CN104881677B - Method is determined for the optimum segmentation yardstick of remote sensing image ground mulching classification - Google Patents

Abstract

The invention provides a method for determining the optimal segmentation scale for classification of remote sensing image land cover. The method includes the steps of multi-scale segmentation and classification of remote sensing images, and optimal scale selection based on entropy information. This method integrates two classification methods based on pixel and object-oriented, and makes full use of sample information to classify remote sensing images. This method effectively overcomes the problem of a large amount of 'salt and pepper' noise produced by traditional pixel-oriented methods, and at the same time realizes the automatic selection of the optimal scale of the object, and provides an effective method for the mapping of land cover.

Description

Optimal Segmentation Scale Determination Method for Surface Cover Classification of Remote Sensing Images

technical field

The invention relates to a processing method for remote sensing images, in particular to a method for determining an optimal segmentation scale for land cover classification of remote sensing images, and belongs to the field of image processing.

Background technique

Land cover refers to the complex of surface elements covered by natural structures and artificial buildings, including surface vegetation, soil, lakes, swamps and various buildings (such as roads, houses, etc.). Land cover is an important forcing factor of global environmental change, which has attracted more and more researchers' attention in recent years.

With the development of remote sensing science and technology, the resolution of remote sensing images is getting higher and higher, which provides feasibility for the mapping of land cover on various spatial scales. The classification of remote sensing images is an important link in land cover mapping, which determines the quality of land cover mapping. At present, the methods for remote sensing image classification are mainly divided into two categories: (1) pixel-oriented methods, (2) object-oriented methods.

Pixel is the basic unit of remote sensing image, and it is the most concise and effective method to classify remote sensing image by using the statistical information of pixel. However, in medium and high-scale remote sensing images, the ground area corresponding to a single pixel is small. Therefore, affected by the complexity of the ground, the use of pixel-based methods for remote sensing image classification will produce a large number of 'pepper and salt' noise. This greatly reduces the accuracy of remote sensing image classification.

In order to overcome the 'salt and pepper' noise produced by the pixel-based method, a classification method combining ground texture and spatial information has emerged quietly - object-oriented classification. Different from pixel-based methods, object-oriented methods first segment remote sensing images into spectrally homogeneous and spatially continuous independent objects, and then classify the objects. This can effectively remove the 'salt and pepper' noise and improve the classification accuracy. However, the size of the object obtained by image segmentation depends on the segmentation scale parameter of the image. A larger segmentation scale will cause the ground objects to be under-segmented, and on the contrary, a smaller segmentation scale will cause the ground objects to be over-segmented. Obviously, it is proved that both over-segmentation and under-segmentation of images will lead to a decrease in classification accuracy, see Liu D, XiaF. Assessing object-based classification: advantages and limitations[J]. RemoteSensing Letters, 2010, 1(4): 187- 194.. Unfortunately, the determination of the optimal scale often requires experimenting with different segmentation scales and relying on experience to determine the optimal segmentation scale. This not only requires a lot of manpower, but also it is difficult to guarantee the accuracy of the optimal scale obtained.

In recent years, many researchers have devoted themselves to solving the problem of choosing the optimal scale for image segmentation. Lucian Drǎgut (2009) proposed to use local variance to automatically determine the optimal segmentation scale of the image, see L, Tiede D, Levick S R.ESP: a tool to estimate scale parameter formultiresolution imagesegmentation of remotely sensed data[J].International Journal of Geographical Information Science,2010,24(6):859-871. Objects use the same optimal scale. 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 (2008) proposed a method to automatically determine the optimal segmentation scale for different objects, see Esch T, Thiel M, Bock M, et al.Improvement of imagesegmentation accuracy based on multiscale optimization procedure[J] .Geoscience and Remote Sensing Letters, 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 taking a house as an object to be extracted, but a relatively small scale is required when taking a car as an object 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

For this reason, the present invention provides a method for determining the optimal segmentation scale for classification of remote sensing image land cover, which can reduce or avoid the problems mentioned above.

In order to solve the above problems, the method for determining the optimal segmentation scale for the classification of remote sensing image land surface coverage provided by the present invention comprises the following steps:

Step A, multi-scale segmentation and classification of remote sensing images;

In this step, the segmentation software capable of generating multi-scale segmentation results is used to obtain multi-scale segmentation results, and the average spectrum of each image object after segmentation is calculated by fusing the pixel-level sample information, and the average spectrum is classified to obtain various segmentation results under different segmentation scales. The classification result of the object and the posterior probability vector;

Step B, optimal scale selection based on entropy information;

In this step, according to the multi-scale posterior probability vector obtained in step A, the entropy of the posterior probability is calculated for each object in the order of increasing segmentation scale. The entropy calculation formula is as follows:

Where P _i represents the probability that the object belongs to the i-th category, n represents the number of categories, select the segmentation scale with the smallest entropy value as the optimal segmentation scale of the object, and let the classification result of the object under the optimal segmentation scale be the final category of the object .

