Remote sensing of complex land use change trajectories—a case study from the highlands of Madagascar

Agriculture, Ecosystems and Environment 115 (2006) 219–228 www.elsevier.com/locate/agee

Remote sensing of complex land use change trajectories—a case study from the highlands of Madagascar Tor-Gunnar Va˚gen * Norwegian Centre for Soil and Environmental Research, Norwegian University of Life Sciences, Department of Plant and Environmental Research, Jordforsk, Frederik A. Dahls vei 20, 1432 A˚s, Norway Received 17 September 2004; received in revised form 14 December 2005; accepted 19 January 2006 Available online 3 March 2006

Abstract Madagascar is often portrayed as a global environmental hotspot with widespread deforestation and environmental degradation. Quantitative and spatially explicit data on ecological change are, however, scarce and current estimates are often based on simplistic representations of deforestation and land use change. Significant uncertainties in current estimates therefore remain. The present study was conducted to assess deforestation and other important complex land use change trajectories in the eastern highlands of Madagascar. A timeseries of satellite imagery dating from 1972 to 2001 was used to analyse overall change and rates of change between different land use types in the study area. Forest cover in the study area was approximately 8060 ha in 1972 and 4278 ha in 2001. Rates of deforestation were not, however, constant throughout this period, but varied from 52 ha yr
1 (1972–1992) and 341 ha yr
1 (1992–1999). The increased rates in the 1990s were attributed to turbulent political conditions on the island during the latter period and shows the complex relationships between social, political and ecological processes governing deforestation and land use change processes. Accessibility (distance to villages and roads) and elevation were shown to be the most important predictors of deforestation risk in the study area. Intensive cultivation of slopes (tanety) increased by about 3400 ha (  65%) during the study period, a significant part of which came from cultivation of grassland savanna (net increase  1700 ha). These trends were found to be indicative of increasing pressure on available land resources in the region, leading to extensive cultivation of marginal grasslands and ultimately significant soil fertility decline. # 2006 Elsevier B.V. All rights reserved. Keywords: Madagascar; Deforestation; Land use change; Remote sensing; Image classification

1. Introduction There are significant uncertainties in current estimates of deforestation and land use change on Madagascar, mainly due to simplistic representations of deforestation and ecological change and a general lack of quantitative, spatially explicit and statistically representative data on change in land cover and -use. Important confounding effects in current estimates include the definitions used for deforestation and the scale at which deforestation assessments are made, as well as types of forest under consideration. There are often subtle differences between land cover modifications and conversions (deforestation) * Tel.: +47 64948100; fax: +47 64948110. E-mail address: [email protected]. 0167-8809/$ – see front matter # 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.agee.2006.01.007

due to ecosystem resilience, inter-annual variability and complex land-cover change trajectories (Ringrose et al., 1990; Lambin, 1999) making land cover change estimates uncertain if for instance the spatial scale of the imagery used does not match that of the processes studied or an insufficient number of imagery is used (i.e. temporal scale is inappropriate). Several studies have also been based on extrapolations of earlier estimates without proper ground surveys, resulting in propagation of errors from study to study. The majority of available studies have attempted to assess change in natural forest cover, but very few studies have been conducted to assess other and more detailed types of land use change. Despite evidence to the contrary, many authors on deforestation in Madagascar present the highlands as having been completely covered by evergreen forest prior to

220

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

Austronesian settlement on the island between A.D. 300 and A.D. 800 (e.g. Gade, 1996), and this is generally held up as a fact by environmental organizations working on the island. According to recordings made by early explorers and colonialists, there were grassland areas covering significant parts of the highlands in the mid- to late 1800s and earlier (Grandidier, 1905; Kent, 1970; Kull, 1999). More recent studies based on pollen analysis and radiocarbon dating show that the central highlands were probably never completely covered by forest, but consisted of savanna grassland and forest mosaics (Burney, 1997). There are also evidences that show that fire occurred before human settlement on the island (Burney, 1997). The first attempt at describing and classifying Madagascar’s vegetation was made by Baron (1889) and later added on by Viguier (1914), but the foundation for later vegetation mapping was essentially laid by Perrier de la Baˆthie (1921). He divided the island into two main areas (windward flora and leeward flora) and subdivided these areas into the Eastern, Central and Sambirano regions (Lowry et al., 1997). These regions were later divided more finely by Humbert (1951, 1955), who focused mainly on differentiating native vegetation from anthropogenically modified vegetation types. The first spatial analysis of vegetation types in Madagascar was therefore made by Humbert and Cours Darne (1965), resulting in a 1:1,000,000 scale map based on analysis of aerial photographs taken during the 1950s and 1960s. This work showed that very limited areas of Madagascar actually had vegetation that was still undisturbed by humans. The vegetation map of Madagascar currently used as reference in numerous studies is, however, that of Faramalala (1988) who used Landsat MSS imagery from the 1970s in conjunction with the earlier maps of Humbert and Cours Darne (1965). This map was later published in digital form (Faramalala, 1995), and is currently the basis for vegetation maps issued by the national mapping authority Foiben-Taosarintanin’i Madagasikara (FTM), although it is becoming badly outdated. Green and Sussman (1990) estimated the remaining eastern humid forest cover on Madagascar to 3.8 million hectares (Mha) in 1985, based on a time-series of aerial photographs and Landsat images for the period between 1950 and 1985. This represented a reduction of 7.4 Mha compared to original forest cover and 3.8 Mha compared to forest cover in 1950. Deforestation was reported to be highest in areas with population densities of more than 10 persons per square kilometer (19% remaining forest in 1985 relative to original cover). The population density map used by Green and Sussman (1990) was, however, simplified into only three categories so that urban areas relatively far from forest margins disproportionately influenced estimates. Moreover, the population data used were from the 1966 Atlas de Madagascar at subprefecture level, making the population data, aerial photographs and Landsat MSS/TM data used in the study discordant in time, leading to rather dubitable comparisons. Nelson and Horning (1993) estimated

