The Use of Time-Series Satellite Data for Characterization and Monitoring of the Seasonal Forests and Savannas of Central Africa
|
|
The Use of Time-Series Satellite Data for Characterization and Monitoring of the Seasonal Forests and Savannas of Central Africa |
Christopher O.Justice, Ph.D., Nadine Laporte, Ph.D., William Lawrence, Ph.D. [Editor], David Long, Ph.D., Mark Heinicke.
Dept. Geography, University of Maryland, College Park, MD 20742, USA. Fax: (301) 314-9299,
tel: (301) 405-4050, email: justice@kratmos.gsfc.nasa.govTropical countries support some of the largest and most rapidly growing populations in the world. The presence of these human communities has strongly modified the extent, status and survival of natural ecosystems. Although virtually all of Africa's vegetation has been altered or disturbed to some extent by man, anthropogenic pressure is increasing on the remaining areas of relatively natural and undisturbed vegetation due to rapid population growth and the demands of economic development. These ever increasing pressures and concomitant disturbances, if not well managed, tend to lead to permanent conversion of natural forest and woodland cover to agriculture with a low probability of recovery or sustainability. This study provides a new map showing the current distribution of the forest/savanna systems of Central Africa and a preliminary study of human pressure on these systems.
Rationale for the study - Natural ecosystems are under tremendous pressure in Central Africa, but information concerning their distribution, function and level of disturbance inventory is very limited, scattered, often out of date, and may be in error. The mapping of these ecosystems is of critical importance for natural resources management, agricultural planning and biodiversity assessments. Frequent monitoring is also required due to the rate of anthropogenic change and the highly dynamic phenology and sensitivity of these seasonal forests and savannas to climate variation. The current vegetation cover of Central Africa (Congo, Gabon, Equatorial Guinea, Cameroon, Zaire, Central African Republic) and its seasonal dynamics are poorly characterized, at least in comparison to Sahelian and sub-Saharan Africa where severe droughts during the last two decades and resultant famine and loss of life have drawn international attention.
Besides lack of information on the extent and nature of Central African vegetation cover, thousands of square kilometers of the humid savannas are burned each year as part of common land use practices, not only temporarily altering the vegetation cover and physical surface properties, but also impacting regional and global atmospheric chemistry through release of greenhouse gases and particulates. Anthropogenic processes are very important to pattern and function of ecosystems of the forest/savanna transition region of Central Africa. The extent and timing of biomass burning in this region and its ecological and environmental effects have yet to be quantified.
Data Needs - Topography, soils, rainfall, and human and animal pressures are the primary factors controlling changes in vegetation type and distribution in Africa. In order to understand the interplay of these factors accurate, recent, and spatially specific datasets are required. Topography and soil survey information are relatively static and available to some degree (Biodiversity Support Program, 1993). Climatic information is not widely available, but some excellent long-term synopses have been made, from existing precipitation and temperature data from across the continent (Booth, 1993; Nicholson, et al.,1988; Hutchinson and Bischof, 1983). (click here for pictures) Human and animal pressures are poorly documented in the form of infrequent, highly aggregated human census data and livestock inventories. The presence of humans and their needs for food, fuel, fiber, water and shelter have strong impacts on the landscape. (click here for pictures) Domestic and wild animals also significantly impact the environment by grazing. Humans may even intensify this impact by gathering fodder for penned animals. Finally, the type, location and status of current vegetation cover is poorly documented. Most of the existing vegetation maps for the region are old (Devred, 1954; White, 1983; Koechlin, 1961) and difficult to compare due to different typology (Townshend et al., 1993), these differences stemming from methodological considerations and the reasons for which the maps were initially produced. (click here for pictures) The most recent maps published for the Central African region are for Central African Republic (Boulvert, 1986) and Cameroon (Letouzey, 1985). In any case, regional vegetation maps based on anything except remote sensing approaches will be inherently inaccurate and difficult to use in the precise domain of state-of-the-art geographic information system technology (GIS). Satellite image-based methods have recently been developed to derive land cover maps. The newest maps from this source were produced for the humid forests of Zaire and Cameroon for 1989 (Justice et al., 1993; Laporte, 1994) having appeared in publications from the Biodiversity Support Program (1993) and World Conservation Monitoring Centre (1993). (click here for pictures)
GIS Approach/Modeling - When faced with the difficulties in assembling, interpreting and understanding the wide range of data needed for integrative studies, major problems include those of scale, suitability and spatial density of data, missing data, erroneous data, and out-of-date data. In this study we utilized computer-based geographic information system [GIS] technology to integrate and summarize existing datasets and to transform multiple data layers into new information by modeling. GIS tools and capabilities can greatly facilitate this integration, including scale independence, interpolation routines, integration of vector, point, polygon and raster data sets, and multilayer modeling capabilities. Modeling allows the combination of disparate data into new information, based on usersupplied formulae.
