CN106780503A - Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method - Google Patents

Abstract

The invention provides a method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy. The method includes multi-scale segmentation of remote sensing images, and automatic selection of the optimal segmentation scale of a single object based on information entropy change indicators. This method obtains the posterior probability vector of each pixel based on the SVM classifier, and then calculates the average posterior probability vector of each segment at different scales, and then calculates the information entropy of each segment at different segmentation scales to obtain each According to the entropy curve of the segmented body to which the pixel belongs under different segmentation scales, the first-order difference between the entropy value and the segmentation scale is calculated according to the entropy value curve, and the scale with the largest change in entropy value is found, and finally this scale is used as the optimal segmentation of the current pixel scale. This method effectively overcomes the "over-segmentation" and "under-segmentation" problems in object-oriented segmentation, and at the same time realizes the automatic selection of the optimal scale of the object. It is an effective automatic segmentation method for ground objects.

Description

Determination Method of Optimal Segmentation Scale of Remote Sensing Image Based on Posteriori Probability Information Entropy

technical field

The invention relates to a processing method of remote sensing images, in particular to a method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy, belonging to the field of image processing.

Background technique

High-spatial-resolution remote sensing images are important data for humans to accurately monitor land cover. Pixel-based image analysis methods generally have the problem of "salt and pepper phenomenon". Phenomenon. This phenomenon severely restricts the extraction accuracy of ground targets in high spatial resolution remote sensing images. In order to solve this problem, the object-oriented image analysis method is gradually paid attention to and adopted by researchers in the industry.

With the development of remote sensing science and technology, commercial spaceborne high spatial resolution images and airborne high spatial resolution images have been widely used in urban planning and land cover mapping. Image segmentation is to divide the spectral image into a series of segments, and each segment is composed of a group of homogeneous pixels. Among them, the segmentation scale is an important indicator to determine the size and homogeneity of the segmentation body. The smaller the segmentation scale, the smaller the area of the segmentation body and the higher the degree of homogeneity; the larger the segmentation scale, the larger the area of the segmentation body. The degree of homogeneity is also lower. In the current segmentation methods, people mainly rely on experience and trial-and-error methods to choose the optimal segmentation scale, which hinders the universality of the segmentation method and reduces the efficiency of the segmentation method.

In the high spatial resolution images, each house and each piece of lawn has its real physical scale, that is, the size of the segment that completely covers these features. However, in the process of image segmentation, any single segmentation scale cannot simultaneously conform to the real physical scale of all ground objects in the image. Therefore, the multi-scale segmentation method is more suitable for solving the problem of automatic target extraction. Finding the optimal segmentation scale for each ground object in the image will greatly improve the recognition accuracy of ground targets.

In recent years, many researchers have devoted themselves to solving the problem of choosing the optimal scale for image segmentation. LucianDragut et al proposed to use local variance to automatically determine the optimal segmentation scale of images, see "a tool to estimate scale parameter for multiresolution image segmentation of remotely sensed data" (Dragut L , International Journal of Geographic Information Science, 2010, 24(6): 859-871). This method uses the same optimal scale for all objects in the image. However, in most cases, objects in an image have different sizes, and it is obviously unreasonable to apply the same segmentation scale to objects of different sizes for segmentation. In addition, T-Esch et al. proposed a method to automatically determine the optimal segmentation scales for different objects, see "Improvement of image segmentation accuracy based on multiscale optimization procedure" (Esch T, Journal of Earth Science and Remote Sensing, IEEE, 2008,5(3):463-467). However, a large number of parameters are used to determine the optimal scale, which makes the whole algorithm quite complicated. At the same time, for different application purposes, there are different optimal segmentation scales even for the same image. For example, a relatively large scale is required when a house is used as a segment to be extracted, but a relatively small scale is required when a car is used as a segment to be extracted. Therefore, we urgently need a method that can automatically determine the optimal segmentation scale of different objects in the image.

Contents of the invention

Therefore, the present invention provides a method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy, which can reduce or avoid the problems mentioned above.

In order to solve the above problems, the method for determining the optimal segmentation scale of remote sensing images based on information entropy change index provided by the present invention comprises the following steps:

Step A, perform multi-scale segmentation on the remote sensing image and calculate the pixel-level posterior probability vector: this step uses eCognition 8.9 software to obtain multi-scale segmentation results; use the SVM classifier to calculate the posterior probability vector of each pixel, and calculate the different The average posterior probability for each segment at the segmentation scale.