total forest cover on the island (including eastern humid, and southern and western dry forests) to about 11% (  6.1 Mha) based on Landsat MSS and AVHRR-LAC data, while Mayaux et al. (1999) estimated the dense humid forest cover to approximately 5.5 Mha using the SPOT-4 VEGETATION platform (spatial resolution = 1165 m). Dense humid and dry deciduous forest cover in the latter study was estimated at over 9.5 Mha. The AVHRR platform has been applied in numerous studies of biomass burning, but has been shown to have large problems with cloud contamination in tropical regions due to its afternoon overpass (Perrin and Millington, 1997). The more recent ATSR-2 sensor, which has a similar spatial resolution (  1 km) is therefore commonly used to complement AVHRR due to its morning overpass. The link between shifting cultivation, population growth, deforestation and environmental degradation is often made in debates on tropical deforestation. However, shifting cultivation and population dynamics cannot be abstracted from external processes such as colonial capitalism and imperialism (Jarosz, 1993). Studies suggest that population growth and deforestation in Madagascar were not linked until after the colonial period as national growth rates were at or below replacement levels due to widespread malnutrition, famine and disease (Chevalier, 1952). Yet between 3 and 7 Mha of forest were cleared between 1900 and 1940 Hornac, 1943; Boiteau, 1982, mainly due to logging (of exotic species such as palisander (Dalbergia spp.) and ebony (Diospyros perrieri)) and export crop production (the most important being coffee). Also, colonial policies, including the prohibition of traditional shifting cultivation (tavy) while widespread mining of the forest resources was taking place, further exacerbated the situation. To the Malagasy, tavy was and is an important practice, not only for food production but also for a number of cultural reasons (Kull, 2002). Village institutions have often contributed to preserving natural forest areas (McConnell, 2002), but the actions of corrupt officials seeking short-sighted personal gains often undermine these efforts. Deforestation and land use change in Madagascar is therefore a difficult issue due to complex interactions between social, political, economic and ecosystem processes. Quantifying change in land cover types and land use is important for understanding other ecosystem processes, including change in soil quality and processes leading to soil and environmental degradation. Quantitatively describing the spatial and temporal dynamics of these processes is also important for the development and implementation of targeted improved land management policies and interventions, and for monitoring of ecosystem status and trend. In this study land use dynamics were studied in an area of the eastern highlands of Madagascar based on a time series of satellite images from 1972, 1985, 1992, 1999 and 2001. The main objective was to quantify deforestation and various land cover modifications and transitions as part of a study of change in soil quality in relation to deforestation and land use change.

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

221

Fig. 1. Colour infrared (CIR) composite of Landsat ETM+ scene from study area (coordinates are UTM 38s).

2. Materials and methods

2.1. Geometric correction and pretreatment of imagery

The area studied was located in the eastern highlands of Madagascar, covering approximately 476 km2to the east and north east of the town of Ambositra (Fig. 1). The study was based on a time-series of satellite (Landsat) images from 1972 (Multispectral scanner (MSS)), 1985 and 1992 (Thematic Mapper (TM)), and 1999 and 2001 (Enhanced Thematic Mapper (ETM+)), used in conjunction with aerial photographs and a recent high-resolution EROS-A image (Table 1). The Landsat TM image from 1985 had significant cloud contamination, and only small sections of this image were therefore analysed and used for verification of classification results in the period 1972–1992. The study area is dominated by the Betsileo ethnic group, with agricultural production centered mainly on intensive smallholder rice (paddy) production. Tavy is practiced in the eastern parts of the study area, while a range of different crops are cultivated on sloping lands. The study area is also extensively grazed by local Zebu cattle, and grasslands are periodically burnt in the period between September and November, before the onset of rains. Soils of the study area are predominantly Oxisols, are generally acid (pH between 3.7 and 5.5) and inherently nutrient poor. The western part of the study area is driest with rainfall averaging around 1200–1400 mm yr
1, while the eastern escarpment receives about 2000 mm yr
1.

