Surface information has different forms of expression on different scales (time or space span). For example, from Figure 5- 1, two circular objects are distinguished. When the observation distance is prolonged, we see Figure 5-2. At this time, we can immediately distinguish that the circular object on the left is a plate and the circular object on the right is a wheel according to its relationship with adjacent objects. This is a simple example on the spatial scale. The time scale is even simpler. For example, a piece of cultivated land is green in summer and turns yellow in autumn. The above examples show that when we want to correctly identify the target object, we must choose the appropriate scale to achieve the best resolution effect. Traditional pixel-based information extraction methods are all carried out on the same scale, which is the spatial resolution of the image. Because we can't give consideration to the macro and micro features of ground objects, when the image information is very rich (high-resolution image), we often can't get a good extraction effect, and many broken areas appear, which is what is commonly called the "salt and pepper effect" of high-resolution image classification. In order to solve this problem, the object-oriented classification method introduces the concept of multi-scale segmentation.
Figure 5- 1 Two circular objects
Figure 5-2 Plate and Wheel
(A) the concept of multi-scale segmentation
Multi-scale segmentation refers to the process of generating meaningful polygon objects with minimum heterogeneity and maximum homogeneity on any scale on the premise of minimum image information loss. It is a means of image abstraction (compression), that is, the information of high-resolution pixels is left to low-resolution objects, and different types of ground objects can be reflected on objects of corresponding scales (Huang Huiping, 2003). Multi-scale image segmentation starts from any pixel and uses bottom-up region merging method to form objects. Small objects can be merged into large objects through several steps, and the adjustment of the size of each object must ensure that the heterogeneity of the merged objects is less than a given threshold (Rebecca et al., 2009). Therefore, multi-scale segmentation can be understood as a local optimization process, and the homogeneity standard of an object is determined by the color factor and the shape factor of the object, which respectively represent the respective weights of "color" and "shape" in image segmentation, and the sum of them is 1. The "shape factor" consists of smoothness and compactness, and the sum of their weights is 1. These four parameters * * * determine the segmentation effect (Figure 5-3).
Figure 5-3 Parameter Composition of Multiscale Segmentation
(2) Selection of multi-scale segmentation parameters
Uniformity standards include spectrum (color) and shape, in which shape factors include smoothness and compactness. In most cases, the color factor is the most important factor to generate the target, and the shape factor effectively controls the fragmentation degree of the image target, which can avoid the phenomena of "the same thing with different spectra", "foreign objects with the same spectrum" and "salt and pepper effect" and improve the classification accuracy (Tian, 2007). The traditional pixel-based method does not consider the shape factor, but sets the spectral factor as 1, that is, it completely depends on the spectral value of the pixel for information extraction. Smoothing is to optimize the image object through boundary smoothing, which describes the similarity between the object boundary and a square. Compactness is to optimize image objects through the degree of aggregation, and its function is to distinguish compact and non-compact target objects with little difference. Smoothness and compactness interact and influence each other, but they are not completely opposite, that is, the object with smoothness optimization may also have good compactness, and conversely, the object with compactness optimization may also have smooth boundaries.
When setting parameters, we should first make clear the importance of spectral information and make full use of spectral (color) information. If the weight of the shape factor is too high, it will destroy the homogeneity of the object, and an object will contain multiple land types, which is not conducive to information extraction. Therefore, multi-scale segmentation should follow two basic principles: ① use a larger color factor as much as possible; (2) If you encounter an image object with uneven boundary but high concentration, you can try to control it with a larger shape factor.
(3) Selection of segmentation scale
An outstanding contribution of multi-scale segmentation is that remote sensing image information with the same spatial resolution is no longer represented by only one scale, but can be described by multiple suitable scales in the same remote sensing image (Huang Huiping, 2003). Multi-scale segmentation not only generates meaningful image objects, but also extends the original resolution image information to different scales, realizing multi-scale expression and description of information. Multi-scale segmentation refers to the use of different segmentation scale values in the process of image segmentation, and the size of the generated object depends on the segmentation scale values determined before segmentation. The larger this value is, the larger the polygon area of the generated object is, and the fewer the number is, while the smaller the polygon area is, the more the number is. Therefore, the choice of image segmentation time scale is very important, which directly determines the accuracy of classification and information extraction.
