[GC,12Jul06] The OBIA wiki page has now been created. This discussion page is intended to receive comments, suggestions and controversial discussions that otherwise would populate the OBIA article page. Please put your comments, suggestions, clarifications and/or disagreements here, placing them in the relevant section (unless they are general), or create a new section with your topic if it is not related to the existing ones. Put your initials and date of your comment in brakets before the comment itself, and place the comment so each section/thread follows a chronological order. And remember, be bold!
About this page
- OBIA is only about RS/GIS??
- Glad to see the wiki in operation. But in the first paragraph where OBIA is introduced, its not fair to restrict OBIA to GIScience itself. The concept of OBIA has for obvious reasons several applications in any kind of image analysis. Just because we are using it extensively we can't claim the name. It would be a good idea 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. I'm sure such an attempt will bring in contributions from different fields to achieve the same goal i.e., image analysis.
- [GC]. I agree, we cannot claim the name, even more when OBIA is customarily used in e.g biomedical imaging (BMI). However we have an specific pursue (to emulate photointerpreters), in the same way BMI has its own. Hence we have different goals, and we study different fields, even if we use the same means (digital images). For me it's like 'digital classification' and 'pattern recognition'. The first is RS specific, but it uses methods from the 2nd, yet is a different discipline. Image analysis methods using as first step segmentation are well integrated in domains such as BMI or CompVis, and probably for this reason these group of methods don't need an specific name for that. However, in the RS/GIS world, this approach is quite new and is in need of wider recognition, and a name is part of it. We could coin a more specifi name, such as GISOBIA (geographic information specific OBIA), but I wonder if that is a sexy name, I guess no. So in short, counting tumoral cells and locating apples in a robot image are tasks that may use similar methods than the ones used to depict geoobjects from a RS image. Notwitstanding, these three task belong to very different fields of study. Now the question is, what name should we give to this pproach within the wider GIScience mother discipline?--Guillermo 16:01, 17 July 2006 (MDT)
- [Prashanth,18/07/06] What i would suggest is to keep this field as OBIA but add a sub-section for RS/GIS applications. All the theory that is developed and will be developed can be used by anyone be it in medical image processing, astronomy or microscopic image analysis. But by separating the application specific part of OBIA in RS/GIS, we will have our own identitiy. I propose GOBIA (Geo-Object Based Image Analysis). Lets keep the discussion open. Meanwhile i advice changes in the introduction paragraph as it sends wrong signals to newcomers.
- [AU, 21/07/06] I have attended the first day of OBIA in Salzburg. Being myself an image processing person I was quite surprised by the abbreviation "OBIA" since it suggests a general symposium on object based image analysis - however, as you understand it, OBIA is meant to be a subsection of GIS or remote sensing .......from my point of view, this should be reflected in the abbreviation also, e.g. OBRIA or similar, since the current abbreviation is clearly misleading.
- [GeoffHay, 24/07/06] I agree with Prashanth, GOBIA makes much more sense, and really emphasises the geo-object component and basis of this form of image analysis.
- [GeoffSmith, 28/07/06] I think we need to develop the idea that object-based image analysis does not necessarily begin with segmentation. There are lots of object data sets out there which could be used to drive OBIA. This then leads on to the name. When I first encountered this area of work it was being called Integrated GIS (IGIS) - a GIS based approach that integrated any type of data into the analysis (raster, vector, table etc.). I think GOBIA may be a good solution, although it sounds more like a place!
- [TimW, 23/10/06] Good to see this site up and running. While I agree with all that has been said, I would be a bit wary of going the 'geo' path with the name. While there are objects being created/used or whatever, they may or may not have a spatial origin. In other words, an object created from the segmentation of an image based on DNs of spectral bands has a spatial entity but I'm not sure I would call it a geo-object as it might not be representative of a geographic feature. It might be part of a feature or a mixture of features. I also think the term 'geo' could detract from the aim of the task which is image analysis. What about the term OBARSI (object based analysis of remotely sensed imagery)? Italian for 'itself'. Similar to OARS but a bit more exotic sounding.
- [S Lang, 4/11/06] First of all, I am very happy to see OBIA Wiki up and running! And nice that first in-depth discussions take place. A few words from my side regarding the labelling of what we do. I think there are two distinct strategies to follow – and as a forming scientific sub-community we have to decide which path to take. With OBIA, I think we can observe a reforming of scientists working from different areas in an interdisciplinary manner. Some people would call that “transdisciplinary”, because some emergent phenomena rise from this cooperation, namely new insights and new ways to understand reality.
