Home Articles Multi-Threshold Technique for Segmentation of Satellite Imagery for Feature Extraction

Multi-Threshold Technique for Segmentation of Satellite Imagery for Feature Extraction

Multi-Threshold Technique for Segmentation of Satellite Imagery for Feature Extraction

Shattri Mansor, Teoh Chin Chuang, Abdul Rashid Mohamed Shariff, Noordin Ahmad
Spatial & Numerical Modeling Laboratory, Institute of Advanced Technology
Universiti Putra Malaysia, 43400, Serdang, Selangor, Malaysia

[email protected]


The purpose of segmentation is to separate an image into regions corresponding to objects; i.e. to divide the image into regions, which are identifying common properties. Image regions are expected to have homogeneous characteristics (e.g. grey level, or colour), which may indicate that they belong to the same object or facet of an object. Mathematically, segmentation can be defined by the following rules. An image I must be separate or segmented in several different regions Ri, i = 1, 2,…., n

The first property states that the entire image I must be covered by the regions Ri, i = 1, 2,…., n and the second property states that regions are disjoint sets. The logical predicate P(Ri) in the third property must be true(1) over each regions Ri to ensure distinguish one region from another. Finally, the fourth property states that the predicate P(R) must be false (0) in the union of two image regions to ensure every region is distinct from other regions.


How can we divide an image into homogeneous regions? One of the simplest methods and natural way to segment is through thresholding. Thresholding is one of the powerful methods for image segmentation; it is useful in discriminating objects from the background in many classes of scenes. For example, an image contains an object, which is made up of pixel, which has homogeneous grey level and a background with a different grey level. Such an image usually possesses a bimodal histogram is shown in figure 1. This image can be segmented into two different regions by simple thresholding, which classifies all pixels with grey level values greater than T as object pixels and pixels with grey level values smaller than T as background pixels. Most thresholding techniques indirectly utilize the shape information of an image histogram. The ideal case is a bimodal shape of histogram because of grey level at the valley can be direct selected as threshold value for segmentation, but such bimodality histogram is usually unavailable in the real application (Tseng et al., 1993). In general, an image grey level can be divided into several sub-ranges to perform thresholding but these ranges usually overlap one with another and make thresholding difficult. In this experiment we need to separate the test image into four sub-ranges.

Segmentation by thresholding started many years ago from simple beginnings, and in recent years has been refined into a set of mature procedures and the outstanding problem is how to devise an automatic procedure for determining the optimal threshold (Papamarkos et al., 1999). Sato et al, (1999) used threshold operation to extract mangrove forest with TM data of Landsat 5. This method involve two steps, firstly decided provisional range of threshold selection on the histogram for band 4 and 5 and secondly applied Otsu’s automatic threshold selection method to the range as an automatic searching threshold. Chang et al, (1997) proposed a new wavelet-based automatic multi-level thresholding technique for segmentation of thermal images and Tseng et al, (1993) used an automatic thresholding method based on aspects of the human visual system for segmentation.

Methodology and Results

Figure 2 shows the basic procedure for this study. First step, we selected classes have been through analysis to the land use map representing training site. Second step, an automatic threshold selection program is used to determine the optimal threshold values of each feature. Final step, thresholding based on that threshold values to perform segmentation.

Training Site Selection

Training site selection has been defined as process of searching a subset of the original set or areas representing each known land cover classes that appear fairly homogeneous characteristics on the image. For the selection of the training sites, two different ways have been used to collect training site data for this study. Firstly, by using polygons covered the representing pixels of the selected classes. Secondly, by doing on-screen seeding which is selecting pixels one by one from different areas. Four land cover classes of training site data had been identified; they are urban, clear land, forest and water. In this paper, purpose of training site selection is to determine the minimum (Min) and maximum (Max) grey level for each class (feature) from different band of the test image. Table 1 shows all the Min and Max grey level values that is extracted from the test image. Min(i, k) and Max(i, k) is minimum and maximum grey level values of features i in band k, i = 1, 2,…, n = 4, where, 1 is for urban feature, 2 is for clear land feature, 3 is for forest feature and 4 is for water feature. k is number of band, k = 1, 2,…, m = 6.

Automatic Threshold Selection Program

In this program, are divided into two stages. They are pair features overlapping analysis and optimal threshold values selection.

Multi-Threshold Technique for Segmentation of Satellite Imagery for Feature Extraction

Pair Features Overlapping Analysis

The pair features overlapping analysis algorithm can be described in figure 3. In this analysis, 12 logical expressions are used to do the overlapping analysis of the range of Min to Max grey value for a pair of features in one band. The algorithm shown that expression 1 to 8 can be used to extract threshold values and 9 to 12 are no threshold values can be extracted. This algorithm works by overlapping analysis to the one feature ( i, k ) with another feature ( j, k ) in one band k, where, ( i )