Abstract
The ability of space- and airborne instruments to measure the amount of electromagnetic radiation reflected and emitted by the Earth’s surface has proved to
be valuable for the understanding of our environment, as it provides for an
overwhelming flow of data on the appearance and condition of our planet. The data
yielded by
... read more
remote sensing can be subjected to various types of computer-assisted
manipulation, to arrive at derived data sets tailored to different types of application.
Computer-assisted classification of remotely sensed data into qualitative classes, for
example, is useful for extracting information that can be exploited for cartographic
purposes, such as in the generation of thematic maps of land cover types.
For a proper cartographic application, the fitness for use of a set of remotely sensed
data needs be assessed. The practicability of the data and their classification can be
established by means of an accuracy assessment procedure. An error matrix is created
for the classification by matching a random sample and its counterpart from a
reference data set representing the actual environment. Accuracy assessment based
on an error matrix, however, has several drawbacks. Among these is the non-spatial
and general character of a global statement like 95% accuracy for an entire
classification; moreover, accuracy assessment is a time-consuming and cost-intensive
process. As a consequence, it is easily omitted which, of course, is undesirable and
may lead to the use of data that are unfit for the application at hand.
For assessing the fitness for use of a set of remotely sensed data, accuracy is not the
only consideration. More generally, the phrase data quality is used to refer to the
extent to which the characteristics of the data meet the requirements of the
application aimed at by the user. A high quality indicates a relatively high information
value for the considered application - a good fitness for use. Uncertainty is a key-issue
in quality assessment and, therefore, in the assessment of fitness for use of a data set.
During the life cycle of remotely sensed data uncertainties are introduced and
propagated in an often unknown way. For investigating uncertainty, effective
measures need to be designed. To this end, it is relevant to consider the purpose to
which these measures are to be employed. Here, the focus is on an exploratory
perspective. Exploratory analysis of a set of remotely sensed data aims at acquiring
insight into the stability of various possible classifications of these data. For this
purpose, knowledge about the uncertainties underlying these classifications is
imperative. As in exploratory analysis, classification is an iterative process, needing
not only measures for assessing the uncertainty in a classification but also effective
ways to convey this information to the user. Visualisation is generally considered a
useful means of communication of potentially relevant information. In this thesis a
class of measures of uncertainty is presented, tailored to the purpose of exploratory
analysis of remotely sensed data, together with various ways of cartographic
visualisation of uncertainty.
The uncertainty that is introduced during classification of a set of remotely sensed
data is characterised by the probability vectors that are yielded as a by-product of most
probabilistic classification procedures. Here, emphasis is laid on maximum a?x
posteriori classifications where for every pixel in the data a vector of probabilities is
calculated that specifies for each distinguished class its probability of being the true
class. The probability vectors reflect the differences in uncertainty in the resulting
classification and can be stored in a gis to serve as a basis for the derivation of
weighted uncertainty measures such as entropy.
Besides the assessment of uncertainty, efforts can be aimed at the reduction of the
amount of uncertainty present in a remotely sensed data set. The maximum a
posteriori classification rules being dealt with in this thesis allow for the introduction
of a priori knowledge in the classification process, at different levels of sophistication -thereby
exceeding the simple approaches embraced in existing image processing
packages. Another strategy within the realm of dealing with spatial data uncertainty is
based on the idea of decision analysis that allows for an optimal decision-making given
uncertain information classes. Combining probability theory (defining the uncertainty
related to the occurrence of a particular class) and utility theory (defining the
desirability of the consequences resulting from the actions that are taken assuming
that particular class) contributes to the selection of the best decision under the given
conditions. This idea is particularly interesting when dealing with huge data sets
under uncertain circumstances and with far-reaching consequences for wrong
decisions (e.g. agricultural fraud detection by European Union).
Both the probabilistic results from the classification procedure and other quality
information are subjected to cartographic visualisation rules in order to develop a
framework for the communication of this spatial metadata. Static as well as more
dynamic approaches offer grips for the gis user who needs to consider simple but
persuasive maps to assess the fitness for use of a classification.
Commercial gis packages are still failing when the sound consideration of spatial
data uncertainty is at stake, a fact that has incited the participants of the camotius
project to look for the functionality of an “uncertainty-sensitive information system”.
Such a system is valuable for the Dutch situation in which the extra value added by
remotely sensed data is not always beyond all doubt; the explicit evaluation of these
data as well as their inherent uncertainty reveals their true information value. Two
case studies have stressed the role of remote sensing for planning purposes by
demonstrating its ability to monitor changes in the extent of greenhouses over space
and time, and making inventories of their area. The inclusion of uncertainty
information allows for an exploratory approach in which an appeal can be made to
several levels of knowledge in order to improve the processing results. It is stated that
a user will be encouraged to use remotely sensed data if their extra value is clearly
demonstrable. The components that have been scrutinised in the methodological part
of this thesis are formalised in a demonstration programme that could serve as a
blueprint for commercial gis packages. It can be downloaded from:
http://cartography.geog.uu.nl/research/phd
show less
