short when compared wi th those of human inte rpreters (Ha y et al., 2003). This is in
part because these techniqu es (tra ditional ly) do not take into considerat ion or
incorporat e the concept of object within thei r analys is (Casti lla and Hay, 2008).
Yet this is fundam ental to human cognition (Biederman, 1987).
Rec ently, this limitati on has been more widely recogni zed, resultin g in new objectbased techno logi es being develo ped an d appli ed to lands cape analysis (Hay et al.,
2001, 2005; Hay an d Marce au, 2004; Blaschk e, 2010; Chen et al., 2012; Syed et al.,
2012) suppor ting the de velopment of a growing GEOB IA comm unity.
5 In 2000,
eCognition
6 becam e the first commerci ally avail able soft ware for mul tiscale objectbased segm entation (Baatz and Schäpe, 2000). Her e, segm entation follow s a proprietary bott om-up regio n merging techni que starting wi th one-pi xel objects, which are
iteratively merged into larger objects based on a user-de fin ed (spect ral) scale
parameter .
7 A key limit ation of this met hod is that there is no intu itive relat ionship
between the scale param eter in eCognit ion (wh ich is unitless ) and spatial measures
(i.e., area) speci fic to the o bjects o f interest composing a scene. This limitation has
recently been overcome by Casti lla (2003) an d Hay et al. (2005) wi th the develo pment
of size-const raine d regio n mer ging (SC RM), wher e scene delineati on is based on
either manua lly or automati cally [deriv ed from object-bas ed ima ge statistics (Hay
et al., 2005)] de fining the min imum, maximum , and average size of all scene objects.
More recent comm ercial softwar e such as ENV I FX
8 also include d unit less scale and
aggrega tion param eters but with the advant age that as the user changes these
parameter s (alon g a Graphical User Interface (GUI) sliding thresh old bar), the scene
is segm ented in real time allowi ng for the scale/aggr egati on levels to corres pond to the
boundar ies of the scene object s of inte rest, also calle d ima ge object s).
8.1.4 Image Objects
While remo te sensing images are capabl e of provi ding high- resoluti on spatial,
spectral, radio metric, and tempo ral data relev ant for lands cape analys is, these data
are not compo sed of discrete lands cape objects such as trees , fores ts, rivers, and citie s;
instead their fundame ntal primitive( s) are (typi cally millions of) pixels. Fortun ately,
image objects can b e de fi ned wi thin a remote sensing image (Casti lla and Hay, 2008).
In sim ple terms, image object s are groups of pixels with meanin g in the real world.
More speci fically, image objects are indi vidually resolvable enti ties located within a
digital image that are perceptually generated from high-resolution pixel groups (Hay
et al., 2001), which are sufficiently internally “homogeneous” so as to be distinguishable from their surroundings. High resolution (H-res) corresponds to the
situation where a single real-world object (i.e., a tree crown) is visually modeled
5 International conferences have taken place biannually since OBIA, 2006 (Salzburg, Austria); GEOBIA
2008 (Calgary Alberta Canada), GEOBIA 2010 (Ghent, Belgium); and GEOBIA 2012 (Rio de Janerio,
Brazil).
6 http://www.ecognition.com/products, last accessed May 24, 2012.
7 It also includes a user-defined shape heterogeneity parameter.
8 ENVI FX Feature Extraction ( http://www.exelisvis.com/language/en-us/productsservices/envi/enviex
.aspx), last accessed May 24, 2012.
INTRODUCTION
147
part because these techniqu es (tra ditional ly) do not take into considerat ion or
incorporat e the concept of object within thei r analys is (Casti lla and Hay, 2008).
Yet this is fundam ental to human cognition (Biederman, 1987).
Rec ently, this limitati on has been more widely recogni zed, resultin g in new objectbased techno logi es being develo ped an d appli ed to lands cape analysis (Hay et al.,
2001, 2005; Hay an d Marce au, 2004; Blaschk e, 2010; Chen et al., 2012; Syed et al.,
2012) suppor ting the de velopment of a growing GEOB IA comm unity.
5 In 2000,
eCognition
6 becam e the first commerci ally avail able soft ware for mul tiscale objectbased segm entation (Baatz and Schäpe, 2000). Her e, segm entation follow s a proprietary bott om-up regio n merging techni que starting wi th one-pi xel objects, which are
iteratively merged into larger objects based on a user-de fin ed (spect ral) scale
parameter .
7 A key limit ation of this met hod is that there is no intu itive relat ionship
between the scale param eter in eCognit ion (wh ich is unitless ) and spatial measures
(i.e., area) speci fic to the o bjects o f interest composing a scene. This limitation has
recently been overcome by Casti lla (2003) an d Hay et al. (2005) wi th the develo pment
of size-const raine d regio n mer ging (SC RM), wher e scene delineati on is based on
either manua lly or automati cally [deriv ed from object-bas ed ima ge statistics (Hay
et al., 2005)] de fining the min imum, maximum , and average size of all scene objects.
More recent comm ercial softwar e such as ENV I FX
8 also include d unit less scale and
aggrega tion param eters but with the advant age that as the user changes these
parameter s (alon g a Graphical User Interface (GUI) sliding thresh old bar), the scene
is segm ented in real time allowi ng for the scale/aggr egati on levels to corres pond to the
boundar ies of the scene object s of inte rest, also calle d ima ge object s).
8.1.4 Image Objects
While remo te sensing images are capabl e of provi ding high- resoluti on spatial,
spectral, radio metric, and tempo ral data relev ant for lands cape analys is, these data
are not compo sed of discrete lands cape objects such as trees , fores ts, rivers, and citie s;
instead their fundame ntal primitive( s) are (typi cally millions of) pixels. Fortun ately,
image objects can b e de fi ned wi thin a remote sensing image (Casti lla and Hay, 2008).
In sim ple terms, image object s are groups of pixels with meanin g in the real world.
More speci fically, image objects are indi vidually resolvable enti ties located within a
digital image that are perceptually generated from high-resolution pixel groups (Hay
et al., 2001), which are sufficiently internally “homogeneous” so as to be distinguishable from their surroundings. High resolution (H-res) corresponds to the
situation where a single real-world object (i.e., a tree crown) is visually modeled
5 International conferences have taken place biannually since OBIA, 2006 (Salzburg, Austria); GEOBIA
2008 (Calgary Alberta Canada), GEOBIA 2010 (Ghent, Belgium); and GEOBIA 2012 (Rio de Janerio,
Brazil).
6 http://www.ecognition.com/products, last accessed May 24, 2012.
7 It also includes a user-defined shape heterogeneity parameter.
8 ENVI FX Feature Extraction ( http://www.exelisvis.com/language/en-us/productsservices/envi/enviex
.aspx), last accessed May 24, 2012.
INTRODUCTION
147