Information Extraction of Buildings Damaged by Earthquake

The object-oriented information extraction process of earthquake-damaged buildings is divided into three layers from top to bottom, and different segmentation scales (from large to small) are adopted according to the characteristics of different target objects, and their own fuzzy judgment rules are established based on the spectral, shape, texture and other attribute characteristics of land types, which is highly targeted; At the same time, the object-oriented multi-level classification structure system can transfer the unclassified objects in the upper level to the next level as the input layer, form the relationship between parent and child objects, and continue the information extraction and classification operations, which greatly avoids the occurrence of misclassification and omission in traditional pixel classification methods, and has a clear hierarchical structure, which improves the classification accuracy to a certain extent. The specific idea is shown in Figure 5-27.

Figure 5-27 Object-oriented information extraction method for buildings damaged by earthquakes

(1) Selection of data sources

Definitions can integrate different data (grids and vectors) as original data sources to participate in segmentation according to actual needs. In this section, aiming at the information extracted from the damaged buildings in Zipingpu town, the fusion image (4 bands) of post-earthquake standard differential vegetation index (NDVI) image and post-earthquake QuickBird image is imported into the software as the input layer.

(2) Multi-resolution segmentation

The method of "from top to bottom, from big to small" is adopted. Firstly, vegetation and water information are separated, and the weight of NDVI image is set to 3, the weight of other images is set to 1, the shape factor is set to . 3, and the compactness factor is set to . 6. After many experiments, the segmentation scale is determined to be 1.

Figure 5-28 Image segmentation results with different scales (from left to right, the segmentation scales are 48, 8 and 1, respectively)

(3) Segmentation of water body and vegetation information

First, the water body and vegetation information in the image are separated (not within the damage range of earthquake-damaged buildings). After analyzing the characteristics of normalized difference vegetation index, the threshold values (NDVI mean) of water and vegetation information extraction were determined to be . 6 and . 5, respectively. Thus, the extraction rules of vegetation and water information are established: polygonal objects with NDVI MEAN less than or equal to . 6 are water bodies, and those with NDVI mean greater than . 6 are non-water bodies, and the objects between them are classified by membership function (Figure 5-29); For vegetation, the membership function of "Boolean greater than" is adopted, and it is stipulated that the objects with an average NDVI greater than . 5 are judged as vegetation information, and those with an average NDVI less than or equal to . 5 are considered as non-vegetation information. Then merge the results of the two classifications and take the intersection (logical AND), and finally divide the objects into three categories, as shown in Figure 5-3.

figure 5-29 fuzzy membership function of water information classification

figure 5-3 extraction results of water and vegetation information

(4) extraction of temporary simple house information

take the unclassified category (urban area) in the first floor as the input layer, carry out the second segmentation and extract the temporary simple house information. Temporary simple houses are characterized by a fixed shape, which is a regular rectangle and easy to distinguish; The roof is blue-green, also known as "green tent", which is somewhat similar to vegetation; The results of structure and texture are uniform, showing homogeneity. According to the above three characteristics, 8 is selected as the scale of the second segmentation, and the following judgment rules are established:

1) The length-width ratio of the object is calculated as formula (5-7), where eig1( S) and eig2( S) are the eigenvalues of the covariance matrix, S is the covariance matrix composed of the coordinates of each point after vectorization of the object, W is the width of the object, L is the length of the object, and. After analysis, the decision rule γ≥1. 28 is set.

Study on remote sensing information extraction of degraded wasteland

2) The brightness of an object refers to the average brightness of all pixels contained in the object. Set the decision rule Brightness ＞29.

3) homogeneity of gray scale * * * generating matrix. If the texture homogeneity of the object is stronger, the value will be higher, and its calculation formula is as follows:

Study on remote sensing information extraction of degraded wasteland

Where: I is the number of rows; J is the number of columns; Pi, j is the standardized value of I row and J column; N is the total number of rows or columns, and the range is [,9]. After analysis, the GLCMHomogeneity of temporary simple houses is set to be > .18.

carry out logical intersection operation on the above three fuzzy decision rules to get the information extraction result of temporary simple house, as shown in figure 5-31.

Figure 5-31 Information extraction results of temporary simple houses with a division scale of 8

(5) Information extraction of buildings damaged by earthquake

The damaged buildings generally have the following characteristics: ① The shape characteristics of the original rules have changed due to the damage or even collapse of the buildings; (2) Compared with the adjacent features, it is out of harmony. For a completely collapsed building, the white rubble is particularly conspicuous, which forms a strong contrast with the surrounding features; ③ The damage or collapse of a building changes its original uniform texture characteristics. According to the characteristics of the buildings damaged by the above earthquake, the unclassified classes in the upper layer are segmented for the third time, and 48 is selected as the segmentation scale for the damaged building information, and the following fuzzy judgment rules are established:

1) Shape index. The shape index describes the smoothness of an object's boundary. The more regular an object's boundary is, the smaller its shape index value is. When the boundary forms a square, the shape index is 1. Conversely, the more irregular the boundary, the greater the shape index value. Therefore, the shape index can well reflect the shape damage of buildings damaged by earthquake. After analyzing the damaged buildings, the decision rule is set as shape index > 2.8.

