OBIA

Object-Based Image Analysis (OBIA) is a tentative name for a sub-discipline of GIScience devoted to partitioning remote sensing (RS) imagery into meaningful image-objects, and assessing their characteristics through spatial, spectral and temporal scale. At its most fundamental level, OBIA requires image segmentation, attribution, classification and the ability to query and link individual objects (a.k.a. segments) in space and time. In order to achieve this, OBIA incorporates knowledge from a vast array of disciplines involved in the generation and use of geographic information (GI). It is this unique focus on RS and GI that distinguishes OBIA from related disciplines such as Computer Vision and Biomedical Imaging, where outstanding research exists that may significantly contribute to OBIA.

A key objective of OBIA is to develop and use appropriate theory, methods and tools sufficient to replicate (and or exceed experienced) human interpretation of RS images in automated/semi-automated ways, that will result in increased repeatability and production, while reducing subjectivity, labour and time costs.

About this page
This wiki page is an initiatiave derived from the 1st International Conference on Object-based Image Analysis (, Salzburg, 4-5 July 2006). Its goal is to ensure that a common understanding of OBIA as a discipline is agreed and integrated within the research and commercial software that is being developed for object-based image analysis. This integrative effort can draw upon the strong theoretical and application based components that already exist in the established fields of remote sensing, computer vision, and landscape ecology among others. A way to achieve this is by developing a living document – an OBIA guide book – to which practitioners can turn to for understanding and direction. Wikis may be the ideal vehicle to develop such a device. As such, we cordially invite interested parties to contribute to this page so that OBIA may be developed from a user community and scientific perspective, rather than solely from commercial drivers. We intend to move this page to Wikipedia once it reaches enough content and quality.

This page is open to anyone interested to contribute. To edit this page, click on the 'edit' tab at the top of this page (you need to create an user account beforehand). If you only want to modify a particular section or subsection, click on the '[edit]' hyperlink to the right of the relevant header. When you are done, click the 'show preview' button at the bottom of this page to see the new appearance the page would have, and if it is OK click the 'save page' button. Further information can be found at How to edit a page. If you don't agree with the content of some item within the page, you may change or delete it, but please explain your reasons in the discussion page, which can be reached by clicking the 'discussion tab' on top of this page. Further guidance on modiying wiki pages can be found in Wiki Etiquette.

Brief history
The first RS spaceborne sensors were multispectral to compensate for the reduced spatial resolution of the data, with the implicit assumption that different landcover types would behave like distinct surface materials susceptible of being analysed with a spectrometric approach. Hence it was natural to treat each pixel as a sample introduced in a desktop spectrometer, and therefore the individual pixel has been considered the basic unit of analysis in RS since the beginning. Several digital classification methods were developed based on this approach and soon (even before the launch of Landsat-1 in 1972) they were established as the trusted common practice, becoming the accepted paradigm in the analysis of RS images. The fact that pixels are not isolated but knitted into an image full of spatial patterns was left out of the paradigm since it could only be exploited by human interpreters. OBIA has emerged as an alternative to this paradigm, and is based on the assumption that semi-automated object-based methods can emulate (or exceed) visual interpretation, making better use of spatial information implicit within RS images and providing greater integration with vector based GIS.

The ECHO classifier developed by David Langrebe's team at Purdue University in the 70s included a segmentation algorithm that can be viewed as the first OBIA precedent. The image was divided into 'cells' of 2x2 pixels that were subject to a simple test of statistical homogeneity. Cells failing the test were assumed to overlap a boundary and were later classified on a per-pixel basis. Adjacent cells passing the test were selected and subsequently subject to hypothesis testing for statistical similarity. Cells found similar were merged or annexed into regions. 'In this way an object can grow to its natural boundaries, whereupon either the cell selection or annexation test will fail'.

...

Theory
Here are a number of theoretical issues that need to be addressed:


 * There is a lack of consensus and research on the conceptual foundations of this new paradigm i.e., on the relationship between image-objects (segments) and landscape objects (patches). For example,
 * (i) what is the basis to believe that segmentation-derived objects are fine representations of landscape structural-functional units?
 * (ii) How do you know when your segmentation is good?


 * In addition, there exists a poor understanding of scale and hierarchical relations among objects derived at different resolutions.
 * (iii) Do segments at coarse resolutions really ‘emerge’ or ‘evolve’ from the ones at finer resolutions?
 * (iv) Should boundaries perfectly overlap (coincide) through scale? Operationally it’s very appealing, but what is the ecological basis for this?
 * (vi) how should we deal with transition zones in the landscape?