Further speaking, the specific implementation method of the step A is:

Firstly, the remote sensing image is segmented at multiple scales. The selection of multiple segmentation scales is determined according to the DN value of the image or the range of reflectivity. When segmenting, set the shape factor and compactness factor according to the image characteristics;

Secondly, create a rule set in the segmentation software according to the selected segmentation scale range, and segment the scale step by step in the order of scale from large to small or from small to large;

Then, the objects of the multi-scale segmentation results are formed into a numbered file, which is exported as a raster image in turn, and the value corresponding to each pixel in the raster image is the number of the object;

Then, according to the numbered file of the object after the obtained image segmentation and the original multi-band image, the average spectrum of each object at each segmentation scale is calculated separately to obtain multi-scale average spectral data;

Finally, select training samples from the original multi-band images to classify the average spectral images of objects at different scales obtained above, and the selected training samples are required to be typical and random;

At the same time of classification, the posterior probability vector of objects at multiple scales is obtained. For an object on a certain segmentation scale, its posterior probability vector is expressed as follows:

P=(P ₁ , P ₂ ..., P _i ..., P _n )

Among them, P _i represents the probability that the object belongs to the i-th category, and n represents the number of categories.

Preferably, the shape factor is set to 0.2, and the compactness factor is set to 0.5.

Preferably, the selection of the segmentation scale is carried out according to the principle that as the scale increases, the scale interval also increases accordingly.

The method of the invention combines two methods based on pixels and object-oriented methods, and makes full use of sample information to classify remote sensing images. This method effectively overcomes the problem of a large amount of 'salt and pepper' noise produced by the traditional pixel-oriented method, and at the same time realizes the automatic selection of the optimal scale of the object, and provides an effective method for the mapping of land cover.

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,

1 is a schematic flow diagram of an optimal segmentation scale selection algorithm for remote sensing image land cover classification according to a specific embodiment of the present invention;

Fig. 2 is the schematic diagram of the principle of the method provided by the present invention;

Fig. 3 is a classification according to the optimal segmentation scale determination method for remote sensing image land cover classification provided by the present invention, and an overall accuracy statistical diagram of the results obtained by using the pixel-based method and different segmentation scale object-oriented methods for classification;

Fig. 4 is a classification according to the optimal segmentation scale determination method for remote sensing image land cover classification provided by the present invention, and the overall accuracy of the results obtained by using the pixel-based method and the optimal single-scale object-oriented method for classification with the sample size Change chart.

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 method for determining the optimal segmentation scale for land cover classification of remote sensing images according to a specific embodiment of the present invention; referring to Fig. 1, the method for land cover classification of remote sensing images provided according to the present invention will be described in detail below The principle of the optimal segmentation scale determination method, the method includes the following two steps:

Step A, multi-scale segmentation and classification of remote sensing images;

Step B, optimal scale selection based on entropy information.

For remote sensing images, before performing the processing process of the present invention, it is first necessary to perform atmospheric correction on the images to remove atmospheric influences. At the same time, for hyperspectral images, in order to reduce the amount of calculation, the present invention proposes to use Principal Component Analysis (PCA) to perform dimensionality reduction processing on hyperspectral images. Principal Component Analysis (PCA) is a well known method.

Steps A and B are described in detail below:

Step A, multi-scale segmentation and classification of remote sensing images

The present invention combines two methods based on pixels and object-oriented. For the object-oriented method, the remote sensing image needs to be segmented first. Therefore, in this embodiment, eCognition 8.9 software is used to segment remote sensing images at multiple scales. In fact, all versions of eCognition are acceptable. If other segmentation algorithm software can generate segmentation results of multiple scales, it is also applicable to this method. It is empirically determined that when segmenting, it is better to set the shape factor to 0.2 and the compactness factor to 0.5. The selection of the multi-level segmentation scale can be determined according to the DN value of the image or the range of the reflectivity. 460, 520, 580, 640, 700, 780, 860, 940, 1020, 1100 and so on. It can be seen that the scale interval increases with the increase of the scale, because with the increase of the scale, the area of the object continues to increase, and the heterogeneity difference between the objects is also increasing, and the smaller scale interval is basically The segmentation result will not be changed, so this example chooses an increasing segmentation interval scale.

