4.2 Object-based image classification
4.2.1 Classification of main classes
The first aim of this work is to classify the main land-use classes shown in the class hierarchy (see Figure 3.9). The classification strategy also follows the class hierarchy, which begins with the land uses on a higher level. The result has a similar spatial scale like ATKIS data and will be used as a base map for further detection for the detailed landscape elements.
4.2.1.1 Settlement and traffic extraction
It is often the case that rural settlements show no distinct border with their surrounding area and traffic areas are covered by roadside trees in the Remote Sensing image. A sound solution for the artificial area extraction is to use the existing land-use maps which are
manually produced. In the first step of work the official land-use maps in both test sites are incorporated as thematic layers for image segmentation and classification. The chessboard segmentation algorithm is used to produce segmentation level1. In this algorithm the parameter “object size”, which determines the size of segments, should be set sufficiently high to constrain the border of segments identical to the settlement class and traffic class in land-use maps. Then the image objects are classified according to the attribute from the land-use maps (Table 4.2), e.g. in test site1 objects are classified as settlement where the attribute from ATKIS is “2101”.
Table 4.2: Exemplified results of segmentation and classification for the artificial areas in the two test sites.
Segmentation and classification results (level1)
Test site1 (Rathen) Test site 2 (Jiawang) Chessboard segmentation:
land use map as thematic layer; object size 10000
Settlements and traffic roads classification according to the land-use type in thematic layers
4.2.1.2 Assessing scale parameter for segmentation
For classifying the main classes, a lower segmentation level2 is created by means of Multi- Resolution Image Segmentation (MRIS) which aims to supply the “building” objects for all classes (see in chapter 3.2.1.3). It means the scale parameter in MRIS should be small enough to separate the pixels belonging to different classes. The tool Estimation of Scale Parameters (ESP) (Drǎguţ et al., 2010) is employed for capturing the finest scale parameter. It uses local variance (LV) graphs to reveal the multi-scale structure of images. LV equals to the average value of standard deviation of image segments. To observe the optimal scale parameter of the inflection point of LV, rate of LV-change (ROC=rate of change in local
variance between the scale level of interest and previous one) will be graphed. This graph describes the LV value changing along a given scale array. The inflection points on LV graph indicate the potential scale parameter for the image segmentation. However, this tool only applies on a single spectral band. The scale parameter calculated from one band will neglect the spectral information of the others. In practice, the estimation result of ESP can be used as a reference value for the parameter set in MRIS. In test site1, all bands from the RapidEye image of May 25th have been individually tested in ESP. All other parameters are held constant (shape 0.1 and compactness 0.5). Figure 4.3 shows the values of LV and ROC against scale levels. The smallest scale parameter is defined as the first break in the ROC- LV curve after continuous and abrupt decay. Such a threshold can appear as a small peak. For test site1, the finest scale 22 obtained from the blue and red edge band is set in MRIS (Figure 4.3 c, d). In the case of test site2, the same procedure has been applied on the image of May 14th and the finest scale 12 is observed. Table 4.3 gives an impression of segmentation results after applying MRIS in the two test sites with the parameters estimated from ESP.
Figure 4.3: Outputs of scale parameter estimation of ESP tool: (a) on red band: scale ≈ 24; (b) on green band: scale ≈ 23; (c) on blue band: scale ≈ 22; (d) on red edge band: scale ≈ 22; (e) on near infrared band: scale ≈ 27.
Table 4.3: Parameter sets and exemplified results of applying MRIS for RapidEye images on level 2.
Segmentation on level2
Test site1 (Rathen) Test site2 (Jiawang)
MRIS parameter
Scale=22; Shape=0.1; Color=0.9; Compactness=0.5; Smoothness=0.5
Scale=12; Shape=0.1; Color=0.9; Compactness=0.5; Smoothness=0.5
Segmentation result examples
4.2.1.3 Classification process
Figure 4.4 shows the applied processing chain in test site 1, which follows “left-right”, “top- down” sequence. For both test sites the settlement will be firstly classified on the segmentation level1. Then the rest main land-covers are classified on level2. The classification follows the class hierarchy from top to down (Figure 3.9). After the extraction of settlement on segmentation level1, rule-based classification is carried out at two segmentation levels and detailed rule sets are shown under each class. Firstly the land- covers are easily identified, for example the water area can be classified by NDWI (McFeeters 1996) and other land surface is divided as non-vegetation and vegetation by NDVI. Next, the vegetation objects will be further classified into sub-classes (forest, grassland, crops) based on the changing REVI value in May and August. In the end, all adjacent objects of the same classes are merged to form land-cover polygons and the three crops are merged as one farmland class.
Figure 4.4: Process for the classification of the main land-covers in test site Rathen.
Since the main land-cover types are similar in the two test sites, the same classification procedure can be applied for test site2, but the parameters for class memberships should be adapted to the local situation. Using the same segmentation strategy and multi-temporal classification approach, main land-cover maps for both test sites can be produced (see Figure 4.5).
However, the landscape configuration in two test sites remains quite different. In test site Jiawang, there are four types of crops which account for the largest proportion of the landscape. Sparse vegetation covers the north hilly area and many scatted woody patches are distributed in the agriculture area. Compared to the test area of Rathen, the landscape of Jiawang is more affected by human activities.