(Roeland de Kok, 28-11-06)
 * A famous test from Johanneum research in Graz extracts a set of training areas from an optical satellite image and apllies a shift to this training set. After the shift, the selected pixels are not any more a pure representation of the sample classes. However if this shift is minor, lets say one or two pixels, the result of using this shifted training set towards the origional result shows a small deviation of the final classification results. This test, which can be performed by any RS user, shows how neighbouring pixels are likely to have similar spectral values. The observation that neighbouring pixels in the image domain share similar spectral values is the starting point for using image segments as an input into the classification process


 * The assumption of correct representation of segments for patches can be found in 'parcel based classification'. Starting from a cadastral GIS it is easy to demonstrate that a set of segments correctly explains the conditions of agricultural crops on a parcel. The set of segments can be used to reconstruct a landscape object assuming here a patch equals an agricultural parcel. Crucial in this application is that the information on this object allows to express something on the quality of the content not on the correct spatial layout of the parcel. Even an incomplete spatial object reconstruction (parchel or patch) allows a sucessfull explanation (the field contains drying patato plants). For other objects, like road detection an incomplete result destroys the attribute connectivity, an essential feature of the class of road object.

In optical remote sensing, the image represents a photon count over the landscape. However the measurements are only valid for areas containing a population of pixels. The separation of the photon count from image object reconstruction is essential. If real world objects like agricultural parcels exist in a GIS, derived from an updated cadastre, the photon count per parcel is a representative measurement on energy exchange. Reconstructing the parcel from the pixel measurements will always fail at the borders. Only a reasonable approach of the origional parcel layout is possible. It is still sufficient to use the reconstruction to decide which pixels are representative and valid photon counts for the parcel as a whole. Addendum: follow and contribute the discussion on this!.


 * A segmentation is rather good when objects are smooth. This implies that segmenting a large waterbody in optical remote sensing always give good results and segmentation parameters can do no harm in constructing the object class of water. Starting with an existing agricultural GIS, the user can verify per parcel that the pixels behave as belonging to a normal distribution and can be considered rather smooth. As long as a pixel does not exceed a given variance of the neighbourhood, the neighbouring pixels indeed belong to the same distribution. The construction of a segment with low variance therefore is a representative measurement of photons on that particular part of the earth surface. A segment with a too large variance is not a valid measurement of that surface. This does not mean that an object can not be constructed from segments with high variance. Segments with pixels belonging to the same normal distribution approach the real world objects with homogeneous appearance in remote sensing imagery. As long as neighbouring pixels belong to the same gaussian distribution they should be merged to arrive from segments towards patches. An ideal patch is a set of neighbouring pixels within this patch having with neighbourhood patches that have different gaussian distributions or non-gaussian distributions. The relation of a segment towards an agricultural parcel can be a few to one or in optimal case a one to almost one relationship. Remark that this optimal case can hardly be reached without multitemporal imagery.In this case, objects with low variance are good candidates to express the condition of the objects and more easy to handle when constructing the object of interest (the patch) from the segments. Addendum: follow and contribute the discussion on this!.

Reconstructing an object from a digital image and explaining the condition of this object are separated procedures. Object with a large variance like a road can be reconstructed using OBIA. However the large variance of the pixel-population of this road does not allow to derive the smoothness af the asfalt layer for fast driving abilities as a qualification of this road. Objects which are spectral homogeneous can be constructed as well as qualified. A field of grass can be evaluated on the percentage of area whith lower vegetation activity.
 * Spectral and Spatial

Methods
This section gives an overview of known methods of segmentation and classification that are been used within OBIA

Software tools

 * This subsection gives an overview of sofware suites or modules, either commercially available or available as demo, that implement some OBIA method(s). Add a paragraph of lenght proportional to the interest of the tool to OBIA practioners, provide external hyperlinks to further information, and please use no marketing speech, as this page is not an infomercial. Insert your paragraph in alphabetical order

Definiens Developer,which resulted from the continuous advancement of eCognition is an object based image analysis software which features various segmentation algorithms that enble users to generate a networked object hierarchy. These objects can then be classified either rulebased or sample based. Definiens Developer is part of the Definiens Enterprise Image Intelligence Suite which consists of client and server software products which meet the needs of any image-based business process. The client products are role-based and support the needs of the different users. The server products provide a batch processing environment that enables the analysis of tens, hundreds or millions of images.

Free trial version download Definiens User Forum

The ENVI Feature Extraction Module is available as a plugin to the ENVI remote sensing software package available from ITT Visual Information Solutions (formerly Research Systems, Inc.). This add-on module uses object-based image analysis technology to extract features from imagery and automatically vectorize into Shapefile format. The ENVI Feature Extraction Module uses the combined process of segmenting an image into regions of pixels, computing attributes for each region to create objects, and classifying the objects (with rule-based or supervised classification) based on attributes in order to extract features. To request a free software evaluation simply contact your ITT-VIS Representative.