Modeling efforts in this study have demonstrated how information derived from remote sensing images can be coupled with existing regional information for example on precipitation, soils, road networks in a GIS. The study has provided a better characterization of vegetation types, fire dynamics and the aerial extent of burning in the savanna environment. Remote sensing techniques allow us to assess vegetation distribution as well as its seasonal and multiyear dynamics. The extent and timing of biomass burning may soon be monitored routinely using GIS and remote sensing techniques.
Remote Sensing Datasets - Although remote sensing techniques for vegetation assessment have been available for several years there has been limited success at a comprehensive regional surveys. Some assessment work is underway in tropical forests, but there has not been a specific focus on the savanna and forest transition ecosystems of Central Africa. Over the past three years regional assessments have been initiated with respect to the closed canopy moist humid forests, in part due to their significance with respect to climate change and biodiversity. Such projects include the NASA Landsat Pathfinder Humid Tropical Project of University of Maryland and University of New Hampshire, and TREES at the EEC joint Research Center in Italy. There is now a need to develop appropriate techniques to provide an inventory and monitoring of the seasonal forests and savannas of Central Africa.
Most of the savanna cover types are seasonal in the sense of canopy development and the growth and senescence of lower vegetation strata. Monitoring vegetation seasonality at the regional level requires information at high temporal frequency. Only remote sensing techniques allow us to acquire such time sensitive information at a global scale (IGBP, 1992). A single satellite observation in time is of limited use for successfully separating vegetation classes, except in areas of low seasonality. For satellite mapping of the wet and dry savannas, miombo and seasonal forest ecosystems the seasonal dynamics of the vegetation can be used to discriminate vegetation types. The analytical requirement for a spatially comprehensive, regional, long-term, and multitemporal dataset can now be satisfied to a great extent, although not with data of the highest spatial resolution. In this study, we use time series satellite data derived from the National Oceanic and Atmospheric Administration [NOAA] Advanced Very High Resolution Radiometer [AVHRR] instrument to map the distribution of the seasonal forests and savannas of Central Africa.
The characterization of Central African vegetation in terms of seasonality, as developed for use in this project, was based on the multitemporal analysis of vegetation greenness indices in conjunction with rainfall data (Justice et al., 1985; Tucker et al., 1985). For purposes of this research, 'seasonality' is the change in density, color, or canopy architecture of a vegetation cover that manifests itself as changes in land surface characteristics discernible by satellite instruments. This greenness or seasonality signal is extracted from a spectral index, such as the Normalized Difference Vegetation Index [NDVI], (Tucker, 1983) which is calculated as the ratio of infrared to red reflectance from the viewed surface. NDVI and other related vegetation indices have been shown to be directly related to many ecosystem parameters and processes such as leaf area index, biomass, production, and absorbed photosynthetically active radiation.
Vegetation seasonality, as observed through monthly composites of NDVI values, shows changes primarily due to climatic influences (time and intensity of rainfall, local temperature, day length, and sun-surface geometry) and their effect on vegetation canopy development and appearance (see Prince, 1991). The availability of a long term data satellite data set permitted the mapping of Central African cover types by seasonality criteria. The satellite derived vegetation stratifications were enhanced through the use of local climate data. The regional map of vegetation cover generated in this study was used in support of our research into anthropogenic pressure. In addition to providing baseline information for resource management and environmental monitoring, the vegetation maps generated are of use for addressing the developing issues of biodiversity, global climate change and the human dimensions of global change. The benefit of the map based primarily on satellite data is that it reflects current conditions and is more spatially and temporally accurate than existing maps.
The maps generated from the remotely sensed data were incorporated in a Central African GIS developed by the project and compared with digital versions of existing maps for the region. The new, satellite data-based maps were developed and validated using high spatial resolution imagery (Landsat MSS, TM and SPOT), through consultation with local experts, and with field observations.
II. Project Research Activities
Continued international concern over conversion of tropical forests, savanna and transitional ecosystems and its impact on global environment and human habitability has lead us to examine the potential use of remote sensing and GIS methods for regional environmental studies, management and planning applications. These types of applications are not new but are certainly under continued development. The previous study (BSP, 1993) focused on mapping dense humid forest in Central Africa using NOAA AVHRR data and also examined Landsat as a source of high resolution spatial information for regional assessment of land cover change at the scale of a few hectares. The results of this initial study brought forth several priority research issues:
- the recommendation to include drier ecosystems in future research (savanna, dry forest, transitional areas);
- the need to evaluate the utility of multitemporal satellite datasets for better characterization of vegetation through phenological signals;
- the utilization of phenology in vegetation dynamics studies, particularly those related to carbon cycle, drought detection, plant productivity;
- the assessment of better vegetation classifications using vegetation phenology/dynamics especially forest/savanna boundary;
- the practicality and value of integrating several disparate datasets with GIS techniques and models to extract new information for assessment, management and planning such as indicators of anthropogenic pressure likely to lead to rapid land cover change.