A satellite differential GPS (DSGPS) was used in field (accuracy  30 cm) to collect georeferenced information for geometric correction of the satellite imagery. Easily recognizable landscape features, roads and major buildings were recorded and used to rectify the image from November 2001 through an affine mapping procedure in TNTMips (Microimages, Inc.). This procedure involves the performance of a least-squares fit to determine the best overall transformation. This corresponds to an orthographic or parallel plane projection from a source XY-plane onto a target XY-plane. The images dating from 1972, 1985, 1992 and 1999 were co-registered with the 2001 image using the latter as reference in combination with DSGPS data and available maps. Co-registering was performed to reduce the residuals as much as possible (to within 5 m), and the older images were finally resampled to match the spatial extents and pixel size of the 2001 Landsat ETM+ image using nearest neighbor resampling to preserve spectral information. Due to differences in atmospheric conditions, sun angle, etc. between image dates, individual images were corrected for these artifacts prior to analysis of change by computing the exoatmospheric reflectance of each image pixel using the following equation:

Table 1 Sensor types and image dates for imagery used in study Sensor type

Date(s)

Spatial resolution (m)

Aerial photograph Landsat MSS Landsat TM Landsat ETM+ EROS-A

1957 and 1991 December 15, 1972 80.00 March 08, 1985 and April 28, 1992 28.50 October 17, 1999 and November 07, 2001 28.50 September 06, 2003 1.94

rp ¼

pLl d 2 ; ESUNl cos us

where Ll is spectral radiance, d the earth–sun distance, ESUNl mean exoatmospheric irradiance (derived from Landsat Technical Notes, 1986) and us is solar zenith angle (degrees). Calibrations were conducted in ENVI version 4.0 (Research Systems Inc.). 2.2. Change detection At the scales sampled by the Landsat sensor, vegetated ecosystems are comprised of mixtures of surface materials,

222

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

including different canopy components, bare soil, water, and shadow, and the spectrum measured by the sensor is therefore a mixture of each of these components (Roberts et al., 1998). A number of methods are available for temporal land use change detection, including: (i) postclassification comparison, (ii) classification of multitemporal data sets, (iii) principal components analysis (PCA), (iv) temporal image differencing and ratioing, (v) change vector analysis and (vi) spectral mixture analysis, to mention some. The main emphasis of the study was on change in natural forest cover (i.e. deforestation) and areas under intensive cultivation. The present paper reports the findings of post-classification comparisons between the image dates included in the study. Ground truth (training) data were collected in field between September and November 2001, while training data from 1972 to 1992 were derived from aerial photographs, maps and directly from the MSS and TM imagery. This information was compared to information from farmers and local experts in the study area. Supervised classification was conducted using Maximum Likelihood (ML) classification, which is one of the most common classification methods used in remote sensing image analysis. The ML classification is based on the estimation of probability distributions for the land cover types in the training data through a quantitative evaluation of category spectral response pattern variance and covariance (Lillesand and Kiefer, 2000). Classification accuracy was estimated for all years by computing confusion (error) matrices between reference data and classification results relative to ground truth data, and overall accuracy and Kappa index values were computed (Cohen, 1960; Congalton, 1991). Supervised classification was conducted in TNTMips version 6.9 (Microimages, Inc.). 2.3. Analytical procedures 2.3.1. Factors determining the rate of deforestation Roads and villages with more than five houses were mapped for the eastern section of the study area based on existing maps, GPS recordings made in field, aerial photographs from 1991 and a high resolution EROS-A satellite image from September 2003. Population data were acquired from the Malagasy national bureau of statistics (Institut Nationale de la Statistique et de la Recherche Economique, INSTAT) for the periods 1960, 1970, 1980, 1990 and 2000, but due to poor spatial representation, these data were not included in the analysis of deforestation trends, but rather served as a basis for assessing the general population trends in the study area. Data on population density from the African Population Database (Deichmann, 1994) were therefore used in conjunction with the INSTAT data and the change in population density between 1960 and 1990 was computed and used as input for the assessment of deforestation trend. Topographical information (elevation and slope) was derived from the Shuttle Radar Topography

Mission (SRTM) data (NASA, JPL) at a horizontal resolution of 90 m (3 arc-s). 2.3.2. Modelling deforestation risk The deforestation pattern was studied using distance to roads and villages, elevation, slope, and change in population density as independent variables. Levels for the dependent variable (deforestation) were selected based on the postclassification comparison of forest cover for different image dates and a multinomial logistic model (Nelder and Wedderburn, 1972; McCullagh and Nelder, 1989) was then applied to predict the probability of deforestation. The multinomial logistic probabilities are given by: Pi j  PðY ¼ jjXi Þ ¼

exp ½b0j Xi  Di

P P where Di ¼ Jj¼1 ½exp ðb0j Xi Þ and b0j Xi represents bk j Xik . The unknown, which is estimated, is b j (see Aldrich and Nelson (1984)). A mask was created prior to fitting the logistic model to restrict model predictions to areas with forest in 1972 (i.e. candidates for conversion). A nested modelling strategy was adopted where the independent variables were progressively separated (Hosmer and Lemeshow, 1989), based on significance tests in the logistic model using Wald tests (Fahrmeir and Tutz, 1994). Maps of predicted risk of deforestation (log-odds) and ordinary residuals (Y
P) from the model predictions were then created and compared to the actual pattern of deforestation. Computations were conducted using the Design and Hmisc libraries (Harrell, 2003) for S-Plus (Insightful, Inc.), while the GIS application of the model was implemented through a customized script in SML for TNTMips (Microimages, Inc.).

