Geoscience Reference
In-Depth Information
The validation of the GLC 2000 products incorporated confidence building based on a com-
parison with ancillary data and quantitative accuracy assessment using a stratified random sampling
design and high-resolution sites. First, the draft products were reviewed by experts and compared
with reference data (thematic maps, satellite images, etc.). These quality controls met two important
objectives: (1) the elimination of macroscopic errors and (2) the improvement of the global
acceptance by the
200 km) was
systematically compared with reference material and documented in a database containing intrinsic
properties of the GLC 2000 map (thematic composition and spatial pattern) and identified errors
(wrong labels or limits).
This design-based inference had the objective of providing a statistical assessment of the
accuracy by class and was based on a comparison with high-resolution data interpretations. It was
characterized by: (1) random stratification by cover class, (2) a broad network of experts with local
knowledge, (3) a decentralized approach, (4) visual interpretation of the higher-resolution imagery,
and (5) interpretations based on the hierarchal classification scheme (Di Gregorio, 2000). Both the
confidence building and design-based components occurred sequentially. Confidence building
started with problematic areas (as expected by the map producer). This allowed for the correction
of macro-errors found during the check. Then, a systematic review of the product using the same
procedure was conducted before implementing the final quantitative accuracy assessment.
customers associated in the process. Each validation cell (200
¥
3.3 VALIDATION OF THE MODIS GLOBAL LAND-COVER PRODUCT
A team of researchers at Boston University currently produces a global LC product at 1-km
spatial resolution using data from the MODIS instrument (Friedl et al., 2002). The primary product
is a map of global LC using the IGBP classification scheme, which includes 17 classes that are
largely differentiated by the life-form of the dominant vegetation layer. Included with the product
is a confidence measure for each pixel as well as the second-most-likely class label. Input data are
MODIS surface reflectance obtained in seven spectral bands coupled with an enhanced vegetation
index product also derived from MODIS. These are obtained at 16-d intervals for each 1-km pixel.
The classification is carried out using a decision tree classifier operating on more than 1300 global
training sites identified from high-resolution data sources, primarily Landsat Thematic Mapper and
Enhanced Thematic Mapper Plus (ETM
). The product is produced at 3- to 6-mo intervals using
data from the prior 12-mo period (http://geography.bu.edu/landcover/userguidelc/intro.html).
The validation plan for the MODIS-derived LC product incorporates all approaches identified
in section 3.1.2. Confidence-building exercises are used to provide a document accompanying the
product that describes its strengths and weaknesses in qualitative terms for specific regions. A Web
site also accumulates comments from users, providing feedback on specific regions. Confidence-
building exercises also include comparisons with other datasets, including the Landsat Pathfinder
for the humid tropics, United Nation's Food and Agricultural Organization (FAO) forest resource
assessment, the European Union's Co-ordination of Information on the Environment (CORINE)
database of LC for Europe, and the U.S. interagency-sponsored Multi-Resolution Land Character-
istics (MRLC) database.
Model-based inference of classification accuracy is represented by the layer of per-pixel con-
fidence values, which quantifies the posterior probability of classification for each pixel. This
probability is first estimated by the classifier, which uses information on class signatures and
separability obtained during the building of the decision tree using boosting (Friedl et al., 2002)
to calculate the classification probability. This probability is then adjusted by three weighted prior
probabilities associated with (1) the global frequency of all classes taken from the prior product,
(2) the frequency of class types within the training set, and (3) the frequency of classes within a
200-
+
200-pixel moving window. The result is a posterior probability that merges present and
prior information and is used to assign the most likely class label to each pixel. The posterior
¥