The formulation of this current project, focused on several of these issues. In particular, we specifically address the extension of the Central African study to include savanna and transitional ecosystems, the utilization of multitemporal satellite data for assessment of phenological status and vegetation dynamics, the use of phenology for enhancement land cover classification, the integration of both climatic and remote sensing based datasets for land cover classification, and integration of population, road network, and land cover classification with GIS modeling techniques to derive spatially explicit delineations of areas susceptible to rapid land cover change due to anthropogenic pressures.
The specific activities of this project that support these research foci in Central Africa are:
- vegetation characterization in terms of seasonality;
- seasonal forest and savanna mapping;
- mapping of the areal extent, distribution and timing of fires;
- development of indicators of anthropogenic pressure; and
- development of a Central Africa GIS.
These objectives, a brief discussion of methodology and datasets, as well as an overview of results are presented below.
Objective 1. Vegetation characterization in terms of seasonality.
The development of this methodology provides the underpinning for the vegetation mapping activity (Objective 2). For this objective, vegetation community types were studied to see how their phenology and seasonal changes in foliar structure are represented by a relatively low resolution satellite instrument (ca. 8 X 8 km pixels). Basic criteria for classification were the temporal fluctuation of greenness (NDVI), from the long-term (1982-1987) NOAA AVHRR GAC NDVI database. Climatic, soils, topographic and other information were used in this analysis to develop the best stratification model.
Vegetation cover types were separated into classes of wet and dry savanna, dry evergreen forest, dry seasonal forest, and dense humid forest. A principal components analysis of the six year (1982-1987) GAC NDVI monthly composites was used to discriminate between classes in an unsupervised classification. (click here for pictures) This long-term record was used to reduce problems based on year-to-year seasonal differences. A total of 12 classes resulted from the classification, which were then checked against known cover types (from maps, air photography and high resolution satellite data) and then assigned to either dense humid forest, or wet and dry savannas. (click here for pictures) To eliminate some confusion between cover types and to further separate subclasses, rainfall indicators were also utilized (from Nicholson et al., (click here for pictures) 1988 and Booth, 1993 ( click here for pictures)). We found that rainfall criteria limited dense humid forest to areas with annual precipitation > 1500 mm, dry evergreen seasonal forests to areas with precipitation < 1400 mm, wet savannas to where precipitation > 1200 mm, a transitional savanna where precipitation ranged from 800 to 1200 mm, and dry savanna on sites where annual rainfall was < 800 mm. All closed canopy forests are located within dry, humid or hyper-humid climates. The combined use of NDVI seasonality plus precipitation was needed to extract all vegetation types. Once the rainfall and NDVI classes were known they were used to assign vegetation cover type to all of the Central Africa area.
Observations of NDVI showed that all forest types, even the dense humid forests, have seasonal changes in NDVI signal. In comparing wet and dry forests, the dry forests have very high NDVI maxima, but the NDVI range is higher than in wet forests. In dry years, dry forest communities have much reduced NDVI maxima and minima as compared to those of humid forests. Dry seasonal forests have a single annual NDVI peak, but wet forests have two peaks in wet years [the result of the 'little' dry season], but only a single NDVI peak in drier years. The use of seasonality and rainfall together provided a powerful classification tool even at the very coarse resolution of the GAC dataset, for cover type identification. Neither NDVI nor rainfall alone could be used to successfully discriminate vegetation cover types.
Objective 2. Seasonal forest and savanna mapping and monitoring.
The approach adopted for this objective was use of an existing multiyear database of monthly NDVI composites from coarse resolution (8 km) satellite data (NOAA AVHRR Global Area Coverage (GAC)) to determine vegetation seasonality and to use their seasonality to classify vegetation land cover type. High resolution Landsat MSS or TM satellite imagery was used for assistance in cover interpretation, validation and classification model refinement. The results of the coarse resolution mapping were evaluated by comparison with existing, conventional vegetation maps and those developed from high resolution satellite data. This objective was supported through continuing collaboration with Belgian colleagues at University Catholique of Louvain (UCL), African national forest services and NGO'S.
