Modeling deforestation to REDD+ Project: a case study in Alto Mayo Protected Forest, San Martin Region, Peru

Transcription

1 Modeling deforestation to REDD+ Project: a case study in Alto Mayo Protected Forest, San Martin Region, Peru Fabiano Luiz de Oliveira Godoy 1 Eddy Hoover Mendoza Rojas 2 1 Conservation International - CI 2011 Crystal City Drive, Suite 500 Arlington, VA, fgodoy@conservation.org 2 Conservation International Peru Office CI-Peru Avenida Dos de Mayo Miraflores - Lima, Peru emendoza@conservation.org Abstract. Alto Mayo Protected Forest, a natural protected area in the Peruvian amazon, is recognized for its importance for biodiversity and environmental systems but also is highly threated to rapid conversion of forest land to coffee plantation and illegal logging. The conservation of Alto Mayo forests is crucial to maintain the provision of environmental services to local population as well global community though maintaining the natural carbon cycle and avoiding greenhouse gas emission to the atmosphere. Reducing (greenhouse gases) Emissions from Deforestation and Degradation (REDD+) is being established as a mechanism to mitigate climate change as well as finance framework of conservation projects. In this study, we classified satellite images and analyzed the historical deforestation rates for the period of , developed a model to predicted the rates of forest loss based on the major driver of deforestation, and using neural networks developed a land change model from IDRISI to predict the location of future deforestation. Potential scenarios demonstrated increasing deforestation pressure mainly in northern region of protected area. The results showed that this approach can be considered as a reliable alternative to estimate the potential forest loss in Alto Mayo and a credible baseline for REDD+ project. Palavras-chave: remote sensing, deforestation, REDD, protected area, peruvian amazon, sensoriamento remoto, desmatamento, area protegida. 1. Introduction The Alto Mayo Protected Forest (AMPF) covers approximately 182,000 hectares of land in the Peruvian Amazon of extremely high value for biodiversity conservation and watershed protection (Figure 1). Its forests are also recognized for their importance in providing clean and abundant water supplies, preventing soil erosion, protecting soils in the lowland areas from torrential flows and floods, and for their scenic beauty (INRENA, 2008). The Alto Mayo forests also store a significant amount of carbon, whose release in the atmosphere through deforestation results in the emission of large quantities of greenhouse gases which contribute to climate change. Alto Mayo forests are critical for global climate change mitigation, biodiversity conservation, and the provision of ecosystem services to the local population. Despite of its environmental importance and the legal designation as Protected Area, AMPF has one of the highest rates of deforestation rate in Peru. The threats to the area have increased in the last decade with the linking of the infrastructure development projects and the mainly by the conversion of forest to unsustainable coffee plantation. In response, since 2008 Conservation International and its local partners designed and implemented a Reduced Emission from Deforestation and Degradation (REDD+) project, whose main goal is to promote the sustainable management of the AMPF and its ecosystem services for the benefit of the local populations and the global climate. The financial mechanism of this project is based on the avoided greenhouse gas (GHG) emissions that would be released to the atmosphere through the deforestation in the absence of the project (baseline scenario) and the actual deforestation that has happened in the same period. In this 7233

2 study remote sensing and GIS processing were used to project the deforestation in AMPF for the period of and estimated the potential forest loss in the project area in the absence of any activity to mitigate the deforestation. Figure 1. Location of the Alto Mayo Protected Forest, established in Data and methods 2.1 Remote Sensing Data An analysis of land-use and land-cover change of the AMPF and surrounded area was conducted for the reference period ( ) using medium resolution satellite imagery and validated using high-resolution satellite images. All data sources used in these analyses are listed on Table 1. Table 1. Data used for historical LU/LC change analysis. Resolution Coverage Acquisition date Satellite Sensor Spatial Spectral (km 2 ) (DD/MM/YY) Landsat 5 TM 30m µm 185 x 172 km 8-Jun-96 Landsat 7 ETM+ 30m µm 185 x 172 km 30-Jun-01 Landsat 5 TM 30m µm 185 x 172 km 8-Sep-06 Landsat 5 TM 30m µm 185 x 172 km 4-Sep-96 Landsat 7 ETM+ 30m µm 185 x 172 km 24-Aug-01 Landsat 5 TM 30m µm 185 x 172 km 15-Sep-06 Landsat 5 TM 30m µm 185 x 172 km 25-Aug

