Assessing Forest Cover in the Uplands of Southeast Asia

Stephen Leisz *Δ, Christian Tottrup*, Michael Schultz Rasmussen*, Tran Duc Vien Δ & Kjeld Rasmussen*
*) Institute of Geography, University of Copenhagen, Øster Voldgade 10, DK-1350 Copenhagen K, Denmark
Δ) Center for Agricultural Research and Ecological Studies, Hanoi Agricultural University, Gia Lam, Hanoi, Vietnam

Abstract

Remote sensing data and digital image processing is widespread being used to map tropical rainforest and its changes. The purposes differ from global and regional assessment of forest cover in a global change context to local needs for natural resource planning support. While good accuracies can be obtained using the right methodology (overall accuracy superior of 90 %) for a limited area, the broad worldwide or regional maps may work at the large scale, but will inevitably have a bias at the local scale. Despite these shortcomings, these latter forest classifications are more and more being used as planning support because of their extended coverage and readily availability. The present paper assesses the pertinence of using broad global classifications in a local context in mountain forests in the Ca river basin in northern Vietnam. In order to understand the land-cover and the production system, field studies were conducted and it appeared that there was no or insignificant on-going clear-cutting of the remaining forest and shifting cultivation was important in the area. Settlements varied much in size from less than a hectare to over 10 hectares and the fields were cultivated from one to three years at the time, leaving the soil fallow for a period of a few years up to 20 years. Four broad land cover classifications available for the Ca river basin were compared to determine how well they characterized the land-cover and the production system found. The three broad classification schemes, Pathfinder (Tropical Rainforest Information Center, 1998), FAO (FAO, 2001) and Tropical Ecosystem Environment Observation by Satellite (TREES, 2001) where compared to Tottrup’s verified land-cover maps derived from multi-date classification of Landsat ETM data (Tottrup 2002). It was shown that the broad classifications over- and underestimated the forest area with a magnitude in difference and hereby fail to capture the important balance between the agricultural system and the forest. Specifically land-cover classes associated to the exploitation of the forest, the degraded forest, bamboo, shrub, grass and farmland are not allocated consistently to either forest or non-forest. These classes are of primary relevance to local forest management and it is concluded that broad classifications should not be used in a local context without previous verification. The paper includes discussion on how lack of understanding of the production system affects the classification result and equally how unintentional selections of remote sensing data from different periods can have a significant influence on the classification result. It is concluded that using multi-date classifications of Landsat or equivalent high-resolution data provide the necessary high quality information to characterize the forest dynamic.
Keywords: Remote Sensing, Deforestation, Land cover, Tropical forest, Landsat, Shifting cultivation

Introduction

The mountainous area of Southeast Asia is considered one of the most important deforestation and land cover change “hotspots” in the world.  Along with the Amazon Basin and the Congo Basin, the forests of Southeast Asia are considered the most vulnerable to deforestation.  Land cover change, including deforestation, is considered a very important factor affecting ecological systems (Vitousek 1994).  It is often considered environmentally harmful and one of the driving factors behind global climate change. When a mountainous area is deforested it may affect the area’s hydrological cycle (Konnick 1999), leading to erosion, flooding, sedimentation, and changes in water quality that negatively impact on downstream agriculture, aquaculture activities, and hydropower generation (Douglas 1999, Lørup et al 1998).  When this same area is deforested, the carbon that has been stored in the forest is released into the atmosphere and the area goes from being a carbon sink, to a producer of carbon in the short run, contributing greenhouse gases to the atmosphere (Skole and Tucker 1993, Foody et al. 1996).  Furthermore, land cover change is also expected to be the most significant variable affecting biodiversity for next century (Chapin et al. 2000).

Deforestation, and more generally land cover change, can also affect the prospects of economic development and local livelihoods. On the positive side the clearance of forest for agricultural purposes and timber production, among other activities, may help to raise national, local and household incomes. This must however be balanced by the adverse effects caused by reduced supplies of wood for energy and housing, decreased amounts of non-timber forest products, such as fruits, rattan, wildlife and medicinal plants (Lambin 1994), and perhaps rising prices on forest goods. Thus, the published accounts that the mountain areas of Southeast Asia are undergoing widespread deforestation are a cause for ecological, economic and social concerns. 
However, are these types of areas really going through a rapid deforestation, or is the structure of the vegetation covering the land, the forest morphology, merely changing overtime? If this is the case, is the change in structure permanent or part of longer-term historical patterns?  The FAO defines deforestation as “the transfer of forest land to non-forest uses and includes all land where the forest cover has been stripped and the land converted to such uses as permanent cultivation, shifting cultivation, human settlements, mining, and building of dams” (Rao 1989, Holmgren and Davis 2000).  By this definition deforestation means the complete clearing of forest and the permanent change in the land-use so that the forest does not regenerate.  A non-permanent change in the land-cover, e.g. in situations where forest land is temporarily cleared and then allowed to regenerate, or where forest is thinned of trees, may be variously thought of as “degradation” (Rao 1989) or, less judgmentally, as a change in forest morphology.