3. Results and discussion The major (dominant) land use types in the study area were broadly natural forest, savoka (regrowth, including natural fallow), eucalyptus (Eucalyptus sp.), tanety (dryland cropping on slopes), paddy and grassland savanna. In addition, recently burnt areas were assessed for each image date, including both burnt grassland and natural forest (tavy). According to 1993 census figures, the population in Ambositra was 21,350 (INSTAT), and it was estimated at 28,000 in 2001. Population density increased throughout the study area between 1960 and 1990 (Table 2). 3.1. Accuracy assessment Due to lack of aerial photographs from the early 1970s, the low resolution and relatively poor quality of the MSS sensor, ground truth data for the 1972 image were associated with more undertainty than for the rest of the image dates. This was particularly the case in eucalyptus and savoka areas, which were difficult to differentiate straight from the

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

223

Table 2 Population density estimates for Ambositra and Ranomena for the period 1960–1990

Table 3 Summary of classification performance for four of the image dates, showing ground truth and classification accuracy for each land use type

Year

Land use

Population density (persons km
2)

Ambositra 1960 1970 1980 1990

47 61 63 83

Ranomena 1960 1970 1980 1990

25 34 36 47

Year Accuracy (%) 1972

1992

1999

2001

Natural forest

Ground truth Classification

98.5 99.9

96.6 98.1

99.5 98.6

97.7 99.8

Savoka

Ground truth Classification

78.6 90.9

58.8 93.0

91.8 72.4

97.3 85.1

Eucalyptus

Ground truth Classification

88.1 55.3

78.6 43.8

85.6 93.7

97.7 87.6

Tanety

Ground truth Classification

88.1 95.9

52.3 75.5

81.9 79.8

100.0 98.8

Paddy

Ground truth Classification

93.9 92.8

90.2 74.9

95.8 99.3

99.9 99.9

Grassland

Ground truth Classification

93.2 84.5

91.0 85.3

98.3 83.9

98.3 99.9

Source: Deichmann (1994).

MSS image, as reflected in the accuracy assessment (Table 3). The user accuracy for tanety and grassland savanna was relatively good (i.e. low error of commission), while eucalyptus showed confusion with both savoka and tanety and savoka areas were also inaccurately classified as eucalyptus. Producer’s accuracy was also poorest for eucalyptus due to confusion with forest transition zones and to some extent natural forest. This is partly due to the structure of the eucalyptus vegetation in the study area as it is dense with significant undergrowth near forest margins and boundaries between eucalyptus, savoka and natural forest are therefore gradual in some areas. Also, the low resolution and limited number of wavebands (4) in the MSS sensor makes it difficult to spectrally distinguish between different types of dense green woody vegetation. In other areas eucalyptus vegetation generally had a more open structure (i.e. open woodlands) with grass undergrowth and classification results were generally more accurate, although there was some confusion with grasslands and tanety. Natural forest was classified with a high degree of accuracy, relative to ground truth data (Table 3). The Landsat TM image from 1985 had significant cloud contamination, particularly in eastern (i.e. natural forest) regions of the study area, and despite high classification accuracy, deforestation trends could not be assessed between 1972 and 1985. Areas with excessive cloud cover were excluded from the analysis, while it was suitable for assessing cover of eucalyptus, grassland and paddy. Images from 1985 and 1992 were taken during the latter stages of the rainy season (March and April), contrary to the other image dates, which complicated their classification somewhat as otherwise easily distinguishable land use types (i.e. grassland and tanety) were more difficult to accurately classify (Table 3). Land use classifications for the latter image dates (1999 and 2001) had high degrees of accuracy (Table 3), while natural forest was classified with a high degree of accuracy at all image dates.

Kappa coeff.