3 Landsat 5 TM 30m µm 186 x 172 km 8-Sep-06 CBERS CBERS-2 2.5m µm 113 x 113 km 6-Jul-08 CBERS CBERS-2 2.5m µm 113 x 113 km 27-Aug-08 RapidEye RapidEye 5m µm 77 x 77 km 21-Set-10 RapidEye RapidEye 5m µm 77 x 77 km 21-Set-10 RapidEye RapidEye 5m µm 77 x 77 km 21-Set-10 RapidEye RapidEye 5m µm 77 x 77 km 21-Set Map of land-use and land-cover classes Land-cover change data for the reference region were mapped, via analysis of Landsat-5 and Landsat-7, for the period of 1996 to The map was further refined and updated using additional imagery acquired in circa 2001, and circa 2006 to create a multi-temporal map with minimal cloud cover. The result is a map with five classes including forest cover and loss, non-forest, cloud and water. Forest types were further characterized by several parameters: elevation above sea level, number of trees per hectare, canopy density, topographic position and minimum mapping unit. Elevation turned out to be the main characteristic for distinguishing between forest types. Forest land was therefore stratified by elevation, namely pre-montane forest, found below 1200 mabsl; cloud forest, located between 1200 and 2500 mabsl; and dwarf forest with shorter vegetation found above 2500 mabsl. The multi-temporal analysis of forest cover and loss has been filtered to a MMU of 2 ha, thus yielding a conservative estimate of forest area, and a conservative estimate of the area in which forest loss could potentially be recorded. One broad class of non-forest land use was used due to the high uncertainty in distinguishing areas covered by each of the non-forest classes present in the reference region (i.e. coffee plantation, pastures and fallow). 2.3 Analysis of agents, drivers and underlying causes of deforestation. The agents, drivers, and underlying causes of deforestation in the project area were conducted to understand the spatial variables that could explain the geographical distribution of deforestation (e.g. proximity to urban centers or commodity markets) and the future likely trend. The agents and drivers of deforestation were identified from the expert opinions gathered through a revision of socio-economic studies, interviews with local experts (such as park guards, government officials and community leaders), and was completed by a participatory workshop following the Open Standards for the Practice of Conservation methodology. Based on these sources, seven agent groups were identified. Coffee producers represent the main agent group responsible for deforestation in the AMPF region, followed by a number of less significant agent groups including cattle farmers, subsistence farmers, local politicians promoting the illegal construction of infrastructure, illegal loggers and timber merchants, land traffickers, and firewood collectors (AIDER, 2011). The following spatial data, available in digital format, were used to represent these drivers and the underlying causes (e.g. coffee production is mainly cultivated in low slopes, medium altitude, near to market centers): Access to the forest: - Settlements: a Euclidean distance map was created based on proximity to the nearest settlement, subdivided into capitals, towns and villages; - Rivers: a Euclidean distance map was created based on proximity to the nearest rivers, subdivided into primary and secondary rivers; 7235

4 - Roads: a Euclidean distance map was created based on proximity to the nearest roadways, subdivided into major roads, secondary roads and trails; Terrain: - Elevation: obtained from the Shuttle RADAR Topography Mission (SRTM) of NASA, with a 90-meter horizontal resolution and 10-meter vertical resolution; - Slope: derived from the SRTM dataset; Factor maps for discrete data, like soil type and administrative units, were generated by evidence likelihood analysis. The evidence likelihood analysis estimate the probability of deforestation in each unit (polygon) based in the historical deforestation and uses an index of deforestation 2.4 Projection of future deforestation Projection of the annual areas of baseline deforestation Forest conversion to coffee plantation is the major driver of deforestation in the AMPF. As the results confirmed deforestation is directly related to coffee production. The conventional coffee production techniques used by the majority of coffee producers are highly unsustainable and after a few years of production the yields decrease substantially. Subsequently, new forested areas are converted to coffee plantations while some of the old coffee fields become pasture land (GORESAM, 2011). Therefore annual areas of deforestation can be projected by correlating coffee production and observed deforestation in the historical period. Data on annual coffee production in San Martin were gathered per province (Rioja, Moyobamba and Huallaga). The total coffee production per year was available since 1997 (MINAG; 2011) and is shown in Figure 2. The data covers the historical period until 2010 and confirmed an increasing pattern in production (R2 = 0.86). Figure 2. Annual coffee production in the Rioja, Moyobamba and Huallaga provinces. This data on coffee production were then grouped in two periods ( and ) in accordance with the historical deforestation periods used in the forest cover and loss analysis. An increasing trend with direct relationship is observed between coffee production and deforestation (Figure3) across provinces and periods (R2 = ). Thus the future annual rate of deforestation (in ha) was a function of coffee production per year (Y = * ( x ) ) 7236

