Geoscience Reference
In-Depth Information
a
f
()
u
an accuracy value
(i.e., for the classification of
Figure 11.4d). These accuracy values were mapped in Figure 11.5b. The accuracy map of Figure
11.5b exhibited much higher values than the corresponding map of Figure 11.5a, indicating an
increased confidence in classification due precisely to the consideration of contextual information.
In addition, the low accuracy values (~0.4-0.6) of Figure 11.5b were found near class boundaries,
as opposed to the low accuracy values of Figure 11.5a, which just corresponded to pixels classified
as shrub and rangeland. This latter characteristic implied that contextual information yielded a
more realistic map of classification accuracy, which could be useful for designing additional
sampling campaigns.
for the particular class reported at pixel
u
11.4 DISCUSSION
A geostatistical approach for mapping thematic classification uncertainty was presented in this
chapter. The spatial correlation of each class, as inferred from a set of training pixels, along with
the actual locations of these pixels, was used via indicator kriging to estimate the location-specific
probability that a pixel belongs to a certain class, given the spatial information contained in the
training pixels. The proposed approach for estimating the above preposterior probability accounted
for texture information via the corresponding indicator covariance model for each class, as well as
for the spatial proximity of each pixel to the training pixels after this proximity was discounted for
the spatial redundancy (clustering) of the training pixels. Space-derived preposterior probabilities
were merged via Bayes' rule with spectrally derived preposterior probabilities, the latter based on
the collocated vector of reflectance values at each pixel. The final (fused) posterior probabilities
accounted for both spectral and spatial information.
The performance of the proposed methods was evaluated via a case study that used realistically
simulated reflectance values. A subset of 0.14% (314) of the image pixels was retained as a training
set. The results indicated that the proposed method of context estimation, when coupled with
Bayesian integration, yielded more accurate classifications than the conventional maximum likeli-
hood classifier. More specifically, relative improvements of 10% and 34% were found for overall
accuracy and the Kappa coefficient. In addition, contextual information yielded more realistic
classification accuracy maps, whereby pixels with low accuracy values tended to coincide with
class boundaries.
11.5 CONCLUSIONS
The proposed geostatistical methodology constitutes a viable means for introducing contextual
information into the mapping of thematic classification uncertainty. Since the results presented in
the case study in this chapter appear promising, further research is required to evaluate the perfor-
mance of the proposed contextual classification and its use for mapping thematic classification
uncertainty over a variety of real-world data sets. In particular, issues pertaining to the type and
level of spatial correlation, the density of the training pixels, and their effects on the resulting
classification uncertainty maps should be investigated in greater detail.
In conclusion, we suggest that the final posterior probabilities of class occurrence be used in
a stochastic simulation framework, whereby multiple, alternative, synthetic representations of land
cover maps would be generated using various algorithms for simulating categorical variables
(Deutsch and Journel, 1998). These alternative representations would reproduce: (1) the observed
classes at the training pixels, (2) the class proportions, (3) the spatial correlation of each class
inferred from the training pixels, and (4) possible relationships with spectral or other ancillary
spatial information. The ensemble of simulated land-cover maps could be then used for error