Applications
This section is intended to give an overview of problems where OBIA has been successfully applied, with emphasis on improvements achieved compared to other approaches. For particular examples or success stories, insert the references in the 'Further reading' section


 * Landcover mapping and monitoring


 * Forestry


 * Urban studies


 * Disaster prevention, assessment and relief


 * [ propose another field of application... ]

Key Terms
No discipline can be considered such without a consistent ontology. An ontology is an specification of a conceptualisation of a knowledge domain, that is, an ontology is a (typically hierarchical) taxonomy that describes the objects relevant to a given field and their mutual relations in a formal way. The aim of this section is to help develop an OBIA ontology with contribution from peers. We believe that wikis may allow to reach consensus much faster than through formal publications. Please add your proposals of key terms (include the ';' wiki markup before the term)and/or new definitions (include the ':' wiki markup before it). In order to avoid 'authority bias', please do not include authorship of definitions. Insert new terms so that the list remains alphabetically ordered.


 * Feature
 * A term to be avoided within this discipline, as it is ambiguous (it can refer both to a ground object and to a dimension of the hyperspace where the classification operates, i.e an attribute used to discriminate between different classes of objects.


 * Image-objects
 * Delimited regions of the image that are internally coherent and different from their surroundings
 * Delimited regions of the image that are internally coherent
 * OBIA
 * A sub-discipline of GIScience devoted to partitioning remote sensing imagery into meaningful image-objects, and assessing their characteristics through spatial, spectral and temporal scale.
 * A sub-discipline of GIScience that studies the territory using as basic units computer-delineated regions derived from remote sensing imagery whose meaning is elucidated by assessing their characteristics and mutual relations at multiple scales.
 * A sub-discipline of GIScience that studies the territory using as basic units computer-delineated regions derived from remote sensing imagery whose meaning is elucidated by assessing their characteristics and mutual relations at multiple scales.


 * Scale
 * Shows a very specific selection of spatial information in a chosen ratio between real world objects with spatial features and their presented counterparts and their spatial features. The selection compromises on the importance of the shown information as well as on it's area features in the real world. The spatial neighbourhood or relationships of all objects should be preserved. For example, the road west of the river might be shown using a line thickness that would not match the proper width of the road, however it should always remain west of the river and showing a correct lenght using the ratio ;;mapping scale:real world meters;;.

GIS is sometimes referred to as scale-less. In the sense that the compromise and selection of the information can be revered instantly and is not 'fixed' as in a paper print.
 * Segmentation
 * Spatial Resolution of a Classification
 * The minimum size of the circle, expressed by its diameter, over which the surroundings of a particular point on the earth have to be observed in order to determine the class (e.g. landcover type) from a list (legend) that should be assigned to that point.
 * The minimum size of the circle, expressed by its diameter, over which the surroundings of a particular point on the earth have to be observed in order to determine the class (e.g. landcover type) from a list (legend) that should be assigned to that point.


 * Structural signature
 * Recurrent spatial patterns observable at certain scales that can be readily associated to specific thematic classes.

Key Issues
Here are a number of strategic issues that need to be addressed:

OBIA name and definition
Following the discussion on whether the name (OBIA) we have chosen is misleading given the definition as a subdiscipline of GIScience, there seems to be two courses of action:
 * Take the concept of OBIA as an aspect of all the fields associated with image analysis and make a sub-section for OBIA in RS/GIS (so this wiki page would be renamed OBIA for RS and GIS)
 * Change the name so that it explicitly accounts for the link with RS/GIS. In this case we need a new name.Suggestions so far are:


 * GEOBIA (Geo-Object Based Image Analysis: pronounced GEE-OH-BE-UH) - This better distinguishes us from non-geographic areas of research i.e., biomedicine.
 * GOBIA (Geo-Object Based Image Analysis; NB. Gobia is a city in Ivory Coast)
 * OARS (Object-based Analysis of Remote Sensing images; NB. Oars are what you use to propel a small boat)
 * GIOA (Geographic Image-Object Analysis)
 * OBARSI (Object-Based Analysis of Remotely Sensed Imagery; NB. Obarsi = 'itself' in Italian, 'origin'in Romanian)

More suggestions or thoughts on this?

SWOT
In order to provide a better understanding of the current state of OBIA, and potential strategies to achieve the stated objective, a SWOT can be undertaken. SWOT Analysis is (one of many possible strategic planning tools) used to evaluate the Strengths, Weakness, Opportunities and Threats involved in a project, or any other situation requiring a decision. Consequently, our objective is to apply this method of planning early in the life cycle of OBIA with the intent that concepts described here can be used to strengthen and guide this emerging discipline.