- Coming back to the strategies: either we go broadly and consider OBIA a comprising methodological approach, which is not restricted by any application or scale domain. In other words, no matter if we look at geographical, medical or technical realms, we see reality as being structured of objects and in order to understand this reality, we need to handle these objects. Of course, much has been accomplished in these areas by other disciplines by a range of optimised solutions and algorithms for particular issues. And, from a spatial science perspective, we could provide valuable input for gaining more flexibility through spatial analysis techniques (in its broadest sense).
- The other strategy is to focus on "geo" as such. Personally, I feel kind of comfortable with it since demarcating a certain field of expertise – if you want, a territory. I think it is worth it. It doesn’t mean we leave behind the scientific discourse, we will of course interact thoroughly and communicate with neighbouring disciplines (i.e. adjacent scale domains ...). I like GEOBIA, it is a handy abbreviation considering the fact, that the “o” can both be part of "geo" or "object" (if one considers the actual Greek word for Earth, "gé", then “ge-o” is nothing else but an earth-object). A comment to Tim's remark: I see your point, you are perfectly right, and I would not dare to claim that everything being automatically derived is necessarily a geographical object. But on the other hand, delineating or modelling geo-objects is the aim of what we do! Why using remote sensing data or any other earth observation techniques, if not for providing relevant units? No doubt that this is challenging and partly controversial, but that’s why we are here. The only thing that puzzles me a bit is the fact, that we are not necessarily restricted to images as such; we use other continuous information for object delineation as well. On the other hand, applying a more general concept of “image” (Latin: “imago”), it could be a synonym for any pseudo-continuous representation of a continuous phenomenon.
- Talking about terminology, I am just about compiling an article where I would like to propose the word “geon” for a systems-theory driven, scale-dependent and hierarchical 'earth-object'. In some way in accordance to other word creations using the Greek "on" and attributing to Koestler’s holon concept (like econ for an ecological unit). “Geon” would be general, though. In fact, this term has already been used in cognitive psychology by Biederman. He describes geometrical ons, which we use for mentally compiling more complex structures in the process of perception. Of course, there are some parallels, so let’s see …
[GC,12Jul06] Should we delete this section and/or move it to Key issues?
- photon counts and image-object reconstruction
- [GC] This is an interesting point that would need further clarification from the author [Roeland]. At least to me it is not clear to what what s/he means by 'measurements' and 'reconstruction'. S/he also seems to assume that the real object (cadastral agricultural parcel), in order to be reconstructed (delineated accurately from the image?) have to be homogeneous (meaning that the DNs of the pixels within it fit well a normal distribution with low [how low?] coefficient of variation). But for example if the parcel is an olive tree plantation and is imaged at Quickbird resolution, the set of pixels belonging to the parcel will show a multimodal distribution and yet an algorithm taking advantage of the recurring spatial patterns present in that area, could correctly delineate the boundary of the parcel providing it is surrounded by a different crop. Was it what you meant when you streesed the need to separate phothon count (digital classification?) from object reconstruction (outlining patches of terrain that are semantically homogeneous and different from their surroundings?)? Then I agree, but I would like to understand better your point. Mine is that shape is far more important than color for meaning, i.e that in order to understand and image we rely more on spatial patterns than on spectral ones. --Guillermo 23:31, 13 July 2006 (MDT)
For certain remote sensing applications, shape is irrelavant. Software like eCognition even produces objectborders on lake surfaces. Something that does not happen with QuadTree segmentation techniques. For a water object surrounded by water objects, only the spectral properties matter. (RdeKok_dec06)
- Boundaries and transitions
- [JRadoux] In GIS, simple objects can be defined as point, lines or polygons. Image classification often yields polygons when converted into GIS. From an ecological point of view, lines (e.g. ecotones) or points (e.g. isolated tree) also have meaningfull properties which can be derived from OBIA.In my opinion, considering edges instead of areas is a good way to handle "natural fuzzyness" or "detection accuracy" in vector maps.
- Classification and Deliniation
Delineation and classification are somehow connected in the sense that there is a sequence of assigning class labels to pixels and using the labels to make a local decision on deliniation. Afterwards, deliniated pixel groups are selected to be assigned to the same class. Dr. Werner Schneider from BOKU, Vienna pointed this out in earlier apllication notes and presentations in the late 90th. What is important in the early development of OBIA is the notion on the strong focus of per parcel classification as a hot topic in the 90th which can be found for example by Dr. Lucas LF Janssen.