Secondly, use the MRS (multiresolution segmentation) algorithm to create a rule set in the eCognition 8.9 software according to the selected segmentation scale range (the creation of the rule set is the basic function of the eCognition software, specifically referring to different segmentation parameters (shape factor, compactness factor) The settings are divided step by step according to the order of the scale from large to small or from large to small.

Then, the objects of the multi-scale segmentation results are formed into a numbered file, which is sequentially exported as a raster image, so the value corresponding to each pixel in the raster image is the number of the object. The process of forming numbered files is a function of the eCognition software itself and is also a known method.

Then, according to the numbered file of the object after the image segmentation obtained in the above steps, and the original multi-band image, the average spectrum of each object at each segmentation scale is calculated respectively, and the multi-scale average spectral data is obtained. The calculation of the average spectrum is to average the spectra of all the pixels inside the object, which is a well-known process.

Finally, training samples are selected from the original multi-band images to classify the average spectral images of objects at different scales obtained above. The selection of training samples requires that the selected samples are typical and random. The selection of training samples occurs in all supervised classifications and is a well-known process.

Either a maximum likelihood classifier (SVM) or a support vector machine classifier (MLC) can be used for the above classification process. For hyperspectral imagery, the present invention proposes to use SVM classifier, because SVM classifier has better resolving power to hyperspectral data; The classification speed is conducive to improving the overall efficiency.

The posterior probability vector of the object (the attribution probability of the object for different classification categories) can be obtained at the same time when the classifier is used for classification. For an object in a certain segmentation scale layer, its posterior probability vector can be expressed as follows:

P=(P ₁ , P ₂ ..., P _i ..., P _n )

Among them, P _i represents the probability that the object belongs to the i-th category, and n represents the number of categories. By using the above classification methods for all objects of different segmentation scales, multi-scale classification results and multi-scale posterior probability vectors can be obtained.

Step B, optimal scale selection based on entropy information

The posterior probability vector obtained according to the process described in step A can be obtained: when the object is over-segmented (that is, when the segmentation scale is small), the number of pixels contained in the object is less, which is affected by the 'salt and pepper' noise in the object , the uncertainty of object classification increases. When the object is in low segmentation (that is, when the segmentation scale is large), the object contains more pixels. Although the impact of 'salt and pepper' noise can be ignored, the mixture of different ground features within the object also increases the uncertainty of object classification.

Based on the above ideas, the present invention proposes to use entropy information as a standard for measuring the inaccuracy of object classification. According to the multi-layer posterior probability obtained in step A, the entropy values of the object posterior probability vectors are respectively obtained on different segmentation scales. The calculation formula of entropy is as follows:

Among them, P _i represents the probability that the object belongs to the i-th category, and n represents the number of categories. When the object is in an over-segmented state, the classification uncertainty caused by 'salt and pepper' noise will lead to an increase in the entropy value; similarly, when the object is in a low-segmentation state, the classification uncertainty caused by the mixture of various ground features will also lead to entropy increase in value. Its rules are shown in Figure 2. It can be obtained from this: when the posterior probability vector of the object at a certain segmentation scale has the minimum entropy value, its classification stability is the highest. Therefore, the segmentation scale with the smallest entropy value is selected as the optimal segmentation scale of the object where the pixel is located. At the same time, the object classification result under this scale is used as the final classification result of the object. As mentioned above, by implementing the above process on all objects, the optimal segmentation scales and final classification results of all objects can be obtained.

In order to better illustrate the technical effect of the present invention, for a hyperspectral image, the optimal segmentation scale determination method for remote sensing image land cover classification, the single pixel-based method and the optimal single-scale object-oriented method proposed by the present invention are used respectively. The method classifies the images, and then compares the classified results. All classification processes are completed by SVM classifiers, and the optimal single-scale object-oriented classification selects the scale with the highest accuracy among the classification results of multiple scales. The classification training sample selected by the above comparison method is completely consistent with the classification training sample of the method provided by the present invention. At the same time, in order to verify the stability of the method, the present invention has carried out multiple tests under different sample sizes. The population sample size increases from 1000 pixels to 10000 pixels at intervals of 1000. Ten experiments were performed simultaneously for each sample size, and the samples for each experiment were randomly selected from the reference classification map. In order to reduce the amount of calculation, before performing the processing of the method provided by the present invention, the principal component analysis algorithm (PCA) is first used to reduce the dimensionality of the image, and the first 8 bands are extracted for the processing of the method provided by the present invention.

Compared with the pixel-based method, the classification error caused by 'salt and pepper' noise is effectively removed by using the optimal segmentation scale selection method for remote sensing image land cover classification provided by the present invention to classify remote sensing images. ; Compared with the optimal single-scale object-oriented classification results, the method provided by the present invention can extract more detailed information and improve the classification accuracy.