0.90

0.81

0.93

0.97

3.2. Land cover change from supervised classification 3.2.1. Change in forest cover The total area classified as forest was 8060, 7024, 4633 and 4278 ha in 1972, 1992, 1999 and 2001, respectively. The deforestation trend in the study area was therefore clear, but there was a complex pattern of conversion and regrowth during the study period (Table 4), particularly since the rate of recovery of forest strongly depends on the degree of disturbance (i.e. length of tavy period) and spatial pattern of clearing (i.e. at forest edges or within forest areas). An important consideration is of course the differences in spatial and spectral resolution of the MSS and TM/ETM sensors, which may have lead to an overestimation of forest cover in 1972 in cases where only small patches of forest were cleared. Conversion rates were estimated at about 52 ha yr
1in the 20-year period between 1972 and 1992, 341 ha yr
1 between 1992 and 1999, and about 178 ha yr
1 between 1999 and 2001, as an average for these periods. In other words, there was a dramatic increase in deforestation rates after 1992 and throughout the 1990s. This increase may in part be explained by the turbulent political situation in Madagascar during the period from 1991 and throughout the 1990s, which resulted in widespread turmoil and a weakening (of an already weak) forest management system (Andrianorofanomezana, personal communication). Another explanatory factor may be increasing population pressure in the region, which almost doubled between 1975 and 2001 in Ambositra, and increased dramatically near forest margins in Ranomena (Table 2), although the areas with highest deforestation rates were not those with highest population densities. The varying rates for different time periods also show that deforestation is not a linear process, but rather a result of interactions among complex

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

224

Table 4 Forest cover change trajectories during study period (1972–2001) Change trajectory

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Description

1972

1992

1999

2001

Forest Forest Forest Forest Forest Forest Forest Forest Nonforest Nonforest Nonforest Nonforest Nonforest Nonforest

Forest Forest Forest Nonforest Nonforest Forest Nonforest Nonforest Forest Nonforest Forest Forest Nonforest Nonforest

Forest Forest Nonforest Nonforest Forest Nonforest Forest Nonforest Forest Forest Forest Nonforest Nonforest Nonforest

Forest Nonforest Nonforest Nonforest Forest Forest Nonforest Forest Forest Forest Nonforest Forest Forest Nonforest

political, social and ecological processes, making the prediction of future forest cover difficult without taking all of these factors into consideration. Particularly erroneous results may result from studies where only two dates (i.e. before and after) are taken into consideration, which is relatively common. Earlier studies in Madagascar have also reported similarly complex deforestation dynamics (Jarosz, 1993; Kull, 1999). This also illustrates the importance of establishing national frameworks for monitoring of vegetation change dynamics at proper scales to better manage remaining forest areas. About 11% of the area classified as forest in 1972 had been converted to eucalyptus in 2001, while 22.5% was under savoka. In 1999, slightly more than 21% of the area classified as forest in 1972 was classified as a combination of savoka and recent burn. An analysis of the various forest

Stable primary or secondary forest Very recent forest clearing Recent forest clearing Old (permanent) forest clearing Old clearing with regrowth Recent clearing with regrowth Repeated clearing and regrowth Old clearing with recent regrowth Old regrowth Old regrowth Old regrowth with recent clearing Repeated clearing and regrowth Recent regrowth Stable grassland or perm. agr.

Coverage (ha)

(%)

3467.4 771.2 1887.2 1323.4 17.5 522.7 50.4 19.7 164.2 9.5 153.4 58.4 18.8 39095.1

7.3 1.6 4.0 2.8 – 1.1 0.1 – 0.3 – 0.3 0.1 – 82.2

cover change trajectories in the study area showed a general trend in the direction of natural forest to eucalyptus plantations and tavy. The latter accounts for the significant fragmentation observed in remaining natural forest areas (Fig. 2). The pattern of deforestation in the study area was generally from west to east, but was complex due to the high level of fragmentation (Fig. 2), illustrating the importance of studying these processes at appropriate spatial scales. 3.2.2. Nonforest land use change trajectories There was a decline in the total area classified as grassland savanna between 1972 (approximately 20,000 ha) and 2001 (15,340 ha), although the estimate from 1972 is associated with some uncertainty. This corresponded to a decrease of about 4,400 ha if recently burnt grasslands in 2001 were taken into consideration. There was increased

Fig. 2. Change in forest cover between 1972 and 2001 (coordinates are UTM 38s).

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

225

Fig. 3. Change in cultivated area on slopes (tanety) between 1972 and 2001 (coordinates are UTM 38s).

cultivation of previous grasslands further away from paddy fields (Fig. 3) and a general increase in cultivation of slopes (tanety), particularly between 1992 and 1999 (Fig. 4).About 3800 ha of the areas classified as grassland savanna in 1972 were cultivated in 2001, while approximately 2100 ha of former tanety was under grassland in 2001, making the net transition in the direction grassland to cultivation 1700 ha (89.5 ha yr
1). An estimated 4000 ha grassland was converted to eucalyptus during the study period, and about 920 ha former eucalyptus was converted to grassland. About 1440 ha former paddy fields were converted to grassland savanna, mainly due to insufficient amounts of water for irrigation after clearing of forest (according to local farmers). Similar reductions in streamflow following clearing of natural forest have been reported in other

studies (Moraes et al., 1998). However, differences in data on streamflow response to clearing are large and can only partly be explained by differences in rainfall (Bruijnzeel, 1998), and this is therefore an area that warrants more research in Madagascar, the results of which could help in better understanding the impacts of deforestation and ecological change on watershed level processes. The above land use change trajectories and trends indicate significantly increasing pressure on available land resources in the study area, leading to the cultivation of increasingly marginal areas (i.e. savanna grassland), which again leads to dramatic soil fertility decline, as shown by Va˚gen et al. (unpublished). It is imperative that these trends are taken into consideration when developing strategies for agricultural development in Madagascar, as focusing solely on alternatives to slash-and-burn agriculture will not help in mitigating current degradation trends in vast areas with marginal savanna grasslands. 3.3. Predictive modelling of deforestation risk

Fig. 4. Estimated change in areas of selected nonforest land use types during the study period.

