CHAPTER 2 LITERATURE REVIEW
2.7 Remote Sensing and Its Roles in Landslide Hazard Assessment
2.7.7 Image Classification for derivation of Land Use and Land Cover
A good explanation on how this image classifier works was given by Gibson and Power [94]. Once the number of classes defined, the computer software specifies arbitrary DN means to each class and have pixels moved to the nearest mean of classes. The software then recalculates new class means. These values are used to re- evaluate each pixel. The software then relocates pixels to the closest new class means.
This procedure can be repeated several times until meet the number of iterations set by the user or achieve the threshold value set by the user. The author said that at a threshold of 0.98, the adjacent iterations left less than 2% of pixels moving. Included as unsupervised classification techniques are K-means, ISODATA (Iterative Self- Organizing Data Analysis Technique), and histogram based clustering as described by Gao [165].
2.7.7.2 Supervised Classification
Unlike the previous method, supervised classification requires the user to provide spectral signatures of known land cover lands categories such as forest, urban, barren land, bushes, etc., prior to do image classification. In doing so, such spectral signatures are sampled from the image by delineating several areas of a homogenous type, e.g. forest, then computing statistical parameters, e.g. mean, standard deviation, etc., of their spectral signatures. These areas are called as training areas. The other categories undergo the same procedures so that the statistical parameters of spectral signature of all desired land use land cover categories is made available. The image processing software then compares the DN of each pixel with the statistical parameters and do image classification using statistical techniques so called image classifiers. There three classifiers as described by Liao [47], Lillesand, et al. [166], Gao [165], and Gibson and Power [94] namely, from low to high accuracy, parallelepiped, minimum distance-to-mean or nearest neighborhood, and maximum likelihood classifiers.
Parallelepiped classifier constructs ‗boxes‘ using statistical parameters of training areas with mean as the center of boxes and maximum/minimum and standard deviation for determining the width and length of the boxes (Fig. 2.15a). Any pixel falls within a particular box, for example box of forest category, will be grouped to
Fig. 2.15 Illustrations of supervised classification classifiers Source: CCRS [178]
that category. This method offers fast classification but has poor accuracy. Using this method, a pixel can be belongs to two boxes or none of them. In Fig. 2.15b, pixel a
c)
a) b)
X
belongs to cluster/class 4 due its location within box 4. Meanwhile, pixel b could be unclassified or has no class since there are no boxes containing this pixel. Using Minimum distance-to-mean (Fig. 2.15b), pixel b can be grouped to class 3 according to the shortest distance between the distance from pixel b to all class mean pixel.
Maximum likelihood is considered as the most useful classifier admitted by Gao [165]. The so-called Gaussian maximum likelihood classifier evaluates any pixels using variance and covariance of spectral signature of land use land cover categories.
This method applies an assumption that the distribution of points forming each category, e.g. forest, water, crop land, etc., derived from training areas is Gaussian (normally distributed). Fig. 2.15c shows how point/pixel X belongs to forest type rather than to agriculture classes. The reason is that the probability of pixel X being grouped to forest class is higher than to agriculture class. The graph of probability density of forest looks more normally distributed than that of agriculture which appears to be a bit narrow.
2.7.7.3 LULC Classification Scheme
Anderson, et al. [179] designed a classification system for LULC for use with remote sensing data that satisfy the needs of the majority of users. This scheme does not differentiate between land cover and land use. The scheme starts with Level I category for example; urban or built up land, agricultural land, forest land, etc. Urban or built up land is broken down for level II into such as residential, commercial and service, industrial, etc. One thing to take note is that the various types of land use land cover in level II do not necessarily available. Hence, one may reduce or modify it to fit with what available in the study area. The scheme, shown in Table 2.7, contains 9 levels I while the total of level II is many. Level III and IV are left open-ended to users such as federal, regional, state, and local agencies so they can have flexibility in developing more detailed land use land cover according to their particular needs.
LULC classification for level I to IV requires different specification of satellite imageries. Anderson, et al. [179] did not specify detail of satellite mission in accordance to various level of classification. However, the authors only stated about
Table 2.7 Image classification scheme
Category
Number Level 1 Level II
1 Urban or built-up land 11 Residential
12 Commercial & services 13 Industrial
14 Transportation, communications & utilities 15 Industrial & commercial complexes 16 Mixed urban or built-up land 17 Other urban or built-up land 2 Agricultural land 21 Cropland & pasture
22 Orchards, groves, vineyards, nurseries &
ornamental horticultural areas 23 Confined feeding operations 24 Other agricultural land 3 Rangeland 31 Herbaceous rangeland
32 Shrub & brush rangeland 33 Mixed rangeland 4 Forest land 41 Deciduous forest land
42 Evergreen forest land 43 Mixed forest land
5 Water 51 Streams & canals
52 Lakes 53 Reservoirs 54 Bays & estuaries
6 Wetland 61 Forested wetland
62 Nonforested wetland 7 Barren land 71 Dry salt flats
72 Beaches
73 Sandy areas other than beaches 74 Bare exposed rock
75 Strip mines, quarries & gravel pits 76 Transitional areas
77 Mixed barren land
8 Tundra 81 Shrub & herbaceous tundra 82 Herbaceous tundra
83 Bare ground tundra 84 Wet tundra 85 Mixed tundra 9 Perennial snow or ice 91 Perennial snowfields
92 Glaciers
Source: Anderson, et al. [179]
the altitude of remote sensing platform. As classification goes deeper (e.g. level IV), low attitude platform with higher resolution is required. As illustration, global
MODIS (The Moderate Resolution Imaging Spectroradiometer) satellite can provide data for level I with resolution range 250 m to 1.1 km. Level II (resolution range of 80 m to 250 m) can be fulfilled by Landsat Thematic Mapper. Level III (30 m to 80 m) can be fulfilled by Landsat 7 ETM+. Level IV (3 m to 30 m) can employ SPOT and aerial photograph to provide such data, and level V, if required, can utilize IKONOS or QuickBird images.