Results/Discussion
The evergreen and seasonal vegetation types are the most easily discriminated due to high contrasts in both their floristic and phenological characteristics. Although there is some difficulty in differentiating between savanna types because of year-to-year climatic variation, their highly mosaicked, fine-scale distribution patterns, land use practices and soil diversity, multiyear mean NDVI could be used to separate the main climatic and azonal woodland savanna types from one another. Subtypes within woodland savanna classes could be better separated by a tree cover density criterium if bioclimatic information were available describing their distribution. The 8 X 8 km GAC resolution NDVI data are too coarse for this fine level of discrimination.
Many of the difficulties we had in deriving results directly comparable to existing vegetation maps were attributable not to methodological errors, but to the differences between the floristic classifications of traditional vegetation analysis and the spectral and phenological approach used in our remote sensing classification. There are basic, perhaps irreconcilable, differences between regional, satellite-based classification and that of the traditional field-based map maker. In many cases even marked species composition differences cannot be discriminated from space, because of lack of any physical manifestation of floristic differences in the spectral nature of the foliage, its density, growth, or response to seasonal climatic changes. A good example, is the limit between forest/secondary grassland and the wet miombo (as classified by White, 1985) which is difficult to detect with remote sensing. Although this is a clear floristic zonation, it is neither spectrally nor phenologically distinct. Miombo has large areas with low tree density (like secondary forest grassland), and both are dominated by an overstory of broad-leaved trees whose cover density can vary due to a human disturbance and local soil conditions. This makes them appear very similar in the AVHRR data. Conversely, gradients in soil characteristics that affect water relations and/or nutrient status can appear as distinct vegetation classes even when floristics remain constant. In the future, some convergence in terminology between the botanical and remote sensing communities would be useful in defining new cover descriptions of use to a broad range of users.
Another set of methodological challenges to the GAC NDVI classification methodology came about because of its very coarse resolution (> 45 km2 per pixel). Gallery forest, although extensive in the transitional forest zones, was difficult to detect because it is spatially distributed in narrow bands along rivers. More importantly, deforestation, shifting cultivation and/or degraded forests in dense humid forests cannot be detected in the GAC NDVI dataset because of their size and their mixture with remaining forest fragments and areas of secondary forest growth. Disturbed areas were uniformly classified as intact forest unless they filled a large portion of the GAC pixel. Mixtures of secondary forest with grassland often fell into forest classes if forest blocks were large and fragmentation low. The remote sensing community is just beginning to understand the scale and mixture issues of using coarse resolution satellite data.
During the validation exercise using high spatial resolution data (usually Landsat MSS, ca. 60 X 60 m resolution) several general observations were made. The most important was that GAC classification always overestimated humid forest by up to 33% of a single MSS scene. Most of the confusion came from failing to capture areas of degraded forest (17%) or in the classification of tree savanna as forest (13%). Mixes of forest with nonforest types produce intermediate vegetation classes that are difficult to uniquely identify; these usually occur in transitional areas where there are often gradients in cover type. The overestimation of forest by GAC NDVI classification procedures seems independent of the dominant cover types, although our analyses with MSS data did not investigate the effect of size or shape, level of fragmentation, nor distribution of the cover types within the MSS images. For example, in two forest to savanna transition scenes, MSS was 37.9% forest, 56.5% non-forest to GAC 40% forest and 60% non-forest. In a second MSS scene, forest was 77.7% to 15.7% non-forest to GAC 100% forest (2% to 22.3% overestimation of forest). In two MSS scenes from humid forest areas, the predominant classes were 95.1% forest and 4.4% regrowth which was classified as 100% GAC forest, and in another validation scene, 86% forest and 11.6% regrowth to GAC 98% forest and 2% non-forest.
Most of the problems with GAC NDVI classification would be moot given higher resolution data and the use of seasonal NDVI plus reflective or thermal bands. Such a dataset could in future being produced for the IGBP by the U.S.G.S. EROS Data Center.
Objective 3. Mapping the areal extent, distribution and timing of fires.
The approach adopted for this objective used two existing data bases to derive very different products. The 1 km AVHRR Local Area Coverage [LAC] for southern Africa was used to obtain the frequency and timing of active fires occurring in 1989 and 1992. The fire detection technique uses data from the middle and thermal infrared. Since this data only shows active fire fronts at the time of satellite overflight, an additional assessment of the total areal extent of burning was provided by analysis of the multiyear GAC NDVI time series. Validation of both of these products has been carried out using Landsat MSS data. Fire validation has been part of the Cameroon and UCL collaboration.