The term “forest” is also open to different interpretations.  For example, one definition states that forest is “generally, an ecosystem characterized by a more or less dense and extensive tree cover,” and “a plant community predominantly of trees and other woody vegetation, growing more or less closely together,” (Ford-Robertson 1971).  Another states that it is an “ecosystem with a minimum of 10 percent crown cover of trees and/or bamboos, generally associated with wild flora and fauna and natural soil conditions, and not subject to agricultural practices” (FAO 1999). 

Fox et al. (2000) suggests that deforestation, referring to a permanent change in land cover from forest to non-forest, is not taking place to the extent previously suggested for the northern mountain regions of Vietnam. Rather, given the farming systems and logging practices, what is actually happening is that the land-cover is temporarily being cleared of trees and used for active swidden, and the general class of “forest” is quickly regenerating. Our experience in the Ca River Basin (CRB) seems to support the analysis that the land-cover change that is taking place is not one of “forest” to “non-forest,” but rather one of changes in the quality of the broad category of “forest.” It appears that the changes taking place are in the amount of area that is found in the “degraded” forest, brush, bamboo, and grass types of land-cover and the movements of areas between these categories, rather than in wholesale movement of areas from “forest” to “non-forest” or the “other land cover” categories.

In order to confidently state that deforestation is going on at a rapid pace in Southeast Asia, it must be shown that forest areas are being permanently converted to other, non-forest, and land-cover types. This requires primarily a good understanding of the relevant production and ecological systems, in order to identify pertinent land cover classes and assess how remotely sensed data can contribute to map the changes. The choices of remote sensing tools and methods have to be sensitive enough and appropriate to distinguish between important forest types and different types of non-forest land-cover.  These tools and methods also have to produce results that are consistent over time with repetitive mapping exercises.

The work described in this paper is based on an environmental assessment study in the CRB, an activity within the Resource Policy Support Initiative (REPSI) project (WRI, 2002, Institute of Geography, 2002). Figure 1 shows the geographic location of the study. The objective of this paper is to assess the information contents of different remote sensing forest-mapping approaches relative to the ecosystem and production system complexity in the study area. It is believed that in order to map relevant changes in forest cover, land cover maps should be produced with a high accuracy and with a high temporal frequency.

Since quality information on forest cover in areas like Southeast Asia is scarce or often is collected in a decentralised manner; a number of decision makers are willing to use the coarse and gross forest maps derived from remotely sensed data. These are intended to be used for larger regions or even at the global scale and are easily made available through the Internet or freely distributed on CD ROM. Consequently, a second objective is to assess the quality of these coarse maps for the case of the CRB, both in terms of the actual information, how reliable the map is itself, as well as their capabilities to predict changes in time from a series of maps. It is anticipated the coarse maps may not capture details, but will they be accurate enough to capture the general situation? Furthermore, are the methods applied for the regional or global mapping robust enough to identify significant changes in forest cover in time or are the biases of these maps varying too much over time to be used in a local context? The lessons learned will be relevant for other areas of the Southeast Asia mountain region.

Remote Sensing

Remote sensing of reflected electro-magnetic radiation from the earth’s surface has been used intensively to map forests over the past 30 years since 1972 when the first earth resource satellite was launched. The sensor system carried by earth observation satellites is often characterized by three distinctive resolutions. The sensors spectral resolution refers to the number and dimensions of specific wavelength intervals recorded by the sensor. Temporal resolution refers to the time interval between repetitive recordings of a specific area. Spatial resolution refers to the smallest detectable area, by the sensor, on the earth’s surface.  Spatial resolution is often referred to as the pixel size of the image (pixel is short for picture element). The choice of sensor system for a specific study (and the methods that are used to interpret the information provided by the sensor) often involves a balance between the specific objectives of the study and the cost and availability of data. The first Landsat instrument was the Multispectral Scanner (MSS) with 79-meter pixels and was followed by the Thematic Mapper (TM) with 30-meter pixels in 1984. The latest Landsat 7 satellite with the Enhanced Thematic Mapper (ETM+) instrument has a spectral resolution of 7 bands in the visible and near-infrared and mid-infrared parts of the spectrum, the temporal resolution is 16 days and the multispectral spatial resolution is 30 meters and 15 meters for the single pan-chromatic band.