Various factors determining deforestation rates have been proposed and studied in earlier work from Madagascar and in other parts of the humid tropics, including topography (elevation and slope) (Green and Sussman, 1990; Nagendra et al., 2003; McConnell et al., 2004), population density (demography) (Green and Sussman, 1990; McConnell et al., 2004) and distance to roads (Nagendra et al., 2003) and village centres (Mertens and Lambin, 2000; McConnell et al., 2004). Nagendra et al. (2003) reported higher rates of deforestation in easily accessible areas in a study in Honduras, but these patterns were altered with changes in government policies promoting mountain coffee production. Farmers and local experts

226

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

Table 5 Proportional odds model parameter estimates, results of Wald tests and standard errors Variable

Estimate

Standard error

Wald statistic

p-Value

Intercept Distance to road Distance to village Change in population density Elevation (m) Slope (deg)

26.4420 -0.0006 -0.0016 -0.0803 -0.0165 -0.0218

0.41 0.00 0.00 0.01 0.00 0.00

63.74 -48.8 -100.8 -10.7 -54.1 -9.8

0 0 0 0 0 0

in the study area also identified accessibility as the major reason for not clearing some of the remaining primary forest areas when asked in informal interviews. As the post-classification comparisons yielded two periods with distinctively different rates of deforestation, three ordinal levels were chosen for the dependent variable in the logistic model; 0: forest (no change), 1: deforested between 1972 and 1992, and 2: deforested between 1992 and 2001. Distance to roads and villages, elevation and slope,

and population increase all had significant influence on deforestation risk predictions in the logistic model (Table 5). The Wald statistics in Table 5 indicate that distance to village centre was the strongest predictor of deforestation, followed by elevation and distance to nearest road, respectively. Green and Sussman (1990) reported selective clearing of forests on flat land. The present study showed a weak correlation between slope steepness and deforestation pattern, but the parameter estimates indicated that steeper

Fig. 5. Estimated probability of deforestation (upper map) showing log-odds of deforestation. Lower map shows residuals ðY
PÞ. Coordinates are UTM 38s.

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

slopes had a higher probability of being cleared than flatter areas. Similar findings were reported by McConnell et al. (2004), and could of course be partly due to the fact that little forest on flatter land is available for clearing, as indicated by the median slope of 8  for remaining forest areas in 1972 and areas cleared during the study period, and the fact that remaining forest areas are located on the steep eastern escarpment. The weak influence of increasing population density (Table 5) was not surprising given the relatively poor spatial representation of available data for one, and also the fact that areas with most significant deforestation were those with less change in population density (i.e. areas with the lowest overall population). This was also indicated by the logistic model parameter estimate for population change, and again contradicted the findings of Green and Sussman (1990), underlining the importance of spatial scale in such assessments. One could, however, probably argue that distance to village centre is a proxy for population density. One of the primary reasons for the felling of exotic tree species (e.g. palisander) in the study area is demand from the wood-carving industry in Ambositra, which is the nearest town and Madagascar’s ‘‘centre of art’’, and distance to roads was also identified by farmers in the area as a major reason for not clearing remaining forest areas due to difficulties in getting the cut timber transported out of these areas. The model parameter estimate in Table 5 does indeed show that the odds for deforestation decrease with increasing distance to roads, a finding reported in several other studies including Mertens and Lambin (2000) who studied land cover change trajectories in Cameroon. The spatial representation of estimated deforestation risk shows higher risk in western areas, as expected given that they are at higher elevations and generally more accessible (Fig. 5). Areas predicted to have a low probability of being cleared, but which were still cleared (Fig. 5- lower map), have high residual values (i.e. dark). The majority of these areas were located within remaining forest, corroborating the importance of fragmentation through tavy, as discussed earlier. The map of model residuals also shows that significant areas of western forest at lower altitudes were cleared despite low probabilities for conversion, while certain areas at higher elevations were not cleared despite high deforestation risk estimates. The latter was mainly due to inaccessibility of some remaining forest areas as they were east and south of some of the major tributaries to the Mananjary river, reducing the rate of conversion in these areas.