Specifically, firstly, use eCognition 8.9 software to segment remote sensing images at multiple scales to obtain multi-scale segmentation results. Second, select pixel-level training samples, and use the SVM classifier to calculate the posterior probability vector for each pixel of the original multi-band image. Third, according to the multi-scale segmentation results, the average posterior probability of each segment at each segmentation scale is calculated.

Step B, select the optimal segmentation scale of a single segment based on the information entropy change index: this step calculates the information entropy based on the posterior probability as the category uncertainty evaluation index of the segment, and determines it according to the change of entropy value of a single object at different segmentation scales its optimal segmentation scale.

According to the multi-scale posterior probability vector obtained in step A, the posterior probability information entropy is calculated for each segment, and the entropy calculation refers to the following:

Where P _j,i represents the probability that the j-th segment belongs to the i-th category under the s-segmentation scale, n represents the number of categories, and E _s,j represents the information entropy value of the j-th segment under the s-segmentation scale.

Based on the information entropy at each scale, the first-order difference ΔE of information entropy changing with adjacent segmentation scales is calculated, and the segmentation scale at the maximum position of the obtained information entropy change is taken as the optimal segmentation scale.

The beneficial effects that the present invention has are:

The method of the invention automatically determines the optimal segmentation scale of each segment based on the information entropy change index of the segmented body of the ground object at different scales.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

Description of drawings

The following drawings are only intended to illustrate and explain the present invention schematically, and do not limit the scope of the present invention. in,

Fig. 1 is a schematic flow chart of a method for determining a segmentation scale of a remote sensing image based on posterior probability information entropy according to a specific embodiment of the present invention;

Figure 2 is a schematic diagram of the inclusion relationship between segments at adjacent segmentation scales of eCognition;

Fig. 3 is the schematic diagram of the principle of the method provided by the present invention;

Fig. 4(a) and Fig. 4(b) are effect diagrams of two examples of methods for determining optimal segmentation scales of remote sensing images based on posterior probability information entropy provided by the present invention.

detailed description

In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific implementations of the present invention are now described. But those skilled in the art should know that the following examples are not the sole limitation to the technical solution of the present invention, and any equivalent transformation or modification made under the spirit of the technical solution of the present invention should be considered as belonging to the protection of the present invention scope.

Fig. 1 is a schematic flow chart of a remote sensing image segmentation scale determination method based on posterior probability information entropy according to a specific embodiment of the present invention; with reference to Fig. 1, the following describes in detail the remote sensing based on information entropy change index provided by the present invention The principle of the image optimal segmentation scale determination method, the method includes the following two steps:

Step A, perform multi-scale segmentation on the remote sensing image and calculate the pixel-level posterior probability vector.

Step B, selecting the optimal segmentation scale based on the information entropy change index.

For remote sensing images, before performing the processing process of the present invention, it is preferable to perform atmospheric correction on the image to remove the atmospheric image. In addition, for hyperspectral remote sensing images, in order to reduce the amount of calculation, it is preferable to use principal component analysis (PCA, Principle Component Analysis) to perform dimensionality reduction processing on hyperspectral images before performing the processing process of the present invention. In the process of image feature extraction, too many feature dimensions extracted often lead to too complex feature matching and consume system resources. Therefore, feature dimensionality reduction methods are often used to process hyperspectral images. The so-called feature dimensionality reduction means using a low-dimensional feature to represent a high-dimensionality. Feature dimensionality reduction usually uses PCA for feature extraction for dimensionality reduction. The above preferred steps are all adopted before the processing steps of the present invention. For example, the atmospheric correction can be performed first and then the dimensionality reduction treatment can be performed, or the dimensionality reduction treatment can be performed first and then the atmospheric correction can be performed. similar.

Steps A and B are described in detail below:

Step A, perform multi-scale segmentation on the remote sensing image and calculate the pixel-level posterior probability vector

In this step, the multi-scale segmentation of the remote sensing image is first performed, and then the posterior probability vector of each segment in the multi-scale segmentation result is calculated.

The present invention uses eCognition 8.9 software to first segment remote sensing images in multiple scales. For segmentation, the shape factor was set to 0.2 and the compactness factor was set to 0.5. In this example, the step size is 10, increasing from 50 to 1000, and the selected multi-level segmentation scales are 50, 60, 70, 80, ..., 970, 980, 990, 1000 (see Figure 2 for the segmentation principle of eCognition software). Among them, the segmentation scale is positively correlated with the area of the segmented volume, and negatively correlated with the homogeneity of the types of pixels inside the segmented volume.