There are two main applications of satellite remote sensing to forest mapping, the first, and most widely used, is the classification of multispectral data from such satellites as Landsat, Spot or Ikonos. The classification is done visually or automatically using digital image processing.  The principle in these cases is to characterise the spectral signatures (a unique combination of reflected electro-magnetic radiation in different windows of the visible, near-infrared, and mid-infrared parts of the electromagnetic spectrum) for the land cover or forest classes that should be mapped. There are two commonly used methods for doing this. One is the supervised classification method.  This method consists of the use of training areas, which are identified in the field, that are used to derive the spectral signatures prior to the digital image processing software generalising this information for the study area. The other widely used method is the unsupervised classification.  This method allows the digital image processing software to initially subdivide any given area into a user-defined number of classes.  Next information from fieldwork or from other ancillary data is used to help label the classes and possibly merge (or split) some of them in order to obtain a generalized map of land cover.

The second main application of satellite remote sensing to forest mapping has been developed over the past 20 years using time series data from the NOAA Advanced Very High Resolution Radiometer (AVHRR) to capture the dynamics of changes in vegetation from vegetation index data. The AVHRR instrument has only two bands in the visible and the near infrared part of the spectrum and the spatial and temporal resolution is 1.1 km and 12 hours respectively. This is however sufficient to compute a vegetation index. The vegetation index is a measure of the photosynthetic capacity, and annual and inter-annual variation can be used to identify forest communities with the same photosynthetic dynamic. This approach also needs additional field checking in order to produce a generalised forest or land cover map. To learn more about remote sensing of forests, see Tottrup (2002) and Lambin et al (1997).

Remote Sensing of Tropical Forests

The majority of what is known about land cover in Vietnam comes from one of two sources: periodic land-cover analyses carried out by Vietnam’s Forestry Department and/or General Direction of Land Administration, or large area, small scale, analysis of land-cover done by international research projects focusing on tropical forest cover assessments and change.  These programs include the Landsat Pathfinder Project (Tropical Rainforest Information Center 1998), the FAO Forest Resource Assessment (2001), and the “Tropical Ecosystem Environment Observation by Satellites” (TREES) program (2001).  These programs also often collaborate with institutions within Vietnam, such as the Department of Forestry, in order to carry out their assessments.

Landsat Pathfinder has carried out assessments of forest cover in the tropics on a decadal basis for the 1970s, 1980s, and 1990s.  For our study area the most recent assessment was done using 1992 Landsat TM data.  The FAO Forest Resource Assessment is also done on a decadal basis and for our study area the most recent assessment was done using AVHRR data from the year 2000.  The TREES program is also trying to do decadal analysis of forest cover, although only one has been done to date.  For our study area the assessment was done using AVHRR data from 1992/1993.

The government of Vietnam draws on the Department of Forestry’s land-cover analyses for many of its planning needs.  International non-governmental organizations (NGOs) and multi-lateral organizations working in the fields of development and conservation / biodiversity rely on either the Forest Department’s land-cover assessments, assessments that have already been done by the aforementioned projects, or assessments done by groups using methods similar to those used by the aforementioned projects.  The varying land-cover analyses are used to decide on such things as where biodiversity is most threatened, how much deforestation is going on, where conservation priorities should be focussed and what government land use policies should be pursued.  The information derived from these analyses is also used to determine changes in land-cover over time and how humans (and their agricultural systems) impact land-cover. All of these analyses rely on remote sensing. 

In order to measure world-wide or regional scale tropical deforestation, studies using remotely sensed data, especially NOAA AVHRR and Landsat MSS/TM data have been used.  Some of these studies, such as the Landsat Pathfinder Project, the FAO Forest Resource Assessments and TREES, can be viewed as operational forest monitoring systems. In order to assess deforestation rates, the procedures followed in these studies have been to simplify the input classes to just one or only a few broad forest classes. For example, Pathfinder classifies the Landsat imagery into four classes: forest, non-forest, water and clouds (Brunner et al. 1998) and TREES divides land-cover into evergreen forest, seasonal forest and non-forest. The simplified classification scheme is used to compare the classified images from different dates on a pixel by pixel basis. This simplified classification scheme provides the user with high individual scene accuracy and helps maintain the accuracy of the final analysis (Salas et al. unpublished report).  In these simplified classification schemes land cover types such as primary forest, secondary forest, bamboo forest, and many combinations of woody brush are usually considered forest, and barren areas, grassland, rock and agriculture are usually considered non-forest. 