The AVHRR l km active fires show the timing and distribution of burning within the region. (click here for pictures) Issues associated with the delectability of fires using this data set and the diurnal sampling of fires are discussed in Justice et al. (1991) and (1995). There is a general trend in the evolution of the fire season from north west to south east across southern Africa. The peak burning within the central African region for 1989 was shown to occur from late June to early August (southern hemisphere), although preliminary analysis shows considerable inter annual variability in the timing of fires. The boundary between the dense humid forest and savanna is clearly demarcated by the fire distribution. The most extensive areas of burning occur in southern Bandundu, Kasai and Shaba Provinces. The GAC NDVI bum analysis was based on a simple methodology to separate burned areas from those of only sparse vegetation and low annual NDVI that were unburned. An unsupervised classification was made of the minimum NDVI time series for a single dry season. By inspection of the results we determined that a minimum change threshold could be used to differentiate between areas of low vegetation cover (low NDVI) and actual burned areas. Separate dry seasons were used in northern (1986) and southern (1989) portions of Africa for this analysis to match the validation data base and field validation efforts of collaborators. The cycle of burning with subsequent regrowth was considered in the discrimination.
The monthly GAC NDVI burn classifications for southern Africa 1989 show that greatest extent of fires (over 20,000 km2) occurs in the wet savanna/forest transition areas, followed by areas in forest to savanna transition (14,500+ km2), dry savanna (14,196 km2) and the wet/dry savanna transition areas (14,047 km2). Although we do not know the sensitivity of the GAC pixel to actual burned area, we know from comparison to MSS images that only large-scale uncontrolled savanna burns can be detected using GAC data. Small agricultural clearing fires are too limited in area for routine detection. Ancillary datasets proved useful in the analysis of burn extent. The distribution of ranches (from Atlas Jeune Afrique) has a strong positive relationship to burned areas for several areas like the Oriental Kassai (in Lulua, Yangweshi and Mazia Mpota; and in Kabinda, Kambaye and Kamiji) and Shaba (Haut Lomani Mitsha, Luniemu and Kyabukwa). Maps of population density (Atlas Jeune Afrique) were also overlaid with burn extent. In regions with high population densities fires are reduced, perhaps due to the prevalence of subsistence agriculture that is associated with small fields and small fires. This pattern can be contrasted to more pastoral or low population regions where fire use is much less constrained and burns tend to cover very large areas. Wet savanna is more likely to be burned than dry savannas due in great part to lower fertility and the relative low population densities.
Other population and land management characteristics can be seen in the pattern of burns. North of the Equator, the largest burns are found in the wet savanna regions of the Central African Republic. This is an area where some of the largest fires ever reported have occurred. Other large fires have been reported for the wet savanna of Guinea (Malingreau and Laporte, 1989). The recurrent droughts of the last few years have strongly affected the wet savanna. Drought has enabled the Bororo nomadic herdsmen to burn large areas of wet savanna to improve grazing, thus putting them in conflict with local agricultural communities. Future analysis of fire scars would be helped by the availability of l km time series data for the region. The future research and development could explore the development of a hybrid burn detection methodology detecting both active fires and burn scars. Estimates of the areal extent of fires can be used to provide improved estimates of trace gas emissions from burning.
Objective 4. Development of indicators of anthropogenic pressure (identification of areas potentially sensitive to land cover change).
This effort has used several datasets from diverse sources to try to identify 'hot spots' susceptible to near-term human disturbance. Datasets used include the regional maps of burns, vegetation type, climatic data and ancillary data indicative of population density such as census data and road networks. (click here for pictures) These have been combined using GIS techniques to explore and model the relationships between variables and their utility in gauging anthropogenic pressure.
Many factors affect human pressure and impact. We have defined the most critical of these and examined model sensitivities to those factors. As access is one of the greatest controls on human activities, we focused attention on the relationship between roads and forest degradation. We have used the long-term GAC dataset throughout this study in the absence of a higher resolution dataset offering regional, cloud-free composites. The major drawback of this data [see Objective 2 results above] is the lack of sufficient resolution to observe changes in secondary forest or deforestation areas which would be a direct indicator of anthropogenic pressure. Since disturbance cannot be directly observed in these coarse resolution datasets, we have sought to develop and evaluate appropriate model surrogates.
Results/Discussion
This component focused on understanding the factors driving forest disturbance and forest or savanna conversion to agricultural use. Several indicators were developed from both remote sensing and GIS data layers. We first derived what we considered to be an appropriate indicator of access. To simulate accessible areas along roads we created a 1 km buffer around all 'roads' from the Central African portion of the Digital Chart of the World. This area was then divided into 30' X 30' gridcells, and the fraction of buffered road area in each gridcell was used as a road index. A fragmentation index was also derived from the LAC derived forest maps by calculating the ratio of forest patch perimeter to area within each of the 30'X 30' gridcells. This ratio increases as patches become smaller and the edges more complex.