Although the eCognition software can produce segmentation results of different segmentation scales (eg 50, 60, ..., 980, 990, 1000), these segmentation results all use the same segmentation scale to segment all objects in the image. In actual situations, different types of ground objects have different physical scales, namely geometric parameters such as area, perimeter, and diameter. For example, houses and grasslands have different patch areas in the image, so they should be segmented at different segmentation scales during image segmentation. To sum up, the segmentation results directly generated by eCognition software cannot be directly used for ground target recognition of remote sensing images.

In order to solve this technical problem, the present invention calculates the information entropy change index and comprehensively utilizes the segmentation results of eCognition from "over-segmentation" to "low-segmentation" (corresponding to a segmentation scale of 50 to 1000), and automatically selects each feature in the image. The optimal segmentation scale finally produces a multi-scale segmentation result, in which, as shown in Figure 2, the segmentation results of adjacent segmentation scales with scales from large to small have a relationship between the full set and the subset.

That is, when segmenting, create a rule set in eCognition 8.9 software according to the selected segmentation scale range (the method of creating a rule set is described in detail in the eCognition software manual), and segment from small to large scales step by step. After segmentation, the segmentation results of different scales generated by eCognition software are exported as several raster images (Tiff format or img format), and the segmentation results of each segmentation scale correspond to a raster image. In each raster image, the pixels in each segment are assigned the same label value DN, that is, the DN value is the label of the segment. In a preferred embodiment, the selection of the multi-level segmentation scale can also be determined according to the range of the DN value or the reflectivity of the image.

Then calculate the posterior probability vector for each pixel. The present invention uses a support vector machine (SVM, support vector machine) classifier to calculate the posterior probability vector of a single pixel. The SVM classifier is a problem of converting complex classification tasks into a problem of constructing a linear classification hyperplane in a high-dimensional feature space through kernel function mapping. The optimal classification hyperplane can be obtained by solving a quadratic programming problem. The vertical Euclidean distance between sample points of different categories and the hyperplane can be converted into the posterior probability of this type of samples, and the vector composed of the posterior probabilities of all categories is called the posterior probability vector. SVM classifier is a well-developed classification method. The code and principle of this method can be obtained free of charge through public network platforms. Journal of the Chinese Academy of Sciences, 2010, 24(6): 859-871).

That is, before using the SVM classifier to calculate the posterior probability vector of each segment in the multi-scale segmentation results, the sample training is used to obtain the posterior probability vector of each pixel. The specific steps are to select the pixel-level training samples and use The SVM classifier calculates the posterior probability vector of each pixel, for example, selects 3000-5000 pixels from each land cover type in the original multi-band image as a training sample, and uses the SVM classifier to calculate the posterior probability vector of each pixel vector of posterior probabilities.

Then, based on the multi-scale segmentation result in the form of a grid obtained in the above steps, and the posterior probability vector of each pixel, the average posterior probability vector of each segmented body at each segmentation scale is calculated respectively. Assuming that the j-th segment contains n pixels at the s-th segmentation scale, the average posterior probability vector of the segment The calculation method is as follows:

in Indicates the average posterior probability of the k-th category of the current segment, C represents the total number of categories, n represents the number of all pixels in the current segment, p _i,k represents the k-th category of the i-th pixel Posterior probability. Following the method described above, the posterior probability vectors for each object at all segmentation scales are calculated.

Step B, choose the optimal segmentation scale based on the information entropy change index

According to the multi-scale posterior probability vector obtained in step A, it can be obtained: when the segmented body is over-segmented (that is, when the segmentation scale is small), the pixels contained in the segmented body are basically of the same category, and the category of the internal pixels is Greater homogeneity. When the segmentation body is under-segmented (that is, when the segmentation scale is large), the segmentation body contains more pixels of different types, and the category homogeneity of the internal pixels is lower. Based on the above ideas, the present invention proposes to use information entropy as an index to measure the homogeneity of categories within a segment. According to the multi-layer posterior probability obtained in step A (the segmentation results of the adjacent segmentation scales from large to small scales have the relationship between the complete set and the subset, and the segmentation results of different segmentation scales are expressed in layers, see Figure 2), for each A segmentation body calculates the entropy of the posterior probability in the order of increasing segmentation scale. The entropy calculation formula is as follows:

Where P _j,i represents the probability that the j-th segment belongs to the i-th category under the s-segmentation scale, n represents the number of categories, and E _s,j represents the information entropy value of the j-th segment under the s-segmentation scale.