4. Conclusions Despite the focus on Madagascar as a global hot-spot for biodiversity conservation, studies of deforestation and land use change on the island are often based on the simplistic view that the island was once completely covered by forest. Assessments of change at inappropriate scales or simple

227

extrapolations of earlier estimates are also common, making current estimates uncertain. This study has shown that deforestation and land use change trajectories in the eastern highlands of Madagascar are highly complex, both in space and time due to a high degree of fragmentation of remaining forest areas and varying rates of conversion during different time periods. Change in natural forest cover was assessed with a high degree of accuracy for all image dates (1972, 1992, 1999 and 2001) included in the study, while accuracy was poorer for savoka and eucalyptus areas, particularly in the image from 1972. Deforestation rates were relatively low during the first 20 years of the study period (on average 52 ha yr
1), but increased dramatically between 1992 and 1999 (341 ha yr
1) probably due to the turbulent political situation on the island during this period coupled with increasing pressure on forest resources due to population increase. Accessibility (distance to nearest village and road) and elevation were stronger predictors of deforestation risk than change in population density and slope. The most important nonforest land use change trajectories were cultivation (  3800 ha) and planting of eucalyptus (  4000 ha) in former grassland savanna areas, contributing to an overall reduction in the area under grassland savanna of about 4400 ha during the study period. Significant areas with former paddy cultivation had also been converted to grassland, mainly due to insufficient amounts of water available for irrigation. The increased cultivation of marginal grassland areas is an important indicator of pressure on land resources in the area, and has been shown by Vagen et al. (unpublished) to lead to significant soil fertility decline. Acknowledgements The author would like to acknowledge the assistance of Masy and Salmata Andrianorofanomezana and the President Fokontany of Ranomena during preparations for the study and in collection of ground truth data. The collaboration of Foiben-Taosarintanin’i Madagasikara (FTM) in making aerial photographs and maps available was also important for the progress of the study. This work was carried out as part of a PhD study funded by the Norwegian Research Council.

References Aldrich, J.H., Nelson, F.D., 1984. Linear probability, logit, and probit models. Series: Quantitative Applications in the Social Sciences. SAGE Publications, Inc., London, UK, p. 95. Baron, R., 1889. The flora of Madagascar. Bot. J. Linn. Soc. 25, 246–294. Boiteau, P., 1982. Contribution a` l’histoire de la nation Malgache. Reprint from 1958, Paris, Edition Sociales. Bruijnzeel, L.A., 1998. Soil chemical changes after tropical forest disturbance and conversion: the hydrological perspective. In: Schulte, A., Ruhiyat, D. (Eds.), Soils of Tropical Forest Ecosystems—Characteristics, Ecology and Management. pp. 45–61.

228

T.-G. Va˚gen / Agriculture, Ecosystems and Environment 115 (2006) 219–228

Burney, D.A., 1997. Theories and facts regarding holocene environmental change before and after human colonization. In: Goodman, S.M., Patterson, B.D. (Eds.), Natural Change and Human Impact in Madagascar. Smithsonian Institution Press, Washington. Chevalier, L., 1952. Madagascar: Polpulations et resources, Institut national d’e´tudes de´mographiques, Travaux et Documents, Cahier no. 15, Presses Universitaires De France, Paris. Cohen, J., 1960. A coefficient of agreement of nominal scales. Educ. Psychologic. Meas. 20 (1), 37–46. Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 37 (1), 35–46. Deichmann, U., 1994. A medium resolution population database for Africa. In: Database Documentation and Digital Database, National Center for Geographic Information and Analysis (NCGIA), University of California, Santa Barbara. Fahrmeir, L., Tutz, G., 1994. Multivariate statistical modelling based on generelatized linear models. Springer Series in Statistics. Springer Verlag, New-York. Faramalala, M.H., 1988. Etude de la Ve´ge´tation de Madagascar a` l’aide des Donne´es spatiales, Univ. Paul Sabatier de Toulouse. Faramalala, M.H., 1995. Cartographie de la ve´ge´tation de Madagascar, Abstracts of the International Symposium—Biogeography of Madagascar, Socie´te´ de Bioge´ographie, Paris. Gade, D.W., 1996. Deforestation and its effects in highland Madagascar. Mountain Res. Develop. 16 (2), 101–116. Grandidier, G., 1905. Les animaux disparus de Madagascar. Revue de Madagascar (2). Green, G.M., Sussman, R.W., 1990. Deforestation history of the Eastern rain forest of Madagascar from satellite images. Science 248, 212–215. Harrell, F.E., 2003. The design library for S-Plus version 6.x, Functions available on the web at: http://lib.stat.cmu.edu/S/Harrell/. Hornac, J., 1943. Le de´boisement et la politique forestie´re a` Madagascar, Me´moire de Stage, Me´moire de l’Ecole Coloniale ENFOM, Archives D’Outre Mer. Aix- en Provence. Archives Nationales de France. Hosmer, D.W., Lemeshow, S., 1989. Applied Logistic Regression. Wiley Series in Probability and Mathematical Statistics. Wiley and Sons, New York. Humbert, H., 1951. Les territoire phytoge´ographiques du nord de Madagascar. Compte Rendu de Se´ances de la Socie´te´ de Bioge´ographie 246, 176–184. Humbert, H., 1955. Les territoire phytoge´ographiques de Madagascar. Anne´e Biologique (se´rie 3), 31: 429–448; Colloque International du Centre National de la Recherche Scientifique, LIX, ’’Les divisions e´cologiques du Monde.’’ Moyens d’expression, nomenclature, cartographie, Centre National de la Recherche Scientifique, Paris, p. 195– 204. Humbert, H. and Cours Darne, G. (1965). Carte Internationale du Tapis Ve´ge´tal et des Conditions Ecologiques, 3 coupurs au 1/1,000,000 de Madagascar, Travaux de la Section Scientifique et Technique de l’Institut Franc¸ais de Pondiche´ry (hors se´rie). Jarosz, L., 1993. Defining and explaining tropical deforestation: shifting cultivation and population growth in colonial Madagascar (1896–1940). Econ. Geogr. 69 (4), 366–379. Kent, R.K., 1970. Early Kingdoms in Madagascar. 1500–1700. Holt, Rinehart and Winston, Inc..