For development and validation of the disturbance model we used the LAC AVHRR derived forest maps from the previous BSP project (BSP, 1993). These maps do not cover the whole of Central Africa, but they do show degraded (disturbed) forest areas at their 1 X 1 km moderate pixel resolution for Zaire and Cameroon. Selective logging activities are not visible at this resolution, but the disturbance of traditional agriculture, towns, villages, and main roads are clearly depicted.
Regression analysis was then used to compare the road index to the areas of degraded forest in the LAC datasets for Zaire and Cameroon. Other variables such as population density derived from available census data were also incorporated at this step. Once the relationships were described locally for the Zaire and Cameroon LAC maps, then they were applied to the GAC Central Africa cover map to flag areas of high potential for anthropogenic change and to gauge the impacts of human presence by vegetation type.
Some of the basic conclusions of this work show a strong positive relationship between the road index and the area of degraded forest in any one gridcell (r2 = 0.869). Not surprisingly, the road index is negatively correlated to forest cover. Forest fragmentation has only a moderate correlation to either forest or degraded forest cover (r2 = 0.7). The correlation between road index and forest fragmentation is even lower (r2 = 0.569). However, forest is highly fragmented in areas of high population density, probably reflecting short fallow periods and little forest regrowth. The coarse resolution of the GAC dataset may mask the smaller forest patches, reducing the validity of this comparison.
We found that in the absence of high resolution images, a simple linear regression model can be used to predict the percentage of degraded forest. The most reliable model predictor was the road index, which is a rough estimate of human presence and regional accessibility. Results using population data are much less clear. This is due to their high spatial aggregation and low frequency of observation. Higher quality population data are needed. It was noted with some surprise that some of the predicted 'hot spots' of potential human pressure showed no disturbance whatsoever when compared to high resolution images. By examining high resolution Landsat data we found that in many of these cases the road network from the DCW was in error, that either the roads had never been built or had fallen in disuse since the DCW base maps were made. A more current set of road maps would give better results as would a more current and less aggregated population dataset.
Of course, much of this indirect estimation of population-driven 'hot spots' susceptible to change would be unnecessary if higher resolution, frequent regional satellite images were available regionally. With higher resolutions, for example with the l km AVHRR data, change and the location of human impacts could be directly located once a baseline dataset high resolution data set was acquired. Current efforts by the IGBP have begun the collection of a global, l km dataset and this dataset could be directly applied as a baseline for future 'hot spot' identification.
Objective 5. Development of a Central Africa GIS.
There were two basic goals to this objective; the collection of all pertinent spatial data into a single, co-registered digital database system for synergistic data analysis and modeling in this project, and the collation and documentation of this GIS dataset for transfer to the Biodiversity Support Program for use in other projects and for distribution. Major activities undertaken to achieve these objectives included database design, data acquisition from a wide variety of sources, digitizing, data cleaning and editing, quality control, data preprocessing to common spatial projection system, and production of new spatial information by combination and modeling using the base data layers and analysis/integration of satellite datasets. The resultant data system had to be flexible enough to integrate raster, vector and tabular datasets.
Results/Discussion - Now at the conclusion of this project we have produced a Central African Geographic Information System using the widely known ESRI Arc Info GIS environment. A description of the GIS contents, methodologies used in data acquisition and integration, sources and directory structure have also been produced to aid users. The GIS is composed of base maps from many sources, plus the major derived layers from our analysis of satellite data and modeled products from the integration of various data layers into new information.
The simple structure of this dataset belies the many difficulties inherent in its construction. Some of the primary and most time consuming tasks were locating, acquiring and converting existing Central African datasets into a format compatible to our GIS. These datasets were widely dispersed and often in formats that were extremely difficult to ingest. Even the transfer of data between apparently compatible systems was difficult due to varied format standards and system-to-system irregularities with tapes, floppy disks, and file structure. Once data were collected and delivered to the University of Maryland, they had to be identified, converted to digital format when required (digitization of paper maps), methodologies documented, and quality controlled. Then adjustments had to be made to assure identical map projection and suitable georeferencing. Once the data was in-house and on the system, further manipulation was required to insure commonality of attribute descriptions (for relational database manipulation), and to create any legends or color codes required for display purposes. The difficulties of GIS assembly are not to be underestimated.