When the object is in the over-segmented state, the internal category of the segment is single, and the information entropy value is low; similarly, when the object is in the under-segmented state, the segmented body is large, and the mixture of multiple categories leads to a large entropy value. At this time, based on the information entropy of the segmentation volumes at different segmentation scales, the first-order difference of the information entropy changing with adjacent segmentation scales can be calculated (for example, when there are only two categories in the image (the posterior probabilities are respectively p ₁ and p ₂ ), where p ₁ is the posterior probability of the true category of the target object, and p ₂ is the posterior probability of other categories around the target object). The first-order difference of information entropy is expressed as follows:

where ΔE represents the first-order difference of information entropy, It is the partial derivative of information entropy for each category. When _p2 is the smallest, ΔE is the largest. This means that when the change in information entropy is the largest, the segment contains adjacent impurity pixels. Therefore, in the information entropy curve that changes with the segmentation scale, according to the above derivation, the segmentation scale at the position of the maximum value of information entropy change ΔE _max is regarded as the optimal segmentation scale, and the segmented body of this scale is the final segmentation result , that is, in the (50, 60, 70, ..., 990, 1000) multi-scale segmentation results of the previous steps, calculate the first-order difference of adjacent segmentation scale changes, and obtain the maximum information entropy change value ΔE _max :

Finally, the segmentation scale with ΔE _max is used as the optimal segmentation scale. Fig. 3 is a schematic diagram of selecting the optimal segmentation scale of a single object in the present invention.

In order to better illustrate the technical effect of the present invention, for a hyperspectral image, the method for determining the optimal segmentation scale of remote sensing images based on the posterior probability information entropy proposed by the present invention, the multi-scale segmentation algorithm of the variance stable interval and the Experiments are performed on single-scale segmentation algorithms, and then the results after segmentation are compared. Among them, the scale used by the single-scale segmentation algorithm is manually selected: the segmentation results at 50, 60, 70, ..., 990, 1000 segmentation scales are respectively generated, and the scale with the highest segmentation accuracy is selected. In addition, in order to verify the stability of the method, the present invention has carried out multiple tests under different sample sizes and different segmentation step conditions. Various sample sizes were increased from 1000 to 9000 pixels at intervals of 2000. In order to reduce the amount of calculation, before carrying out the processing of the method provided by the invention, at first utilize principal component analysis algorithm (principal component analysis method is a kind of mathematical method that is generally used, and the software, code and principle of this method all can be in public network platform for free) to reduce the dimensionality of the image, and extract the first 8 bands for the processing of the method provided by the present invention.

Compared with the multi-scale segmentation algorithm using the variance stable interval, the "salt and pepper phenomenon" is effectively removed by using the method for determining the optimal segmentation scale of remote sensing images based on information entropy change indicators provided by the present invention to perform multi-scale segmentation results of remote sensing images Due to the segmentation error caused by noise, most buildings and lawns are presented on a real physical scale.

Fig. 4(a) and Fig. 4(b) are respectively the result diagrams of two samples obtained by the method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy provided by the present invention. It can be seen from Fig. 4 that the method provided by the present invention can find the optimal segmentation scale of each ground object, and all the segmentation bodies under the optimal segmentation scale form the multi-scale segmentation result.

Table 1 is the method for determining the optimal segmentation scale of remote sensing images based on the posterior probability information entropy provided by the present invention, the single-scale segmentation algorithm, and the multi-scale segmentation algorithm using the variance stable interval, and the overall precision statistics of the multi-scale segmentation results . Wherein, the algorithm related to the comparison with the method of the present invention and segmentation principles, terms, etc. can be found in "Multi-scale segmentation approach for object-based land-cover classification using high spatial resolution images" (Multi-scalesegmentation approach for object-based land-cover classification using high-resolution imagery) (Zhang et al., Remote Sensing Letters, 2014, 5(1):73-82).

It can be seen from Table 1 that the segmentation accuracy obtained by using the method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy provided by the present invention is the highest. Among them, the multi-scale segmentation algorithm proposed by the present invention is better than the other two methods under the two inspection indexes of F-measure and BCI. For effective evaluation of segmented objects in the three segmentation states, see "Segmentation quality evaluation using region-based precision and recall measures for remote sensing images" (Segmentation quality evaluation using region-based precision and recall measures for remote sensing images) (Zhang et al., ISPRS Journal of Photogrammetry and Remote Sensing, 2015, 102(5), 73-84).

Table 1

Table 2 is the method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy provided by the present invention, and the multi-scale segmentation result accuracy evaluation under different segmentation scale step lengths and different training sample sizes; as can be seen from Table 2 It is concluded that the method provided in this study can maintain relatively stable segmentation accuracy at different segmentation scale intervals; when the sample size reaches 5000 pixels, the segmentation accuracy is the highest. In addition, under all segmentation step sizes and training sample conditions, the segmentation accuracy obtained by the method provided by the present invention is higher than that of the other two methods, which shows that the method proposed in this study is an effective method for image segmentation.