The determination of the optimal scale has always been the research focus of object-oriented classification methods, but the optimal scale is relative to a specific goal or requirement, and the optimal segmentation scale value of a specific variable may not be applicable to other variables, so the optimal scale can only be a numerical range. However, the selection of segmentation scale should follow the following rules: for a specific feature class, the most suitable scale means that the boundary of the segmented object is clear, and this feature class can be represented by one or more objects, which can neither be too fragmented nor too mixed, and a single object can well express the unique attribute characteristics of this feature class, thus being well distinguished from other feature classes (Huang Huiping, 2003). Generally speaking, the smaller the segmentation scale, the more "pure" the object is, and the smaller the probability that different features are divided into a single object, so the higher the accuracy of information extraction; However, the smaller the segmentation scale, the greater the difference between objects in the same feature category, while the lower the heterogeneity between objects in different feature categories, which is not conducive to classification and recognition. Moreover, the number of segmented objects is too large and fragmented, which increases the computational complexity of the computer and reduces the accuracy of extraction, which is not desirable. Therefore, it is necessary to find a balance between segmentation scale and classification accuracy.
(d) Multi-scale segmentation of network level.
Different segmentation scales generate corresponding object layers, thus constructing a hierarchical network between image objects, and expressing the information contained in images with different spatial scales. Each object has its neighborhood (left and right) objects, upper parent objects and lower child objects (Figure 5-4). The hierarchical structure of the object network is arranged from large to small and from top to bottom: the original layer (pixel layer) is placed at the bottom and the largest scale is placed at the top. The smaller the division ratio, the more objects the layer contains, and the fewer pixels each object contains. In a layer with a large division ratio, a single object contains more pixels and fewer objects. In this object network hierarchy, each object contains complex attribute relationships among neighbor objects, subordinate child objects and superior parent objects. When dealing with these relationships, the relationship between superior and subordinate objects is particularly important, because the child object categories can usually be determined according to the attributes of the parent object, the texture attributes of the parent object can be classified according to the average attributes of the child object, and the parent object can be classified according to the child object composition of the determined categories. In addition, adjacent objects are also very important, because if some objects are very similar in spectrum, texture and shape information, it is much easier to extract information if their objects are used as classification criteria.
Figure 5-4 Network Hierarchy Structure Diagram of Multiscale Segmentation
(5) Region merging algorithm based on the principle of minimum heterogeneity.
Multi-scale segmentation adopts a region merging algorithm based on the principle of minimum heterogeneity. Its basic idea is to gather adjacent pixels with the same or similar attributes to form a region polygon (object). First, a seed pixel is found in each region that needs to be segmented as the growth starting point, and then pixels with the same or similar attributes in the neighborhood of the seed pixel are merged into the region where the seed pixel is located, and the above process is continued as a new seed pixel until there are no more pixels that meet the conditions, thus generating a region (object) (Zhang Yujin, 2000). The purpose of region merging algorithm is to realize the minimum weight heterogeneity of segmented image objects. If only the smallest spectral inhomogeneity is considered, the boundary of the segmented object will be broken. Therefore, it is necessary to combine the spectral heterogeneity standard with the spatial heterogeneity standard. Before segmentation, it is necessary to determine the spectral factors and shape factors that affect heterogeneity, because only by minimizing spectral heterogeneity, smooth heterogeneity and dense heterogeneity at the same time can the average heterogeneity of all objects in the whole image be minimized (Dai Changda et al., 2004).
(6) Fuzzy classification
The object-oriented remote sensing image classification method adopts fuzzy mathematical analysis method based on fuzzy logic classification system. Fuzzy theory was put forward by Professor Zade in Berkeley, California in 1965, which is mainly used to deal with fuzzy, imprecise and ambiguous problems. Fuzziness is a common phenomenon in the objective world (Chen Wenkai, 2007). The fuzziness in remote sensing images is mainly manifested in the fact that there may be many types of ground objects in an object (pixel). How to determine its ownership in this case.
Fuzzy classification system generally includes fuzzification, fuzzy reasoning and deblurring. Fuzzification is a process of transforming eigenvalues into fuzzy values, which is essentially a process of feature standardization. Membership function is a fuzzy expression, which can transform any eigenvalue range into a unified range [0, 1]. Fuzzy reasoning refers to the establishment of relevant fuzzy judgment rules for fuzzy sets and the final reasoning. In fact, deblurring is the process of finally determining the result through fuzzy reasoning and comprehensive evaluation methods.
The objects segmented from remote sensing images no longer belong to a specific category of ground objects, but are related to this category to varying degrees. The relationship between them is no longer a rigid relationship between "yes" and "no", but uncertain. Fuzzy classification method is a mathematical analysis method that quantifies uncertain states. Fuzzy classification method has the following three advantages: ① the transformation from eigenvalue to fuzzy value is actually a process of feature standardization; (2) The combination of features is allowed, and even features with different ranges and sizes can be combined as classification rules; (3) Provide flexible and adjustable feature description, and carry out complex classification and information extraction through fuzzy operation and analytic hierarchy process (Zhang Yongsheng et al., 2004).
The technical process of object-oriented classification in this study is shown in Figure 5-5.
Figure 5-5 Technical Flow Chart of Object-oriented Classification Method for Ground Objects