Vol. XXXVIII, Part 7B
In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
A., 2009. "Recent
rie and Radiometric
Л in EARSeL 6
4 from stereo-images
mples with VIR and
; Sensing, vol. 4 (, no.
1993. “Comparative
ctral detectors,” IEEE
A.M., and Stotts, L.B.
and recognition in
:ection and estimation
6, 143-156
COMPARING INFORMATION DERIVED FROM GLOBAL LAND COVER DATASETS
WITH LANDSAT IMAGERY FOR THE HUAMBO PROVINCE AND GUINEA-BISSAU
A. Cabral*, M. Vasconcelos, D. Oom
Tropical Research Institute, GeoDes, Travessa Conde da Ribeira, n°9, 1300 Lisboa, Portugal -
(anaicabral70, maria.perestrelo, duarte.oom)@gmail.com
KEY WORDS: Classification, High resolution, Land cover, Multitemporal, Huambo, Guinea-Bissau
ABSTRACT:
Land cover maps, derived from satellite data, are a valuable tool for various global research studies and are often used in multi
temporal approaches to document the dynamics of processes such as agricultural expansion or deforestation. In this study we show
how the observed land cover change tendencies diverge widely depending on the scale of observation and on the characteristics of
the data sources used. For the analysis we compared land cover changes using two different scale map time-series in the period 1990
- 2009. Two regions were selected, for which there are high resolution imagery and/or ground data available for validation and
verification purposes: the entire country of Guinea-Bissau and the Huambo province in Angola. The first map time series consists of
data available in international projects (IGBP, GLC2000, and MODIS) obtained from classification of 1 Km resolution imagery for
three dates in the study period. The second map-set results from classification and validation of 30 meter resolution images (Landsat
TM and ETM+), covering the same area in approximately the same dates. For the comparisons, the different map legends had to be
aggregated into a common nomenclature to define five common classes: Forests, Savannas/Shrublands, Grasslands, Croplands/Bare
soil and Wetland. The results show large discrepancies in the observed trends in agricultural areas. For example for both regions, the
increase in agricultural land during the analyzed period, which is observed in high resolution maps and confirmed by validation and
field knowledge, is lost in the coarse resolution maps. The deforestation rates reported by the coarse resolution maps are not verified
when high resolution is employed. The consequences of these observations are discussed and future work proposed.
1. INTRODUCTION
In the last years, land cover mapping has become one of the
most important sources of information for environmental
studies. This type of information becomes even more relevant
with the establishment of international agreements such as the
Kyoto Protocol, the International Convention on Biological
Diversity, and the framework Convention on Climate Change,
all of which call for accurate reporting of environmental
variables (McCallum et al., 2006). Having information about
land cover status is essential, as a baseline, in order to evaluate
future changes. Remote sensing data from several satellites
allowed obtaining sufficiently accurate land cover mapping in a
global scale, evenly in remote areas, and has been used to derive
several global land cover maps, that are freely available for a
variety of applications, and which are deemed sufficiently
accurate for different project types. The use of these land cover
maps has been very useful in modelling studies and corresponds
to a great advance in earth system science. Since these maps are
developed by different and independent national and
international initiatives, they were prepared using different data
sources, classification systems and methodologies, which are a
reflection of the different mapping standards adopted and varied
interests. As a consequence, each dataset has some advantages
and limitations and it is important to fully understand their
applicability bounds. One way to do it is by comparison among
different data sets and scales of analysis. This approach helps to
better grasp what data sets should be used for monitoring,
compliance assessment, and trend analysis. Several efforts have
been made in recent years to improve the comparability and
compatibility between land cover datasets. GOFC-GOLD
(Global Observation of Forest and Land Cover Dynamics) in
conjunction with FAO (Food and Agricultural Organization)
and GTOS (Global Terrestrial Observation System) developed a
new Land Cover Classification System (LCCS) in order to
obtain a land cover harmonization methodology (Herold et al.,
2008; Jung et al., 2006). The LCCS allows that land cover
features be defined at any scale or level of detail, with an
absolute level of standardization of class definitions between
different users (Di Gregorio and Jensen, 2000). Several studies
comparing two or more global land cover products were done at
regional (Kalacska et al., 2008) to global scale, which show
significant disagreements and reveal uncertainties (Giri et al.,
2005; Herold et al., 2008; McCallum et al., 2006). Therefore, a
validation and a comparison of these global datasets are
necessary before using them in global and regional studies.
Different approaches are used to quantitatively estimate the
accuracies of the global land cover classifications: confidence
values of the classifier, comparison with other maps, cross-
validation with training datasets and statistically robust spatial
sampling and acquisition of ground reference information (Jung
et al., 2006). The purpose of this study is two-fold: (1) to
characterize land cover change, with a special focus on
deforestation, in two approximately same size regions of Africa;
(2) to assess the effect of using coarser resolution global land
cover maps for producing the same information in the period of
1990 - 2009. To achieve these goals two different map time
series are used. The first map time-series results from
classification and validation of 30 meter resolution images
(Landsat TM and ETM+). The second consists of data available
in international initiatives (IGBP, GLC2000 and MODIS). The
analysis is performed for the entire country of Guinea-Bissau
* Corresponding author.
107