In addition to the above studies related to the classification of the tropical forest a range of studies have considered the spatial patterns detectable in remotely sensed imagery (e.g. Mertens and Lambin 1997; Brunner et al. 1998, Salas et al. unpublished report ). Husson (1995)  proposed a typology in which recognizable spatial patterns of non-forest areas are associated with different processes of deforestation. This typology distinguishes between the following patterns: Geometric (large-scale clearings for modern sector activities), Island (peri-urban areas), Corridor (roadside colonization by spontaneous migrants), Diffuse (smallholder, traditional subsistence agriculture), Fishbone (planned resettlements schemes) and Patchy (high population density areas with residual forest patches). The typology is useful for stratifying larger regions into areas characterized by uniform deforestation patterns.

Some studies suggest that the spatial patterns associated with individual clearings can indicate whether the forest has been cleared for agricultural purposes or for logging. These studies suggest that within Southeast Asia, non-forest areas less than 4 hectares are clearings for agricultural purposes, while areas larger than 4 hectares are clear-cut logging areas (Brunner et al. 1998).  The Pathfinder forest / non-forest classification has also been used to determine shifting cultivation rotations (Brunner et al. 1999).

Leisz et al. (2001) carried out five land-cover analyses as part of a large study on environmental, social conditions and development trends in five communities in the northern mountain region of Vietnam.  One of these studies was in the CRB. In contrast to the methods described above, this land-cover analysis, using Landsat TM imagery from November 1998, is based on a supervised classification of the satellite image using a maximum likelihood classification.  Fieldwork was done to collect 200 ground truth points that were used in the classification. At each location where ground truth data was collected, local informants were interviewed regarding the land cover and the history of land use in the area. Of the ground truth point data, 69 were used to create training sets and 131 were used to assess the accuracy of the final land-cover analysis.  This analysis determined that classes of forest, bamboo forest, brush, tall grass, short grass, wet rice fields, and swidden (upland) agriculture fields, could be delineated with a high degree of confidence. The study results showed that 83% of the ground truth reference points were correctly classified. This result suggests that given good ground truthing it is possible to delineate a more complex land-cover map of the study area than the classifications discussed above.

Tottrup (2002) also used a maximum likelihood supervised classification to derive a forest and land-cover map of the upper CRB. Tottrup used multi-date Landsat imagery and pre-classification image smoothing and achieved a higher classification detail and accuracy than Leisz et al. The smoothing of the imagery decreased the variation of the reflective response within the different vegetation types and provided a better training signature for the supervised classification. This was done as per Hill and Foody (1994) who suggest that the differences of spectral responses of tropical forest and land cover classes is more a function of textural differences within vegetation classes than between the classes. Thus, smoothing decreases the “in-class” differences and accentuates the “between-class” differences, providing the analyst with better training sets for carrying out the land-cover classification analysis. Also, Tottrup’s use of multi-date imagery significantly improved the classification performance by extending the information content to include information on phenological stages and canopy texture. Tottrup’s results suggest that it is possible to accurately classify up to thirteen types of land-cover in the study area: water, rocks, built-up area, wet rice paddy, bare soil, grass, farmland (swidden or dryland fields), shrub, bamboo, degraded forest, deciduous forest, karst, forest, and primary forest.  Notably, this also shows that with the proper ground truth data and methods, different types of forest, such as shrub, bamboo, degraded forest, karst forest, and primary forest can be delineated in areas similar to the study area and an overall classification accuracy above 90 per cent can be achieved for these classes.

Another aspect of Tottrup’s study was to carry out a temporal land cover change analysis of the study area by using ground-truthing derived from historical interviews to create Forest / Non-forest cover maps.  The methodology used was to derive training sets for areas that were known from the interviews and other ancillary data to have not changed land-cover from 1975 to the present.  A supervised land-cover classification for 1975, 1992, and 1998 was then derived from a single image for each time period (Tottrup 2001).  The land-cover was then grouped into Forest / Non-forest categories following Pathfinder’s criteria for forest / non-forest.