Fig. 3 is a statistical diagram of the overall classification accuracy obtained by using the optimal segmentation scale determination method for remote sensing image land cover classification provided by the present invention and the pixel-based method for classification, and using different segmentation scales to classify hyperspectral images , it can be seen from the figure that the method provided by the present invention has higher accuracy than the pixel-oriented method and single-scale object-oriented classification.

Figure 4 is the overall classification obtained from experiments with different sample sizes using the optimal segmentation scale determination method for remote sensing image land cover classification provided by the present invention, and experiments with different sample sizes using the pixel-based method and the optimal single-scale method Accuracy statistics graph. It can be seen from Figure 4 that compared with the pixel-based method, the overall performance of the method provided in this study is higher than that of the pixel-based method under different sample sizes; In the case of 4000, the overall accuracy of the method proposed in this invention is slightly lower than the optimal single-scale method, but when the sample size is greater than 4000, the overall accuracy of the method proposed in this study is higher than the optimal single-scale method. This shows that under the condition of insufficient sample size, the robustness of the method provided by the present invention decreases; under the condition of sufficient sample size, the method proposed in this study is an effective method for image classification.

It can be seen from the above examples that the determination of the optimal segmentation scale for land cover classification of remote sensing images according to the present invention can effectively classify remote sensing images. The optimal segmentation scale determination method for remote sensing image land cover classification provided by the present invention effectively combines pixel-based and object-oriented classification methods, effectively overcomes the problem of a large number of "pepper and salt" noise points generated by pixel-oriented, and simultaneously realizes Automatic selection of optimal scale for objects. Provides an effective method for land cover mapping.

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.

Claims (4)

1. An optimal segmentation scale determination method for remote sensing image earth surface coverage classification is characterized by comprising the following steps:

a, multi-scale segmentation and classification of remote sensing images;

acquiring a multi-scale segmentation result by utilizing segmentation software capable of generating the multi-scale segmentation result, calculating an average spectrum of each segmented image object by fusing pixel-level sample information, classifying the average spectrum, and acquiring a classification result and a posterior probability vector of each object under different segmentation scales;

b, selecting an optimal scale based on entropy information;

respectively calculating the entropy of the posterior probability of each object according to the increasing sequence of the segmentation scale according to the multi-scale posterior probability vector obtained in the step A, wherein the calculation formula of the entropy is as follows:

<mrow> <mi>E</mi> <mo>=</mo> <mo>-</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>P</mi> <mi>i</mi> </msub> <mo>&amp;times;</mo> <msub> <mi>log</mi> <mn>2</mn> </msub> <mrow> <mo>(</mo> <msub> <mi>P</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow>

wherein P is_iRepresenting the probability that the object belongs to the ith class, wherein n represents the number of classes, selecting the segmentation scale with the minimum entropy value as the optimal segmentation scale of the object, and enabling the classification result of the object under the optimal segmentation scale to be the final class of the object;

wherein, for an object on a certain segmentation scale, the posterior probability vector is expressed as follows:

P＝(P₁，P₂...，P_i...，P_n)

wherein P is_iIndicating the probability that the object belongs to the ith class, and n represents the number of classes.

2. The method according to claim 1, wherein the specific implementation method of the step a is as follows:

firstly, segmenting a remote sensing image in multiple scales, wherein the multiple segmentation scales are determined according to the DN value or the reflectivity range of the image, and during segmentation, a shape factor and a compactness factor are set according to image characteristics;

secondly, creating a rule set in segmentation software according to the selected segmentation scale range, and segmenting step by step according to the sequence of the scale from large to small or from small to large;

then, forming a number file of the object of the multi-scale segmentation result, and sequentially exporting the number file to be a grid image, wherein the value corresponding to each pixel in the grid image is the number of the object;

then, respectively calculating the average spectrum of each object under each segmentation scale according to the obtained number file of the object after image segmentation and the original multiband image, and acquiring multi-scale average spectrum data;

finally, selecting training samples from the original multiband images to classify the obtained object average spectrum images under different scales respectively, wherein the selected training samples require typicality and randomness;

while classifying, obtaining posterior probability vectors of the objects under a plurality of scales, wherein the posterior probability vectors of one object on a certain segmentation scale are expressed as follows:

P＝(P₁，P₂...，P_i...，P_n)

wherein P is_iIndicating the probability that the object belongs to the ith class, and n represents the number of classes.

3. The method of claim 2, wherein the shape factor is set to 0.2 and the compactness factor is set to 0.5.

4. A method as claimed in claim 1, 2 or 3, characterized in that the segmentation scale is selected on the basis of an increase in scale, with a consequent increase in the scale interval.