The open source project GeoAIDA integrates existing image analysis operators that are controlled by a-priori knowledge about the processed scene. The a-priori knowledge is modeled by a semantic net. The nodes in the semantic net represent object classes expected in the scene. The semantic net handles properties and relationships of different nodes. One property of a node could be the assignment of an image analysis operator to be used for the classification and/or extraction of a certain object class. If an image analysis operator for a particular task is not available, the node will be identified structurally by its components, i.e. the child nodes. As GeoAIDA transfers the classification and/or extraction task to external operators there is no limitation to the type of input images. Therefore, multi-sensor scene analysis is also possible. The design of GeoAIDA is shown in Figure 1.

Figure1 : GeoAIDA design.

On the input side the system consists of the components database and semantic net. Data processing is handled by top-down and bottom-up operators which are called by the system control unit. The core system control queries the image database, reads the semantic nets as well as project descriptions and generates classification/extraction results by calling top-down operators. The results are evaluated with bottom-up operators and once verified, stored as instance nets with corresponding label images, which describe the position of the related nodes.


The database provides all input information available for the scene interpretation. This includes images of different sensors, like optical, laserscan, or SAR, as well as GIS information or results of an earlier scene interpretation for multitemporal processing. GeoAIDA itself is not limited to any kind of input data - restrictions are only imposed by the attached external image analysis operators, which work on their dedicated input data.

Semantic net

The a-priori knowledge about the scene under investigation is stored in a semantic net. The nodes of the net are ordered in a strictly hierarchical manner, i.e. each node has exactly one superior node. The topmost node is the scene node. Attributes can be assigned to each node. Common attributes are name, class and the associated top-down and bottom-up operators. A top-down operator is capable of classifying and/or detecting objects of the node class in the given input data. For each object and each image analysis operator a result is generated. The bottom-up operator investigates the results generated by the subnodes and groups them into objects of the node class. These objects are represented by instance nodes.

Top-down operators

Top-down operators (s. Figure 2) are external image analysis operators that run a classification/extraction on given input image data and assign the resulting objects to one or more classes. Additionally, the operator is supplied with a binary mask which describes the areas of interest derived from the input GIS. Output of a top-down operator is a list of regions with a corresponding label image, which describes the position of the regions.

Figure 2: Operation of a top-down operator.
Bottom-up operators

Bottom-up operators are used to group a multitude of objects to a smaller quantity of superior objects, s. Figure 3. These operators are also implemented as external programs. Input of a bottom-up operator are the classification results with the corresponding label images, which describe the geometric position of the objects in the scene. The output is a list of instance nodes resulting from the grouping process and a new label image describing the superior objects.

Figure 3: Operation of a bottom-up operator.
Output of GeoAIDA

The output of the GeoAIDA analysis is an instance net, which describes all verified objects of the scene. The ordering of the nodes is strictly hierarchical, i.e. the footprint of inferior (child) nodes is always completely represented in the superior (parent) node. Furthermore, all nodes of the same hierarchic level are disjunct, allowing to describe the position of all objects of a whole instance tree in a two dimensional map.

System control

The main task of GeoAIDA itself is system control. Analysis is accomplished in two major steps. First a top-down pass calling the attached image analysis operators generates classification and/or extraction results about the objects detectable in the scene. According to the semantic net these results are structured in the net. The second step is a bottom-up progression through this net. During this pass an instance net is generated from the results of the nodes on the basis of object properties like size, structural relationship between neighboring objects, etc. The instance net together with the object map is the result of the two pass analysis.


Not available


Not available


Helmholz, P.; Becker, C.; Breitkopf, U.; Büschenfeld, T.; Busch, A.; Grünreich, D.; Heipke, C.; Müller, S.; Ostermann, J.; Pahl, M.; Vogt, K.; Ziems M.: Semiautomatic Quality Control of Topographic Reference Datasets. ISPRS Commission 4 Symposium Orlando, Florida, 2010.

Pahl, Martin: Ein wissensbasiertes System zur automatischen Extraktion von semantischen Informationen aus digitalen Fernerkundungsdaten, Dissertation, ibidem-Verlag, Stuttgart, 2003.

Bückner, Jürgen: Architektur eines wissensbasierten Systems zur Interpretation multisensorieller Fernerkundungsdaten, Dissertation, ibidem-Verlag, Stuttgart, 2003.