I work in the holmes project in clark university and we produce land-cover maps of NE Massachusetts using very high resolution imagery. We use object based image analysis techniques to make these land-cover maps. There are many commercial image segmentation softwares available and each use different algorithms to segment the imagery. OBIA consists of two stages image segmentation and then classification of the segments. Segmentation scale is the most important parameter in the segmentation stage. UVM who gave us the training to use OBIA stated that a scale of 50 works best for them. But their imagery was different resolution from what we use for land-cover mapping. Even the literature on OBIA does not provide any information on the relation between resolution the image and the scale parameter.In our mapping exercise we had a choice of different segmentation algorithms and a choice of scale for each algorithm. The state of art in the choice of scale parameter now is to do eyeballing or visual interpretation. In this presentation I’ll show a method to quantify the selection of appropriate scale of segmentation.
The scale factor which is a unit-less parameter related to resolution of the image is an important characteristic of the segmentation procedure.The heterogeneity of a segment increases with the increase in the scale factor, thus larger scale factors tend to produce larger segments.In this image we want to map the land-cover of the baseball diamond. Fine green areas, impervious surfaces and bare soil are present here.In this segmented image I have used the scale of 30 and as we can see there are far more segments in the image than are necessary. This kind od segmentation is called oversegmentation.In this image the scale is 250 and the number of images is fewer than required and as we can see the segments are missing some patches of bare soil here and here. This kind of segmentation is called under segmentation.This image shows a segmentation of 70 and this is not that oversegmented as compared to scale 30 but still separates the 3 land-cover classes here.Oversegmentation is acceptable as we can always merge the segments after classification but too many segments can slow the process.Undersegmentation is undesirable as is not possible to divide the segments into finer segments during the classification process.So Over segmentation is ok but under segmentation is undesirable in the segmentation process.The scale factor is a relative term used in different softwares a scale of 10 in one software is not the same in some other software.
Watershed Algorithm: The image is regarded as a topographic surface with the gray values converted into gradients. The image is then is divided into a set of high-gradient watershed lines and low-gradient region interiors that act such as catchment basins. These catchment basins correspond to relatively homogeneous segments in the image.Region Growing : This algorithm aggregates pixels starting with seed points and grows into segments through a pair-wise clustering process until a certain threshold is reached which is normally a homogeneity criterion based on color, smoothness and compactness.
No information available on relation between pixel size and scale parameter. When compared to the reference segments, a better segmentation algorithm is the one which produces equal number of segments in the same locations. Selection of optimal algorithm-associated scale parameter is therefore an important step towards high quality segmentation for a given feature type
The segments in map b that have their centroid and at least 50% of surface area in the corresponding segment in map a are selected.
A circular segment would have the value of 1, and as the patch became more convoluted in shape, its shape index would increase in value.
A method to select imagesegmentation scaleRahul Rakshitrrakshit@clarku.eduGraduate School of GeographyClark University hero.clarku.edu/holmes Graduate School of Geography, Clark University 1
Scale Factor: heterogeneity of a segment increases with the increase of scale factor 30 Over-segmented 70 250 Under-segmented Graduate School of Geography, Clark University 2
Segmentation Algorithms Chessboard Quadtree Region Growing Watershed Graduate School of Geography, Clark University 3
Objective: Selection of suitable segmentation algorithm and scale WS10 WS20 WS30 WS40 RG40 RG60 RG80 RG100 RG120 Graduate School of Geography, Clark University 4
Data 3 Bands, 15 cm, 1331x929, Aerial Photo, 2008, MassGIS Graduate School of Geography, Clark University 5
Reference Dataset Hand Digitized Segments, n= 546 Graduate School of Geography, Clark University 6
Segmentation scale comparison: Number of Segments Under Segmented RG120 RG= Region Growing WS= Watershed RG100 RG80 RG60 RG40 WS40 WS30 WS20 Over Segmented WS10 Reference 0 1000 2000 3000 4000 5000 6000 7000 Graduate School of Geography, Clark University 7
Comparison Parameters 1. Circularity 2. Shape Index 3. Over Segmentation 4. Under Segmentation 5. Closeness 6. Hammoude Metric 7. Boundary Matching Graduate School of Geography, Clark University 8
Comparing shape of segments: CircularityA 2D geometric tolerance that controls how much a feature can deviate from a perfect circle RG120 RG100 RG80 RG60 RG= Region Growing WS= Watershed RG40 WS40 WS30 WS20 WS10 Reference 0 5 10 Graduate School of Geography, Clark University 9
Comparing shape of segments: Geometric Feature Shape Index perimeter shapeindex (Neubert et al. 2008) 4 area RG120 RG100 RG80 RG60 RG40 WS40 WS30 WS20 RG= Region Growing WS10 WS= WatershedReference 0 1 2 3 4 Graduate School of Geography, Clark University 10
Over and Under Segmentation oversegmentation2 undersegmentation2 Closeness (Clinton et al. 2010) 2 RG120 Over segmentation RG100 Under segmentation RG80 Closeness RG60 RG40 WS40 WS30 WS20 WS10 0 0.5 1 1.5 2Perfect Match Dissimilar Graduate School of Geography, Clark University 11
Hammoude Metric area(a b) area(a b) H (Marcel 2009) area(a b)RG120RG100 RG80 RG60 RG40 WS40 WS30 WS20 WS10 0.75 0.8 0.85 0.9 0.95 1 Dissimilar Graduate School of Geography, Clark University 12
Distance Metric: Boundary Matching D (r ) r= Boundary pixel of a segment in the reference map D(r)= Euclidean distance between r and any boundary pixel in the segmented map N N= Number of boundary pixels in the reference segment (Delves et. al 1992)RG120RG100 RG80 RG60 RG40 WS40 WS30 WS20 WS10 0 0.5 1 1.5 Perfect Match Dissimilar Graduate School of Geography, Clark University 13
Selection by weighted combination Comparison Parameter Weights Difference in Circularity 2 Difference in Shape Index 2 Under Segmentation 1 Over Segmentation 5 Closeness 25 Hammoude Metric 25 Boundary Matching 40 •Average of the parameter is used •Complement of the parameter (1-parameter) is used Graduate School of Geography, Clark University 14
Results WS20 Reference WS30 WS10 WS40 RG40 RG60 RG80 RG100 RG120 Graduate School of Geography, Clark University 15
Compare Segments tool Graduate School of Geography, Clark University 16
AcknowledgementsAdvisors: Prof. Robert Gilmore Pontius, Jr. Prof. Colin Polskyholmes Team: Albert Decatur Nick Giner Dan RunfolaData: MassGISSoftware Support: James Toledano, IDRISI, Clark Labs Shitij Mehta, ESRIMore Information: email@example.com http://hero.clarku.edu/holmesThis material is based upon work supported by the National Science Foundation (NSF) under grant Nos. BCS-0709685 (Coupled Natural-Human Systems), OCE-0423565 (Long-Term Ecological Research), SES-0849985 (REU Site), and BCS-0948984 (ULTRA-ex), and by theClark University OConnor 78 Endowment. Any opinions, findings and conclusions or recommendations expressed in this material arethose of the author(s) and do not necessarily reflect the views of the funders. Graduate School of Geography, Clark University 17