In practice, once an objective has been established, a multidisciplinary team representing a broad range of perspectives should carry out SWOT analysis; which is typically presented in the form of a matrix

Strengths

 * Partitioning an image into objects is akin to the way humans conceptually organize the landscape to comprehend it.
 * Using image-objects as basic units reduces computational classifier load by orders of magnitude, and at the same time enables the user to take advantage of more complex techniques (e.g. non-parametric).
 * Image-objects exhibit useful features (e.g. shape, texture, context relations with other objects) that single pixels lack.
 * Image-objects are less sensitive to MAUP than units that do not keep a correspondence with the structure of the phenomenon under study.
 * Image-objects can be more readily integrated in vector GIS than pixel-wise classified raster maps.
 * Several OBIA methods/commercial software packages build upon the powerful object-oriented (OO) paradigm.

Weakenesses

 * Under the guise of 'flexibility' current commercial object-based software provides overly complicated options.
 * There are numerous challenges involved in processing very large datasets. Even if OBIA is more efficient than pixel-based approaches, segmenting a multispectral image of several tens of mega-pixels is a formidable task (efficient tiling/multiprocessing solutions are necessary).
 * Segmentation is an ill-posed problem, in the sense it has no unique solution, e.g., (i) changing the bit depth of your heterogeneity measure can lead to different segmentations. (ii) Remember, even human photo-interpreters will not delineate exactly the same things.
 * There is a lack of consensus and research on the conceptual foundations of this new paradigm, i.e., on the relationship between image-objects (segments) and landscape objects (patches). For example, (i) what is the basis to believe that segmentation-derived objects are fine representations of landscape structural-functional units? (ii) How do you know when your segmentation is good? (iii) Is there a formally stated and accepted conceptual foundation?
 * There exists a poor understanding of scale and hierarchical relations among objects derived at different resolutions. Do segments at coarse resolutions really ‘emerge’ or ‘evolve’ from the ones at finer resolutions? Should boundaries perfectly overlap (coincide) through scale? Operationally it’s very appealing, but what is the ecological basis for this?

Opportunities

 * Object-Oriented concepts and methods have been successfully applied to many different problems, not only computer languages, and they can be easily adapted to OBIA. This integration not only includes OO programming, but all the corpus of methods and techniques customarily used in biomedical imaging and computer vision that remain unknown to most of the remote sensing community.
 * While further research is needed, interesting integrative object-based proposals already exist that offers ontological foundations.
 * There are new IT tools (e.g. wikis) that may accelerate consensus and cohesion of OBIA.
 * There is a steadily growing community of RS/GIS practitioners that currently use image segmentation for different GI applications. Thus, as OBIA matures, new commercial/research opportunities will exist to tailor object-based solutions for specific fields, disciplines and user needs i.e., forestry, habitat mapping, urban mapping, mineral exploration, transportation, security, etc.
 * Symmetric multiprocessing, parallel processing and grid computing are recent technologies that OBIA methods may build upon to tackle problems related to the analysis of large datasets.

Threats

 * OBIA is far from been an operationally established paradigm, yet many users of commercial OBIA software do not recognize this fundamental fact. OBIA is not one specific research or commercial software. Much remains to be solved.
 * Trying to make distinct OBIA from other OO concepts and methods (e.g. by using 'based' instead of 'oriented') may contribute to insulation (of users in an esoteric world of 'objects') and isolation (of the concept) rather than to consolidation.
 * The visual appeal of image-objects, their easy GIS-integration and the enhanced classification possibilities have attracted the attention of major RS image processing vendors, who are increasingly incorporating new segmentation tools into their packages. This provides a wider choice for practitioners, but promotes confusion (among different packages, options, syntax, etc) and makes it more difficult to reach a consensus on what OBIA is all about. Will a lack of protocols, formats, and standards lead to a segmentation of the field rather than a consolidation?

OBIA research groups (in alphabetic order)

 * Foothills Facility for Remote Sensing and GIScience - University of Calgary, Alberta, Canada - Contact: Dr Geoffrey J. Hay
 * Z_GIS – Centre of Geoinformatics at Salzburg University

Conferences

 * 1st International Conference on Object-based Image Analysis (, Salzburg, 4-5 July 2006), Flyer (PDF):


 * Remote Sensing and Photogrammetry Society - Annual Conference 2007 (, Newcastle upon Tyne, United Kingdom, 11 - 14 September). If sufficient OBIA-related abstracts are accepted there may be an opportunity for a special session within the conference. Please submit a 500 word abstract using the on-line submission form at the conference website by the deadline of 17:00 (GMT) 26th March 2007. Also, please contact Geoff Smith (gesm@ceh.ac.uk) if you are interested.