5 Figure 3. Deforestation as function of coffee production Projection of the location of future deforestation Several software packages exist that enable spatially explicit modeling of future land use change. The VM0015 methodology (VCS, 2012) references the GeoMod model, which is available as part of the IDRISI software for geographical analysis, however a newer and more robust model, the Land Change Modeler (LCM), is available in the more recent editions of IDRISI. We selected LCM for its relative ease of use and non-reliance on independence among driver variables, as it is based on a neural network rather than on multiple regression analysis (Clark Labs, 2010). LCM uses a neural network analysis to create a land use model and predict the location of future deforestation based on the correlation between the drivers of deforestation (variables) and the observed forest loss (Eastman, 2009). Following this, the model produced a risk map of deforestation, scaled from 0 to 1, where 0 are with lowest risk of deforestation and 1 are areas with the highest risk of deforestation. For the geographical analysis, the model was calibrated using observed data on deforestation from 1996 to 2001, and then validated using the projected deforestation against the observed deforestation for The actual rate of deforestation between 2001 and 2006 was then assigned to the model to predict the deforestation location in Therefore only the distribution of deforestation was projected and tested, and not the amount. The resultant change map ( ) was confirmed with the actual change map produced, by overlapping both maps in GIS. Figure of Merit (FOM) was used to estimate the accuracy of the model. Kim (2010) evaluated the predictive accuracy of the neural network models and found that FOM produces similar results to other statistical methods and provides to be a credible statistical assessment alternative. 3. Results 3.1 Analysis of historical land-use and land-cover change Landsat imagery was used in this project to map the forest cover and loss. For the validation process, high resolution images were used (Table 2). To estimate forest loss change rates were calculated, in percentage per year, for areas that were cloud-free in both time periods, and within the reference region (Table 2). The rate increased from 0.12% y-1 for the period , to 0.36% y-1 between The areas covered by cloud represent less than 0.4% of the reference region. 7237

6 Table 2. Forest cover and loss in forest 1996 forest 2001 forest 2006 change change change ha ha ha ha/y ha/y ha/y Pre-montane 2,301 2,283 2, cloud forest 398, , , , dwarf forest 87,545 87,467 84, reference region 488, , , ,762 1,182 The 2006 land cover classification was validated by visually inspecting a set of 200 randomly generated points within the reference region, against a set of RapidEye and CBERS scenes. The overall accuracy was 95%, while the accuracy of each land cover and land use change was above 90%. 3.2 Risk map of deforestation and projected land use maps The technique assessment - Figure of Merit at polygon level (FOM) - was applied to assess the accuracy of the model in each forest stratum. FOM ranges from 0%, where there is no overlap between observed and predicted change, to 100%, where there is a perfect overlap between observed and predicted change. Within the project area, the estimated FOM was 61% overall, and 54% within the cloud forest areas, which cover almost 95% of the project area (Table 3). The FOM within premontane and dwarf forest were less than 50%, primarily because those classes cover a small area and are located in lower deforestation risk areas. Table 3. Figure of Merit at polygon level. Polygon total area of changes (ha) correct (ha) False alarms (ha) misses (ha) FOM (%) Pre-montane forest Cloud forest 5, ,828 1, Dwarf forest Project area 5, ,964 1, Below is the final risk map. The colors in the map describe the risk of deforestation, with dark red representing high potential, decreasing until low potential represented by dark blue. The left panel (Figure 4) shows the parameter used in the model, including start and end leaning rates and final running statistics. The prediction of deforestation during the project crediting period requires a forest benchmark map for the project start date, The rates estimated in the section 2.4 were applied to the risk map to determine the location of future baseline deforestation maps ( ). Future deforestation is assumed to happen first at the pixel location with the highest deforestation risk value (Figure 5). After the completion of this step, the project area were 7238

