International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999

72

Syntactic

Semantic

models

models

User conjecture modelling

Information fission

Information aggregation

Fig. 4. Simplified diagram of an information fusion system. Heterogenous sources of information are splitted into

elementary sources: the information fission. The information aggregation completes the fusion process.

of information commensurability, e.g. estimated texture

features using co-occurence matrix are not comparable with

parameters of Markov random fields.

• Cluster model: image features or estimated physical

parameters have n-dimensional representations. Due to

observation noise or model approximations, the feature

space is not occupied homogeneously. Thus, another level of

information abstraction, and fusion at the same time, is the

type of feature grouping, i.e. the cluster models, and the

associated parameters. Clustering can represent information

only for one category of the features.

• Syntactic modelling: stochastic grammars are defined to

represent the order of joint spaces, image features and

models. The induced syntax can be interpreted as a

clustering of incommensurable information representations.

Thus, information in incommensurable representations can

be aggregated. In the case of fusion of physical parameters

extracted from different data types, the result of information

fusion is equivalent to a supervised classification, otherwise,

for the case of abstract features, the classification is

unsupervised.

• Semantic representation: due to the impossibility to

estimate the physical parameters, an unsupervised

classification of the data is produced. Augmentation with

meaning requires a higher level of abstraction. Prior

information in form of training datasets or expert knowledge

is used to create semantic networks. Thus, the observations

are labelled and the contextual meaning is defined.

The performance of information extraction depends critically on

the descriptive or predictive accuracy of the probabilistic model

employed. Accurate modelling typically requires high

dimensional and multi-scale modelling. For heterogeneous

sources, accuracy also depends on adaptation to local

characteristics.

6. DATA AND INFORMATION MINING

One of the difficulties in many data fusion applications for

remote sensing image classification is the identification of the

datasets relevant for the user. Here, the information retrieval task

becomes an information mining task. The Bayesian inference can

be applied again, but the goal is to find the set of images,

generally acquired from different sensors, which might contain

the information desired by the user. Thus, the inference is used to

find the posterior distribution over the set of available images

according to the user hypothesis in a given application.

In the present context, data mining will be understood as

searching for datasets with certain properties, and information

mining refers to the search of hidden relationships among image

features, either in one or in multiple image datasets. Data mining

will produce an index, information mining a label.

The concept and hierarchy of information representation allows

the integration of different tools for information mining. The goal

is to explore the information content of the images and decide

which ones are relevant for the user’s application. The tool we

propose is mainly based on interactive decision using data

visualization, e.g. image browsing integrated with image

segmentation, or information visualization, e.g. exploration of

the clusters’ structure in the feature space and the interactive

design of the decision surfaces. The block diagram of a simple

data and information mining tool is presented in Fig. 5.

Recently, solutions of the MAP estimator have been proposed

combined with an expectation-maximization and deterministic

annealing, in which entropy minimization maximizes the amount

of evidence supporting each parameter, while minimizing

uncertainty in the sufficient statistics and cross-entropy between

the model and the data. A similar approach was developed using

two levels of Bayesian inference, model fitting and model

selection, for the estimation of the hidden parameters of a 3-level

hierarchical model. These methods of inference can be integrated

in tools for information mining in the assumption of datasets

explained by multiple models that coexist together.