Geoscience Reference
In-Depth Information
probabilities are then summarized by cover type and region to convey information to users about
the quality of the classification.
A form of design-based inference is used in the preparation of a confusion matrix taken from
the classification of training sites. In this process, all training sites are divided into five equal sets.
The classifier is trained using four of the five sets and then classifies the unseen sites in the fifth
set. This procedure is repeated for each set as the unseen set, yielding a pair of labels — “true”
and “as classified” — for every training site. Cross-tabulation of the two labels for the training site
collection yields a confusion matrix that provides estimates of global, user's, and producer's
accuracies. This matrix is provided in the documentation of the LC data product.
Note that these estimated accuracies will be biased because the training sites are not chosen
randomly, and thus they may not properly reflect the variance encountered across the full extent
of the true LC class. However, in selection of training sites, every effort is made to identify sites
that do reflect the full range of variance of each class. Accordingly, the accuracies obtained are
thought to be reasonable characterizations of the true accuracies, even though they cannot be shown
to be proper unbiased estimators. In a final application of design-based inference, the MODIS team
plans to conduct a random stratified sample of its LC product at regular intervals. The methodology
will be similar to that of the IGBP DISCover validation effort (see section 3.1.3). However, funds
have not yet been secured to support this costly endeavor.
3.4 CEOS LAND PRODUCT VALIDATION SUBGROUP
The lessons learned from previous and ongoing projects point to several areas where LPV can
help with validation efforts. Perhaps most fundamental is that CEOS/WGCV/LPV provides a forum
to discuss these issues and develop and maintain a standardized protocol. Indeed, the authors are
all involved with LPV, and it was through this association that this chapter was developed. There
is also the opportunity to communicate on LC classification systems; although each project will
have its own system, coordination between the two projects results in synergy between the two
systems (Thomlinson et al., 1999; Di Gregorio and Jansen, 2000). Here we present methods by
which LPV can help address the specific lessons learned from IGBP in the context of the two
current projects. Table 3.1 lists the various subgroups and their corresponding URLs.
Table 3.1
CEOS Land-Cover Validation Participants and Contributions
Entity
Role in Global Land-Cover Validation
CEOS
Working Group on Calibration and Validation Land Product
Validation subgroup
http://www.wgcvceos.org/
Coordinates validation activities of CEOS
members
Global Observation of Forest Cover
Global Observation of Land Dynamics
http://www.fao.org/gtos/gofc-gold/index.html
Coordinates regional networks to provide
“local” expertice
European Commission's Global Land Cover 2000
http://www.gvm.sai.jrc.it/glc2000/defaultGLC2000.htm
Produces data
NASA's Global Land Cover product
http://edcdaac.usgs.gov/modis/mod12q1.html
Produces data
EOS Land Validation Core Sites
http://modis.gsfc.nasa.gov/MODIS/LAND/VAL/CEOS_WGCV/
lai_intercomp.html
Sites under consideration for CEOS Land
Product Validation Core Sites
VALERI (VAlidation of Land European Remote sensing
Instruments)
http://www.avignon.inra.fr/valeri/
Sites under consideration for CEOS Land
Product Validation Core Sites
CEOS “LAI-intercomparison”
http://landval.gsfc.nasa.gov/LPVS/BIO/lai_intercomp.html
Sites under consideration for CEOS Land
Product Validation Core Sites
