|
Products
Data Sets
Publications
Protocols
Images
Contribution Series
CAP LTER Symposia
|
|
Land cover classification using Landsat Thematic Mapper (TM) data - year 1991
Identifier:503_1
Publication date:2008
Author(s):
Matthias Moeller
Abstract:
Landsat TM 1991 CAP LTER subset
Keywords:
Phoenix, Arizona, Sonoran Desert, caplter, Central Arizona Phoenix Longterm Ecological
Research, urban, metropolitan area, Landsat, land use, land cover, change detection, National Land Cover Dataset (NLCD), Land-Use and Land-Cover Change, Land Use Changes, Project id 20, caplter created, gis, remote sensing
Temporal Coverage:
1991
Geographic Coverage:
Geographic Description:Central Arizona Phoenix
Bounding Coordinates:
Longitude:-112.783228 to -111.572093
Latitude:33.844150 to 33.194886
Contact:
Data Manager, Arizona State University,
Global Institute of Sustainability,POB 873211,TEMPE
caplter.data@asu.edu
Methods used in producing this dataset:
The land-use/land-cover classification was produced using the
object-oriented approach implemented in the Ecognition software. This
approach consists of three steps. In the first step, the image is
divided into segments basing on gray values of the image bands, chosen
for the segmentation. Segments are clusters with a maximum spectral
homogeneity (M. Baatz, and A. Schaepe, “Multiresolution Segmentation: an
optimization approach for high quality multi-scale image segmentation,”
in: Angewandte Geographische Informationsverarbeitung, Vol. XII, J.
Strobl, T. Blaschke and G. Griesebnaer Eds., Karlsruhe, Wichmann, pp.
12–23, 2000). The user is able to define parameters such as shape size,
shape form and spectral values. Based on these parameters the segments
are calculated. The initial starting points are set randomly by the
software and the segments grow until they reach their bordering
neighbors. An image segmentation can be performed on several levels. It
ranges from a large number of small segments on a low level to a small
number of segments with a large size on an upper level. All levels are
linked to each other in a parent - child relationship and the segments
on each level are also connected to their specific neighbors. In the
second step the classes are named and at least these classes have to be
defined and outlined using fuzzy membership functions, which is the most
important and complex step during the classification [2]. The classes
can be described using a large number of class different parameters.
Starting with the definition of simple spectral properties (e.g. gray
values or brightness values of the bands against the mean of all bands)
the classes can be separated against each other. Also neighborhood
relations can be defined (e.g. relative border of class x is xx% to
class y). Inherited relations from upper to lower segmentation levels
can be used for the description of classes. That enables the performance
of a rough classification on an upper level with large image segments.
With finer segmentation this upper class can be reclassified into a
number of more detailed classes. For example: the class ‘urban’ was
classified manually on a higher level using the click and classify
algorithm. Click and classify is an easy to use method that enables to
achieve rough classification results in a relatively short time. But
these segments did not correctly match the exact border of the urban
area to another LULC class. These overlapping regions may belong either
to areas of farmland or to undeveloped desert. On a lower level with
smaller segments the classification could be refined. For example, the
‘urban’ upper class was split into three more detailed classes: urban
developed sparse, urban developed dense and
commercial/industrial/transportation. However, there are overlap areas
remaining on the finer level which have been classified as ‘urban’ (n
the higher level) but belong definitely to one category ‘farmland’,
either ‘vegetated farmland’ or ‘fallow farmland’. The spectral
properties of these regions were very similar or identical to the class
‘urban vegetation’. These classes could be differentiated using the
neighborhood relations. A segment classified as some kind of vegetation
must have a relative border of at least 50% to other urban features,
leading to classification as ‘urban vegetation’, otherwise it was
assigned to ‘vegetated farmland’. The neighborhood relations were
applied to all classes on the finer level. Checking these neighborhood
relations on the same segmentation level is an essential tool and it
increased the classification results compared to the statistical
approach. The accuracy of classification was checked with randomly
distributed control points.
Entities:
Raster: cap_lter_tm_91_class
download
Attributes:
Attribute:OID
Description:Internal feature number.
Attribute:Value
Description:Thematic value of the cell
Domain:
Enumeration:
0: Unclassified
1: 11 open water
2: 21 low intensity residential
3: 22 low intensity residential
4: 23
commerical/industrial/transportation
24: 84 fallow farmland
29: 81 - 83 cultivated farmland
vegetation
32: 85 urban/residential vegetation
Attribute:Id
Description:ID of class
Attribute:Red
Description:Red value of legend color
Measurement Unit:dimensionless
Attribute:Green
Description:Green value of legend color
Measurement Unit:dimensionless
Attribute:Blue
Description:blue value of legend color
Measurement Unit:dimensionless
Attribute:Class_name
Description:name of landuse class
Domain:
Enumeration:
0: Unclassified
1: 11 open water
2: 21 low intensity residential
3: 22 low intensity residential
4: 23
commerical/industrial/transportation
24: 84 fallow farmland
29: 81 - 83 cultivated farmland
vegetation
32: 85 urban/residential vegetation
Attribute:Opacity
Description:Opacity of legend color
Measurement Unit:dimensionless
Attribute:Count
Description:Number of pixels of each land use - land cover type used to
build a histogram
Measurement Unit:number
|