7 overlaid using GIS to estimate the quantity of deforestation that will happen in the baseline scenario within those boundaries. Figure 4. Deforestation risk map and LCM parameters. Projected Land Cover 2009 Projected Land Cover 2018 Deforestation Figure 5. Projected Land Cover Maps and the total deforestation projected for the period of Conclusions In this study we analyzed the multitemporal dynamics of deforestation in a natural protected area mainly due to conversion to coffee production and migration from other regions to Alto Mayo valley. The land change model developed using neural networks proofed to be adequate framework to predict forest changes in the area of study. The results demonstrate that this area has a high pressure of deforestation and it needs to strengthen the management and control and monitoring actions in order to avoid forest loss. 7239

8 The deforestation rate estimated based on coffee production is pertinent to the reality. Although deforestation is directly correlated to coffee production, an increase of coffee production could not be explained by an increase in productivity, as the production per hectare did not increase substantially. Rather, it is a result of an expansion in the area cultivated with coffee. Between 2001 and 2010, the area occupied by coffee plantations in San Martin increased almost 115% (36,162 ha). However, the total yield showed a similar trend with an increase of 103% or 26,817 tons. Thus, productivity increased only by 4% in the same period, confirming our assumption that increase of coffee production in the region is directly linked with new areas converted to coffee plantations. Empirical studies suggested that a minimal FOM threshold would be set as 50% for frontier type of deforestation (Pontius et al. 2000; 2007; 2008), where forest conversion happens in a front of deforestation in a still huge area of forest. Our results of 61% FOM gives the reliability that the land use model generated is accurate in predicting the areas with high risk to be deforested. 5. References Asociación para la Investigación y Desarrollo Integral (AIDER). Agentes, conductores y causas subyacentes de la deforestación en el Bosque De Protección Alto Mayo (BPAM) y zona de amortiguamiento. Documento de trabajo. Lima, Clark Labs. Modeling REDD Baselines using IDRISI s Land Change Modeler. IDRISI Focus Paper Eastman, R. IDRISI Taiga Guide to GIS and Image Processing. Clark Labs, Clark University, Worcester, MA: IDRISI Production, p. Gobierno Regional de San Martin (GORESAM). Serie Historica de Superficies Existentes, Cosecha y Producción, Available at: < Accessed on: January, Harper, G.; Steininger, M.; Tucker, C.; Juhn, D.; Hawkins, F. Fifty years of deforestation and forest fragmentation in Madagascar. Environmental Conservation, v. 34, p , Instituto Nacional de Recursos Naturales (INRENA). Plan Maestro del Bosque de Protección Alto Mayo Lima, p. Kim, O. S. An Assessment of Deforestation Models for Reducing Emissions from Deforestation and Forest Degradation (REDD). Transactions in GIS, v.14, n. 5, p , Ministerio de Agricultura (MINAG). Series Históricas de Producción Afrícola Compendio Estadístico, Available at: < Accessed on: December, Pontius, R. G. Jr. Quantification error versus location error in comparison of categorical maps. Photogrammetric Engineering and Remote Sensing, v. 66, n. 8, p , Pontius, R. G. Jr.; Boersma, W.; Castella, J.; Clarke, K.; de Nijs, T.; Dietzel, C.; Duan, Z.; Fotsing, E.; Goldstein, N.; Kok, K.; Koomen, E.; Lippitt, C.; McConnell, W.; Mohd Sood, A.; Pijanowski, B.; Pithadia, S.; Sweeney, S.; Trung, T.; Veldkamp, A.; Verburg, P. Comparing input, output, and validation maps for several models of land change. Annals of Regional Science, v. 42, n. 1, p , Pontius, R. G. Jr.; Walker, R.; Yao-Kumah, R.; Arima, E.; Aldrich, S.; Caldas, M.; Vergara, D. Accuracy assessment for a simulation model of Amazonian deforestation. Annals of Association of American Geographers, v. 97, n. 4, p , Verified Carbon Standard (VCS). Methodology for Avoided Unplanned Deforestation VM0015, Version 1.0. Sectoral Scope p. 7240