Kull, C., 1999. Observations on repressive environmental policies and landscape burning strategies in Madagascar. Afr. Studies Quart. 3 . Kull, C., 2002. Empowering pyromaniacs in Madagascar: ideology and legitimacy in community-based natural resource management. Develop. Change 33, 57–78. Lambin, E.F., 1999. Monitoring forest degradation in tropical regions by remote sensing: some methodological issues. Global Ecol. Biogeogr. 8, 191–198. Lillesand, T.M., Kiefer, R.W., 2000. Remote Sensing and Image Interpretation. John Wiley & Sons, Inc., New York. Lowry, P.P., Schatz, G.E., Phillipson, P.B., 1997. The classification of natural and anthropogenic vegetation in Madagascar. In: Goodman, S.M., Patterson, B.D. (Eds.), Natural Change and Human Impact in Madagascar. Smithsonian Institution Press, Washington. Mayaux, P., Vale´ry, G., Bartholome´, E., 1999. Mapping the Forest Cover of Madagascar with SPOT 4-VEGETATION data. Global Vegetation Monitoring Unit, Space Applications Institute, Italy. McConnell, W.J., 2002. Misconstrued land use in Vohibazaha: participatory planning in the periphery of Madagascar’s Mantadia National Park. Land Use Policy 19, 217–230. McConnell, W.J., Sweeney, S.P., Mulley, B., 2004. Physical and social access to land: spatio-temporal patterns of agricultural expansion in Madagascar. Agric. Ecosys. Environ. 101, 171–184. McCullagh, P., Nelder, J.A., 1989. Generalized Linear Models. Chapman and Hall, London. Mertens, B., Lambin, E.F., 2000. Lan-cover-change trajectories in Southern Cameroon. Annu. Assoc. Am. Geograph. 90 (3), 467–494. Moraes, J.M., Pellegrino, G.Q., Ballester, M.V., Maretinelli, L.A., Victoria, R.L., Krusche, A.V., 1998. Trends in hydrological parameters of a Southern Brazilian watershed in relation to human induced changes. Water Res. Manage. 12, 295–311. Nagendra, H., Southworth, J., Tucker, C., 2003. Accessibility as a determinant of landscape transformation in Western Honduras: linking pattern and process. Landscape Ecol. 18 (2), 141–158. Nelder, J.A., Wedderburn, R.W.M., 1972. Generalized linear models. J. Roy. Stat. Soc. A 135, 370–384. Nelson, R., Horning, N., 1993. AVHRR-LAC estimates of forest cover area in Madagascar. Int. J. Remote Sens. 14, 1463–1475. Perrier de la Baˆthie, H., 1921. La ve´ge´tation malgache. Annales de l’Institut Botanicoge´ologique Colonial de Marseille (se´rie 3) 9, 1–226. Perrin, M.C., Millington, A.C., 1997. Monitoring biomass burning using ATSR-2 data. In: Proceedings of the third ERS Symposium—Space at the Service of our Environment, 14–21 March, 1997, Florence, Italy. Ringrose, S., Matheson, W., Tempest, F., Boyle, T., 1990. The development and causes of range degradation features in southeast Botswana using multi-temporal Landsat MSS imagery. Photogram. Eng. Remote Sens. 56, 1253–1262. Roberts, D.A., Batista, G.T., Pereira, J.L.G., Waller, E.K., Nelson, B.W., 1998. Change identification using multitemporal spectral mixture analysis: applications in eastern Amazonia. In: Lunetta, R.S., Elvidge, C.D. (Eds.), Remote Sensing Change Detection: Environmental Monitoring Methods and Applications. Ann Arbor Press, Michigan, USA, p. 308. Viguier, R., 1914. Ge´ographie botanique de Madagascar. Agric. Pract. Pays Chauds 14, 243–249.