

Molinario, G., M. Hansen, and P. Potapov. 2015. “Forest Cover Dynamics of Shifting Cultivation in the Democratic Republic of Congo: A Remote Sensing–Based Assessment for 2000–2010.” Environmental Research Letters 10 (9): 94009. Mon, M., N. Mizoue, N. Htun, T. Kajisa, and S. Yoshida. 2012. “Factors Affecting Deforestation and Forest Degradation in Selectively Logged Production Forest: A Case Study in Myanmar.” Forest Ecology and Management 267 (March): 190–98. Moone, P., B. Verbist, J. Schaepherders, M. Meyi, A. Van Ropaey, B. Muys. 2016. “Actor-Based Identification of Deforestation Drivers Paves the Road to Effective REDD+ in DR Congo.” Land Use Policy 58: 123–32. Nackoney, J., G. Molinario, P. Potapov, S. Turubanova, M. Hansen, and T. Furuichi. 2014. “Impacts of Civil Conflict on Primary Forest Habitat in Northern Democratic Republic of the Congo, 1990–2010.” Biological Conservation 170 (February): 321–28. Pontius, R. and J. Malanson. 2005. “Comparison of the Structure and Accuracy of Two Land Change Models.” Internatioanl Journal of Geographical Information Science 19 (6): 243-65. Pontius, R. and L. Schneider. 2001. “Land-Cover Change Model Validation by an ROC Method for the Ipswich Watershed, Massachusetts, USA.” Agriculture, Ecosystems & Environment 85 (1): 239–48.

TECHNICAL NOTE | June 2017 | 19

ACKNOWLEDGMENTS

ABOUT WRI

The authors thank Naikoa Aguilar-Amuchastegui, Caio de Araujo Barbosa, Tom Evans, Free de Koning, and Giuseppe Molinario for their reviews and helpful comments, which improved the content of this paper. Thanks also to Carin Hall, Maria Hart, Mary Paden, and Jenna Park for providing administrative, editing, and design support. Funding from the U.S. Agency for International Development through the Central Africa Regional Program for the Environment, and from the Norwegian government made this analysis possible.

World Resources Institute is a global research organization that turns big ideas into action at the nexus of environment, economic opportunity, and human well-being.

ABOUT THE AUTHOR Elizabeth Goldman is a GIS Research Associate with Global Forest Watch at the World Resources Institute, USA. Contact: [email protected] Nancy Harris is the Research Manager with Global Forest Watch at the World Resources Institute, USA. Contact: [email protected] Thomas Maschler is the Technical Lead for the African Forests Program at the World Resources Institute, USA. Contact: [email protected]

Our Challenge Natural resources are at the foundation of economic opportunity and human well-being. But today, we are depleting Earth’s resources at rates that are not sustainable, endangering economies and people’s lives. People depend on clean water, fertile land, healthy forests, and a stable climate. Livable cities and clean energy are essential for a sustainable planet. We must address these urgent, global challenges this decade. Our Vision We envision an equitable and prosperous planet driven by the wise management of natural resources. We aspire to create a world where the actions of government, business, and communities combine to eliminate poverty and sustain the natural environment for all people. Our Approach COUNT IT We start with data. We conduct independent research and draw on the latest technology to develop new insights and recommendations. Our rigorous analysis identifies risks, unveils opportunities, and informs smart strategies. We focus our efforts on influential and emerging economies where the future of sustainability will be determined. CHANGE IT We use our research to influence government policies, business strategies, and civil society action. We test projects with communities, companies, and government agencies to build a strong evidence base. Then, we work with partners to deliver change on the ground that alleviates poverty and strengthens society. We hold ourselves accountable to ensure our outcomes will be bold and enduring. SCALE IT We don’t think small. Once tested, we work with partners to adopt and expand our efforts regionally and globally. We engage with decision-makers to carry out our ideas and elevate our impact. We measure success through government and business actions that improve people’s lives and sustain a healthy environment.

Maps are for illustrative purposes and do not imply the expression of any opinion on the part of WRI, concerning the legal status of any country or territory or concerning the delimitation of frontiers or boundaries.

Copyright 2017 World Resources Institute. This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of the license, visit http://creativecommons.org/licenses/by/4.0/

10 G Street, NE | Washington, DC 20002 | www.WRI.org