Based on the experience of assembling the GIS for this project, we can make a few observations and recommendations. Although GIS is a rapidly growing field, ever gaining in importance for international development, planning, global change, ecological and climatological research, it is still in its infancy. Consequently very few sites are prepared for the technical effort and sophistication required to manage and analyze data to its full potential. We recommended that the Central Africa GIS work be continued at some archival site where new data could be entered, quality controlled, documented, analyses carried out to meet the needs of the user community, and new datasets be created. Such a site could also respond to users' needs for new data, such as searching for more existing data sets to add to the GIS, updating existing data sets such as rainfall to fill in gaps of missing information, addition of field notes and results from biodiversity and ecological studies, and addition of new data sets from results of remote sensing interpretation. These results and other GIS data layers could be made available to the research community through a data sharing network or through a metadata browse and physical [not network] distribution of data.
It is not prudent to recommended Unix-based workstations for GIS use in most Central African countries at this time. PC-based GIS's are more appropriate since PC's are already in use and are supported by a substantial and continuously growing user community in the region. Workstations are much more expensive and complex, requiring considerable systems level support both in hardware and software maintenance and in general system administration. In instances where the PC's can not handle a task, ties with outside institutions or a highly capable archival site would allow both analysis and data generation capabilities.
Development of the GIS library/archive site concept as described above should be continued. Data sharing is the best way to reduce the costly data acquisition phase of projects. As a means of establishing a data sharing environment, electronic file transfer [ftp] sites could be set up for data transfer and production of other media, such as CD-ROM or a tape format compatible across platforms, explored for use as a portable library or for data interchange. Methods should be investigated for effective communication with GIS installations on the continent of Africa. Expensive international phone communications and lack of continental networks in Africa hamper GIS expansion and the spread of knowledge both within the countries of Africa and with the rest of the world. Electronic bulletin boards set up around use of telephone communication at off-peak hours could foster electronic contacts, information sharing, and problem solving. A key part of information sharing is the basic communication of what data is available. Before data is requested, an appropriate description, or metadata listing, should be available for potential users to decide if a particular dataset is really useful. Metadata standards and initial listings could go a long way towards beginning a text based information sharing network.
IV. Research Conclusions and Recommendations
This project has demonstrated the power of combining datasets from a variety of sources in spatial models for environmental assessment, planning and management. The synergistic use of climate, population, and data layers derived from remotely-sensed imagery in a geographic information system context is promising for a wide range of disciplines and applications; seemingly bounded only by access to appropriate information collected at suitable scales and frequency.
The results of this project show that long-term, coarse resolution greenness indices [monthly composite GAC NDVI in this project] when combined with sample high resolution data can be used to effectively classify savanna and seasonal forest vegetation types at the regional scale. Although the coarse resolution of the GAC datasets is inadequate to observe forest disturbance directly, a modeling approach can be successfully used in conjunction with the vegetation cover data to delineate areas of strong potential for human disturbance. Models using even simple road networks have good predictive power for forest disturbance as validated using moderate resolution data (LAC AVHRR forest maps).
Other datasets, such as population density, could also be of use in such model efforts, but their resolution and spatial aggregation are too great for effective use even with coarse resolution satellite-based data. A new paradigm of data collection and presentation for census data could be designed to better serve any and all communities that need spatially explicit population parameters. Use of widely available global positioning system technologies in tandem with a GIS approach could pinpoint population centers with a precision useful to practically all users. Even though good results were found using a road index as an indicator of population pressure, the existing road datasets are highly inaccurate and out of date.
Fires and burn areas can be located with both coarse and medium resolution satellite sensors, as long as their frequency of observation is greater than the regrowth rates of burned areas. There remains uncertainty regarding the areal extent of fire or burning required to 'flag' an entire GAC pixel as either burned or unburned. Further work relating to these scaling functions is required before precise areal assessments can be made with the AVHRR type sensor systems. However the accuracy of the current methods is adequate for regional studies showing the extent and timing of burning.
Many of the problems with coarse resolution satellite datasets could be solved by slight improvements in spatial resolution. The LAC AVHRR dataset is superior to GAC and is currently being acquired globally as a 10 day composite product under the auspices of the IGBP. We anticipate the direct observation of human impacts to be possible with the 1 km LAC dataset, once this newly compiled baseline data is made available.
Regional geographic information systems are indispensable for the integrated research undertaken in this study. The effort required to build and maintain a GIS is cost prohibitive for small users. Data, knowledge, and cost sharing will be required to build and maintain the most comprehensive GIS archives. A model of a central archive serving a distributed user community with appropriate datasets and analysis could be effective in getting the most information to the widest possible user community. The costs and overhead of such a service and archive activity should not be underestimated, but it is much less expensive than duplicating data base activities and creating many small centers with limited data and analytical capacity. The concept and structure of a data library for Central Africa was developed as part of this research activity.