Table 2

It can be seen from the above chart examples that the method for determining the optimal segmentation scale of remote sensing images based on posterior probability information entropy provided by the present invention can effectively segment ground objects in high-resolution images.

Those skilled in the art should understand that although the present invention is described in terms of multiple embodiments, and several methods are compared horizontally, not every embodiment only includes an independent technical solution. The description in the description is only for the sake of clarity, and those skilled in the art should understand the description as a whole, and understand the present invention by considering the technical solutions involved in each embodiment as being able to be combined with each other to form different embodiments scope of protection.

The above descriptions are only illustrative specific implementations of the present invention, and are not intended to limit the scope of the present invention. Any equivalent changes, modifications and combinations made by those skilled in the art without departing from the concept and principle of the present invention shall fall within the protection scope of the present invention.

Claims (7)

1. A remote sensing image optimal segmentation scale determining method based on posterior probability information entropy is characterized by comprising the following steps:

step A, carrying out multi-scale segmentation on the remote sensing image and calculating a pixel-level posterior probability vector:

the remote sensing image is segmented by utilizing eCoginization 8.9 software with 10 as segmentation step length and sequentially in (50,60,70, …,990,1000) segmentation scales to obtain a multi-scale segmentation result;

selecting pixel element training samples, calculating the posterior probability vector of each pixel element by using an SVM classifier, and calculating the average posterior probability vector of each segmentation body under each segmentation scale based on the multi-scale segmentation and the posterior probability vector of each pixel element.

B, selecting an optimal segmentation scale based on the information entropy change index:

according to the multi-scale posterior probability vector obtained in the step A, entropy of the posterior probability is respectively calculated for each segmentation body according to the ascending sequence of the segmentation scale, and a calculation formula of the entropy is as follows:

$E_{s,j}=-\underset{i=1}{\overset{n}{\Σ}}{p}_{s,j,i}\×{\log }_2\left(P_{s,j,i}\right)$

wherein P is_j,iRepresenting the probability that the jth segmentation body belongs to the ith class under the s segmentation scale, n represents the number of classes, E_s,jAnd representing the information entropy value of the j-th segmentation body under the s-segmentation scale.

And calculating a first-order difference delta E of the information entropy along with the change of adjacent segmentation scales on the basis of the information entropy under each scale, and taking the segmentation scale at the position of the maximum value of the obtained information entropy change as an optimal segmentation scale.

2. The method according to claim 1, further comprising a process of performing dimensionality reduction processing on the hyperspectral remote sensing image by using a principal component analysis method in advance.

3. The method of claim 1 or 2, wherein the step a further comprises: before an SVM classifier is used for calculating the posterior probability vector of each segmentation body in the multi-scale segmentation result, a training sample is selected from an original multi-band image, and the posterior probability vector of each pixel is calculated by the SVM classifier.

4. The method of claim 1 or 2, wherein the step a further comprises: when the remote sensing image is subjected to multi-scale segmentation by using eCoginization 8.9 software, the selection of multi-segmentation scales is determined according to the DN value or the range of reflectivity of the image, the shape factor is set to be 0.2 and the compactness factor is set to be 0.5 during segmentation.

5. The method of claim 1 or 2, wherein the step a further comprises: and creating a rule set in eCooginion 8.9 software according to the selected segmentation scale range, and segmenting from small to large according to the scale.

6. The method of claim 1 or 2, wherein the step a further comprises: after the division, the generated division results of different scales are exported to be raster images, and the pixels in each division body in each raster image are endowed with the same label value as the label of the division body.

7. The method of claim 1 or 2, wherein the step a further comprises: and respectively calculating the average posterior probability vector of each segmentation body under each segmentation scale based on the multi-scale segmentation result in the grid form obtained in the step and the posterior probability vector of each pixel. Assuming that the jth partition contains n pixels at the s-th partition scale, the partitionVolume average posterior probability vectorThe calculation method of (2) is as follows:

$\stackrel{\‾}{P}=\[{\stackrel{\‾}{p}}_1,...,{\stackrel{\‾}{p}}_k,...,{\stackrel{\‾}{p}}_C\]$

${\stackrel{\‾}{p}}_k=\frac{\underset{i=1}{\overset{n}{\Σ}}{p}_{i,k}}{n}$

whereinRepresenting the average posterior probability of the kth class of the current segment, C representing the total number of classes, n representing the number of all pixels in the current segment, p_i,kRepresenting the posterior probability of the kth class of the ith pixel. According to the above method, a posterior probability vector of each object in all segmentation scales is calculated.