BREC

This integrated suite of software tools is specifically developed for the analysis of urban scenes. Starting from high-resolution radar or optical data the system, named BREC, allows to obtain a representation of the principal objects in the scene, including a GIS interface.

The tools exploit both spatial and spectral information contained in the data and the system extracts and extensively processes the linear features. These latter are progressively defined by the context as road network elements or as the contours of a natural or artificial object. Finally, simple linear features may be connected in higher-level elements using geometric criteria, based on e.g. perceptual grouping techniques.

THE BREC SOFTWARE SUITE

The BREC system includes an integrated set of tools which, starting from different types of remotely sensed images, allows:

  • To classify the land cover, revealing the presence of basic classes such as roads, buildings, vegetation, water, etc.;
  • To extract the basic components required for the recognition of objects, e. g. linear segments and intersections;
  • To combine them intelligently, taking advantage of any additional information in order to highlight possible artificial objects (roads, buildings, bridges ...);
  • To examine the characteristics of the objects in the scene under investigation;
  • To compare the extracted objects in two or more scenes for change analysis.

The program is organized in a modular fashion.

The minimum level is the extraction of basic components, which are then organized into progressively more complex elements up to the identification of the final objects, removing errors and adjusting and refining the result at each step. This enables the user to enjoy the following benefits:

1) Usability of different types of data (e.g. optical, hyperspectral, radar) by switching only the element that accesses the image, while for the rest of the chain the type of data is transparent (flexibility of the source);

2) Possibility of identifying different types of objects (e. g. roads, rather than buildings). These objects have many features in common based on elementary geometric elements. The relationships among these features may help to differentiate different objects at an appropriate level of aggregation, but they till share process at lower levels (flexibility of the aim, efficiency of treatment);

3) Possibility of further development without excessive interference with the existing functions, for example by introducing new geometric primitives, or new objects to be identified, such as piers and boats (expendability);

4) Ability to accept different standards for there presentation of input data and a similar ability to represent different formats of the output data (compatibility, interoperability).

The current phase of development has reached a fully functioning program that extracts and tests satisfactorily building footprints and road networks from aerial images and satellite in urban areas, saving the results into GIS standard format files. The suite addresses the main use of remotely sensed data in urban areas, i.e. human settlement extent identification, land use and land cover mapping inside urban areas, building detection and extraction, with both bi-dimensional and three-dimensional information extraction, road network extraction and urban vegetation mapping. All of these routines are based on techniques developed and published in technical literature.

Example of tool: Building Extraction

The approach followed by BREC for building extraction may follow two different paths:

On the one hand, classification or segmentation followed by contour regularization can be applied.

On the other hand, object recognition through extraction of linear features, fusion into higher level structures and finally footprint extraction has been also implemented.

According to the experience of the authors, a detailed analysis of an urban scene cannot be obtained without considering spatial information. Although spectral and spatial resolution are important, the capability to discriminate among different land covers and, even more important, among different land uses in urban areas relies heavily on the context. That’s why the classification algorithms implemented into BREC always include a spatial element. The basic classifier is the fuzzy ARTMAP neural network, suitable for both optical and SAR data and especially for pixel level data fusion, e. g. to include textural features as additional classification bands. This non-parametric classifier provides a post-processing step able to consider the context by analyzing the classification patterns in the training areas. By learning the association between correct classification classes and surrounding classification pattern the system is able to reduce “salt-and-pepper” classification noise and make the final map more homogeneous. An alternative classifier considered in BREC is a Markov Random Field classifier incorporating in formation about edges in the data. The general idea of the MRF classifier is to take into account the local patterns and minimize a functional which is based on both pixel and context information.

The additional step is that the neighbourhood used to compute the context contribution to the functional are adaptive, taking into account the edges in the scene. Where an edge exists, it is unlikely that correlation patterns are available across it, while they are very likely along it. The same stress on context information is important for the alternative approach to building extraction outlined above. This methodology, labelled “bottom-up analysis” has been implemented starting from linear feature extraction. Indeed, one of the simplest yet efficient geometrical features that one can exploit in human settlements or, generally speaking, when artificial structures are of concern are linear features. They can be used as a basis for higher-level feature extraction, such as parallel pairs or junctions, and combined to obtain even more complex elements, like geometric shapes. They are very flexible elements of the scene, and can be used for a variety of applications, provided an efficient methodology for their extraction is considered, capable of dealing with different VHR images, such as SAR and optical ones. After this step, higher-level feature extraction needs to be at the same time fast and reliable, building over the feature extraction in an application-driven manner. Finally, the ability to manage these spatial features in conjunction with spectral ones must be taken into account, and an eye must be kept on pixel-based measures related to spatial information. The choice of focussing on “segments” in many research works is a result of balancing between advantages and disadvantages. The most important advantage of segments is the (relative) simplicity of the extraction routine, as well as the high flexibility in combining them into various and very different scene elements. The disadvantages are related to problems in discriminating segments belonging to geometrically very similar yet semantically completely different elements of the scene, and the high level of false positives and/or false negatives that any segment extractor produces. The example of higher-level feature extraction starting from linear features which is useful here is the recognition of manmade (artificial) objects in urban areas. To this aim, the idea is to design an algorithm to find, within a segment set, specific combinations suggesting the existence of rectangular structures. The routine is based on two different ideas to aggregate segments and junctions into objects. First of all, it is possible to perform a search for rectangles as combinations of parallel segment pairs. A second option is however to perform the same search looking for combinations of “L-shaped” junctions.

Accuracy results

The accuracy of the results where objects have similar spectral response can be significantly improved by performing an optional segmentation operation. By choosing to perform segmentation, however, we trade a longer processing time in exchange for more accurate results. Much of the additional time is due to the manual action of selecting filters and morphological operation to be performed. The selection can be skipped if one accepts to use the defaults and thus a generally sub-optimal solution. In this latter case the accuracy is to be expected a bit lower, but a significant manpower saving is achieved. Once the manual or automatic (default) selection has been performed, the output segmented map is obtained in few minutes.

Data splitting

Segment extraction routines (not strictly necessary for building extraction) performed on images larger than 6000 × 6000 pixels will require data splitting, due to problems with system memory. For such images, the software activates an image splitting procedure which results in a series of partly overlapping sub-images. Each subimage is then analysed separately and all the results are merged together at a later stage. Each sub-image has a size of 1000 × 1000 pixels or less.

Image input format

The input should be provided as a raw image (ENVI-like) or bmp.

The output may be provided as:

  • BMP image for visualization purposes;
  • SHP files (Arc/info ASCII format) containing the extracted features.
Webpage: 

Not available

Manual: 

Not available

Forum: 

Not available

Publications: 

Gamba, P.; Dell'Acqua, F.; Lisini, G. “BREC: The Built-up area RECognition tool”, Joint Urban Remote Sensing Event, 2009 Year: 2009 , Page(s): 1 - 5