BIBLIOGRAPHY
Biodiversity Support Program. 1993. Central Africa. Global Climate Change and Development. Technical Report, Overview and Synopsis (3 vols.) Biodiversity Support Program, c/o World Wildlife Fund, Washington, DC
Booth, T.H. 1991. A climatic/edaphic database and plant growth modeling system for Africa. Ecol. Modeling 56:127-134.
Booth, T.H. 1993. Nix interpolated rainfall dataset. Data provided by T.H. Booth, CSIRO Division of Forestry, PO Box 4008 Queen Victoria Terrace, Canberra
Boulvert, Y. 1986. Carte phytogeographique de la Republique Centrafricaine au 1:1,000,000. Edition de l'ORSTOM. Note explicative no. 104., Paris, France
Boulvert Y. 1990. Avanc ou recul de la fort central fricaine. Changements climatiques, influence de I'homme et notement des feux. p 353-366. Dans Paysages quaternaires de I'Afrique Centrale Atlantique. Raymond Lafranchiet, Dominique Schwartz (eds). ORSTOM-1990ISBN:2-7099-1022-5.
Devred, R. 1958. La vegetation forestry du Congo Belge et du Ruanda-Urundi. Bulletin de la Societe Royale de Botanie Belgique 65, 4:09-468.
Hutchinson, M.F. and R.J. Bischof. 1983. A new method for estimating the spatial distribution of mean seasonal and annual rainfall applied to the Hunter Valley, New South Wales. Australian Meteorological Magazine 31:179-181.
Justice, C.O., Townshend, J.R.G., Holben, B.N., and Tucker, C.J. 1985. Analysis of the phenology of global vegetation using meteorological satellite data. International journal of Remote Sensing, 6:1271-1381.
Justice, C., Horning, N., Laporte, N. 1993. Remote sensing and GIS contributions to a climate change program for central Africa 1993 - in Central Africa global climate change and development. Technical report -part III. Biodiversity Support Program, c/o WWF, Washington DC. 33 p.
Justice C.O., J.P.Malingreau and A. Setzer (1993). Satellite remote sensing of fires : potential and limitation. In Crutzen P. and J. Goldammer (Ed) Fire In the Environment; Its Ecological, Climatic and Atmospheric Chemical Importance, John Wiley and Sons, Chichester.
Justice C.O., Kendall J., Dowty P.R. and Scholes R.J., (1995). Satellite remote sensing of fires during the SAFARI Campaign using AVHRR data. journal of Geophysical Research (in press)
Koechlin, J. 1961. La vegetation des savanes dans le sud de la Republique du Congo. ORSTOM, Brazzaville, Congo.
Laporte, N., Justice, C.O.J., Kendall, J., 1994. Mapping the dense humid forest of Cameroon and Zaire using AVHRR satellite data. Int. J. Remote Sensing (in press).
Letouzey, R. 1985. Carte phytogographique du Cameroon au 1/500 000 - Ed. Institut de la Carte Internationale de la Vegetation 39, alles Jules Guesde 31400 Toulouse-France.
Malingreau, J.P., N. Laporte and J.M. Gregoire 1989. Exceptional fire event in the tropics: Southern Guinea January 1987. Int. J. Remote Sensing 12:2121-2123.
Nicholson, Sharon E., Jeeyoung Kim and Jon Hoopingamer, 1988. Atlas of African rainfall and its interannual variability. Department of Meteorology, Florida State University, Tallahassee, FL 32306.
Peron, Yves (ed). 1975. Les atlas jeune Afrique. Editions J.A., Paris, France. 47 p.
Prince, S. 1991, Satellite remote sensing of primary production: comparison of results for sahelian grassland 1981-1988-Special issue - Coarse resolution remote sensing of Sahelian environment- Int. J. of Remote Sensing, 12(.6), 1301-1311.
Townshend, J.R.G. (ed). 1992. Improved global data for land applications. A proposal for new high resolution data set. Report #20, International Geosphere Biosphere Program, Paris, France. 87 p.
Tucker, C.J. 1979. Red and photographic infrared combinations for monitoring vegetation. Remote Sensing of Environment 8:127-150.
Tucker C.J, Vanpraet, C., Sharman, M.J, Vanfnttersum, G. 1985. Satellite remote sensing of total herbaceous biomass production in the Senegalese Sahel: 1980-1984. Remote Sensing of Environment 17:233-249.
World Conservation Monitoring Centre. 1993. Conservation Atlas of Africa. Cambridge.
White, F. 1983. The vegetation of Africa (scale 1:5 million). United Nations UNESCO, LaChaux-de-fonds, Switzerland.
 |
WHY BIODIVERSITY FIELD STORIES PARTICIPATE RESULTS MARKETSPACE ABOUT THE NETWORK WHAT'S NEW LEARNING MATERIALS SEARCH LINKS SITE MAP HOME |