The per parcel classification starts with the notion that the majority of pixels within a parcel are similar and their histogram shows a Gaussian Distribution. In this case it is better to take into account the mean and the standard deviation of those pixels within one parcel. This is not the same as the mean and the standard deviation of a class. A parcel with mais as an object has it's own mean and standard deviation in the image domain. The class of all Mais pixels in the feature space domain however might deviate from that. The crucial decision here is that the standard deviation in feature space of the class Mais should not decide if a single pixel belongs to the parcel with mais but the pixel inside the parcel mais must be reviewed knowing in advance it belongs to this parcel of mais. That is why salt and pepper dissappears. The single pixel outside the variance of the class mais in feature space wrongly assigned to another class in pixel based classification methodes still is incorporated within the parcel of mais inside the image domain and is assigned a correct label 'Mais'in an object based analysis.
In accordance with the example used in this discussion, a parcel with olive trees in a Quickbird image, The spectral Mean value of this parcel as well as it's very large standard deviation does not represent the class of 'olive tree plantations' . It is the human eye that assumes a grouping of subobjects (trees) in the 'matrix' of the surrounding field below and arround the canopy. A spectral mean of 'Grass' as well as a spectral mean of 'olive tree' can be found. It is the spatial decision 'olive tree surrounded completly by grass which allows a classification of the plantation.
In the remarks of Dr. Martin Baatz (Definiens), the human eye views imagery as related patches, which according to him makes OBIA analysis closely related to human perception. However, a pure edge detection image in very high resolution (below 1 meter) shows already delineated buildings, roads etc. In this black and white image, a spectral mean value has no meaning at all.
Image edge classification is a smaller problem within OBIA. Image edge classification can not be handled with classical spectral classification methodes like a maximum likelyhood classifier. (Roeland de Kok Sept.-'06)
Shape versus Colour The spatial information in the satellite image is sufficient to approach very closely the parcel of mais. However, during various seasons, this parcel contains patato, sugarbeet, fallow_grass vegetation or bare soil. Only the colour information, it is the counting of photons in various spectral bands will provide the information about the actual vegetation on this single parcel. No spatial information might help us to define this. The Spatial information on this parcel only provide us with it's borders. OBIA provides both information. The pixels belonging to this parcel and the actual vegetation on this parcel
(Roeland, 14.37 29-09-2006)
- are segmentation-derived objects representations of landscape structural-functional units?
Objects of Interest can be linked to objects in the legend of a map. The classical topographic map shows many objects which can be accepted as goals for OBIA to be extracted from a remote sensing image. In classical remote sensing, the image classification delivers a class water In OBIA it is possible to classify Lake, Pond, Channel even Rapids and Waterfall. Those objects all belong to the class Water and can not be classified by traditional multispectral analysis.
If objects like Lake, Pond and Waterfall are part of landscape structural-functional units, than segmentation derived objects can be used to construct these objects of interest from the object primitives.
A classical landscape unit is 'the Island'. A single OBIA rule allows to classify Island; Neigbouring-Non-Water segments completely surrounded by Water. The segmentation derived object island indeed is a perfect example of a landscape-functional unit.
In the same map, the human perception considers objects like lake,island,castle,forest and shopping-mall all as equall objects. To classify a group of segments as island is very much more simple than classifying shopping-mal and this even less complicated than castle.
In anology, we perceive coins from copper, gold, silver, aluminium and platin as coins. However in Napoleon time it was almost impossible to speak about aluminium and platin coins. I assume image objects go the same way. Technological development in robotic vision will be able to produce image objects similiar to platin and aluminium coins.
(Roeland de Kok, 11.50 30-09-2006)
OBIA research groups (in alphabetic order)
Further readings (online)
Hi, maybe we should/could move this section to an extra page and add a lot of literature there? (There are more than three papers out there ;) ...) I'd like to have such a section as a kind of OBIA 'library' with all related literature in it. Then I can have a look if someone has published about a certain aspect that I am dealing with myselfe or try to find out more about ... It could be ordered by year and then by author or by topic (more difficult, since we would have to define them). What do you guys think? – Tillmann 03:33, 14 July 2006 (MDT)