2) gray difference vector (GLDV). Gray level difference vector is the sum of the elements on the diagonal of the gray level * * * generating matrix (GLCM), which shows the absolute difference between the object and the adjacent area pixels. The gray level difference vector contrast (GLDVContrast) reflects the difference between the object and the surrounding objects, and the greater the value, the greater the difference. Among the object-oriented remote sensing image classification methods, the gray difference vector contrast is more suitable for the hierarchical object domain than the gray * * * matrix contrast, which can highlight the shape characteristics of buildings and improve the contrast difference. Therefore, this section selects the gray difference vector contrast to establish the decision rule, and sets the decision rule as GLDV Contrast ＞78.

3) the gray * * matrix entropy (GLCM Entropy). The entropy of gray scale matrix describes the texture uniformity of the object. When the texture of the object is more uniform, the entropy value of the gray * * * generating matrix is larger; On the contrary, when the texture is dense or sparse, the entropy value of the gray * * * generating matrix is smaller. After repeated comparison and analysis, it was finally determined that GLCM Entropy was less than 4. 1.

based on the above three judging rules, the "logical and" operation is performed to obtain the extraction result of buildings damaged by earthquake, as shown in figure 5-32.

(VI) Comparison of the two methods

The information extraction results of earthquake-damaged buildings based on pixel technology and object-oriented technology are compared from three aspects: method principle, visual effect and accuracy evaluation.

1. method principle

the traditional pixel-based information extraction method is a statistical method based on spectral characteristics, which classifies images or extracts information according to the characteristic values of spectral statistics of pixel points (pixels). Pixel is the most basic operation unit, and the characteristic information used is generally only spectral information (variance, covariance, mean, brightness, entropy, etc.), which does not consider the correlation between adjacent pixels, nor does it take advantage of the rich shape, texture, spatial topological relationship and other information in high-resolution remote sensing images. With the continuous improvement of remote sensing image resolution, the amount of information contained in a single pixel is getting less and less, so it is of little practical significance to analyze and classify a single pixel in isolation, which brings great difficulties and challenges to the traditional pixel-based information extraction and classification methods, and affects the application of this method to some extent.

the object-oriented information extraction method is a new method with the continuous development of high-resolution remote sensing image technology, which is essentially different from the pixel-based information extraction method. It does not classify individual isolated pixels, but analyzes and operates polygonal objects composed of homogeneous pixels. In addition to basic spectral information, objects also include useful feature information such as shape, texture, position and context. This method can make use of more comprehensive object features, make up for the shortcomings of single feature information in pixel-based information extraction and classification methods, and is a technological innovation of traditional remote sensing information extraction and classification methods.

2. Visual effect

From the pixel-based information extraction results, there are many tiny patches, the so-called "pepper and salt effect", which seriously affects the visual effect and the practical application of the results to a certain extent. If the extraction results are vectorized, it is difficult to meet the storage requirements without manual modification. The main reason for the "pepper salt effect" is that the local heterogeneity of high-resolution remote sensing images is large, and the contextual relationship between pixels is not fully considered, which is the limitation of traditional pixel-based classification methods (Cao Bao et al., 26).

Figure 5-32 Information extraction results of buildings damaged by earthquake (division scale 48)

From the object-oriented extraction results, it has strong anti-noise ability, can effectively avoid the "pepper salt effect", and contains rich semantic information (You Liping, 27). The classified images are easier to understand, and the results can be directly vectorized and put into storage, which is convenient for the combination of remote sensing data and GIS data. From the visual effect, the object-oriented extraction results are more suitable for the requirements of human vision.

3. Accuracy evaluation

Due to the limitation of conditions, this study did not obtain the relevant reference maps of the damaged buildings in Zipingpu town, the study area. Therefore, firstly, some typical damaged buildings were sketched (classified) in the . 6m QuickBird DOM image after the earthquake by manual visual interpretation, and a total of 3,782 pixel points of damaged buildings were classified. Then, taking these pixel points as reference standards, random samples are generated, and the accuracy of the information extraction results of buildings damaged by earthquake is evaluated by error matrix method. The main evaluation indexes include producer accuracy, user accuracy, overall accuracy and Kappa coefficient.

Through statistical analysis, the total number of reference damaged buildings depicted by hand is 3782; The total number of damaged buildings extracted by object-oriented method is 3845, of which 3563 are correctly classified. The total number of damaged buildings extracted based on pixel method is 3883, of which 3132 are correctly classified. The accuracy evaluation results of the two methods are shown in Table 5-13.

Table 5-13 Accuracy evaluation results of the two methods

The accuracy evaluation results show that the overall accuracy of the object-oriented information extraction method is 9. 38% and the Kappa coefficient is . 8684, while the overall accuracy of the pixel-based information extraction method is 76. 84% and the Kappa coefficient is . 7269. It can be seen that the accuracy of object-oriented information extraction method for earthquake-damaged buildings is obviously higher than that of traditional pixel-based information extraction method. The fundamental reason is that object-oriented information extraction method combines adjacent pixels with homogeneous spectral characteristics to form polygon objects through segmentation of high-resolution remote sensing images, and then makes full use of the comprehensive analysis of the spectral, shape, texture, location, context and other attributes of the objects. The establishment of judgment rules can effectively avoid the phenomenon of "the same thing with different spectra, foreign bodies with the same spectrum" and "pepper salt effect", and obviously improve the accuracy of information extraction from high-resolution remote sensing images.