CN106650790A - Remote sensing image cluster method based on swarm intelligence - Google Patents

Abstract

The invention relates to a remote sensing image cluster method based on swarm intelligence and belongs to the remote sensing image cluster analysis field. The remote sensing image cluster method based on swarm intelligence is provided to overcome disadvantages that a conventional sorting algorithm for remote sensing images is poor in development capability, and cannot obtain a global optimal cluster center or obtain a satisfying remote sensing sorting result when developing an optimal cluster center to realize remote sensing image sorting tasks. The method includes the following steps: determining the sorting number of remote sensing images to be sorted, and randomly distributing each pixel of the remote sensing images to one classification; carrying out swarm intelligence development for the remote sensing images, and ending the development when a cluster index reaches a preset standard; sorting the remote sensing images according to a result of the swarm intelligence development; and if a food source still cannot increase the value of pollen abundance after the preset number of times, utilizing Levy flight to globally search for a new food source in a solution space. The remote sensing image cluster method based on swarm intelligence is suitable for sorting remote sensing images.

Description

Remote sensing image clustering method based on swarm intelligence

Technical Field

The invention relates to a remote sensing image clustering method based on swarm intelligence, and belongs to the field of remote sensing image clustering analysis.

Background

Land utilization/mulch provides basic data for many disciplines, including ecology, geography, and climate. Thus, it has been one of the focuses of scientists. Remote sensing technology has been considered as one of the mainstream means of acquiring land use/cover data because of its many unique advantages, including macroscopicity, presence, repeatability, and economy. Scientists have paid much effort to develop many remote sensing classification algorithms, but accurate remote sensing image classification remains a significant challenge due to the complexity of remote sensing images.

In general, classification methods are classified into supervised classification and unsupervised classification (clustering). Supervised classification (e.g., maximum likelihood) generally has better classification accuracy, but requires a large number of training samples to guide the classification. The collection of training samples is generally time consuming and labor intensive, and is not accessible in many areas. With the rapid development of remote sensing technology, satellite remote sensing data volume acquired by human beings presents a massive scale. In this context, unsupervised classification is of increasing interest and attention by scientists because it does not require sample knowledge and relies only on the statistical properties of the remote sensing images themselves to produce land use data. At present, the unsupervised classification method is widely applied to various remote sensing applications, including global land utilization mapping and the like.

The k-means algorithm is the most commonly used unsupervised remote sensing classification algorithm. The working principle is simple, the efficiency is high, and therefore the method is widely applied. However, k-means is established on the basis that the data object conforms to the Gaussian distribution, and the remote sensing data is quite complex, and the data distribution of the remote sensing data does not conform to the Gaussian distribution. Thus, k-means often cannot reach a globally optimal solution. In addition, the performance thereof is greatly affected by the initial data, and thus the stability is insufficient.

The rapidly developing artificial intelligence provides a new opportunity for the progress in this field. The unsupervised classification problem can be converted into an optimization problem and solved by using an artificial intelligence method. Genetic Algorithm (GA) is the most classical and most commonly used intelligent algorithm, and genetic clustering (GA-clustering) is developed by GA students for remote sensing image classification. The algorithm comprises the following steps:

1. regarding each solution as a chromosome, and realizing optimization through the operation on the chromosome;

2. constructing a fitness function, reserving elite groups through selecting operators, and eliminating part of laggard groups;

3. increasing the population diversity by using a crossover operator and a mutation operator;

4. and (5) outputting an optimal solution when a cycle stop condition is reached, and realizing the non-supervised classification of the remote sensing images.

However, the above genetic clustering algorithm has the following disadvantages:

1 weak development ability: the genetic clustering algorithm develops the existing chromosomes by means of selection and crossover operators, but does not pay particular attention to particularly excellent chromosomes, so that the convergence rate is low, and the genetic clustering algorithm is not suitable for processing complex and large-scale remote sensing data.

2, poor exploitation capability: the genetic clustering algorithm only depends on intersection and variation to realize population development, and the global search and development of solution space are often difficult to realize; due to the weak development capability, the quality of the remote sensing image clustering result is very dependent on the initial population quality, namely the clustering result is not stable enough.

In summary, the existing intelligent algorithm is weak in development and development capability, and in the task of mining the optimal clustering center to realize remote sensing image classification, the global optimal clustering center cannot be obtained frequently, and a satisfactory remote sensing classification result cannot be obtained.

Disclosure of Invention

The invention aims to solve the defects that the existing classification algorithm for remote sensing images is weak in development and development capability, the global optimal clustering center is often unavailable in the task of mining the optimal clustering center to realize the classification of the remote sensing images, and a satisfactory remote sensing classification result cannot be obtained, and provides a remote sensing image clustering method based on swarm intelligence, which comprises the following steps:

step 1) determining the classification number of remote sensing images to be classified, and randomly allocating each pixel of the remote sensing images to one classification; each picture element having a predetermined number of features;

step 2) carrying out intelligent bee colony excavation on the pixels according to the characteristics of the pixels, and specifically comprising the following steps:

step 2.1) initializing control parameters; the control parameters comprise the number of bees, the maximum cycle number and the limited search number; bees include hiring bees, observing bees; the number of the bees is p, the number of the employed bees is p/2, and the number of the observation bees is p/2;

step 2.2) establishing a food source; each food source corresponds to a employment bee; the food source is formed by connecting the clustering centers of each category, and the length of the food source is n multiplied by m, wherein n is the wave band number of each pixel in the remote sensing image, and m is the number to be classified; the first n bits of the food source represent the cluster center of the first category, and so on;

step 2.3) calculating the suitability of the food source; the expression of the fitness function f is 1/(M +1), and M is a clustering index;

step 2.4) searching a new food source; after the suitability of the existing food source is calculated according to the step 2.3), a new food source position is randomly searched around the existing food source;

step 2.5) observing bees according to the random probability P (X)_i) Follow-up of a food source with random probability P (X)_i) The expression of (a) is:

wherein, X_iIs the ith bee food source position, f (X)_i) Is a food source X_iFlower ofPowder abundance, N_eNumber of bees employed;

step 2.6) if one food source can not improve f (X) all the time after the limited search times_i) If the value of (a) is less than the predetermined value, the hiring bee is changed into a scout bee, and a new food source is globally searched in the solution space by using column-dimensional flight; if f (X) can be increased_i) If so, jumping to step 2.4);

step 2.7) when all bees finish searching, comparing the current most suitable food source with the optimal food source of the previous cycle, and selecting the food source with a higher numerical value as the current global optimal food source; and when the circulation reaches the maximum circulation times, stopping circulation and outputting the optimal clustering center.

The invention has the beneficial effects that: 1. the genetic clustering algorithm in the prior art only increases the proportion of the elite population by selecting an operator, but does not perform a strong search around the population, so the development capability of the genetic clustering algorithm is weak. In the invention, in the process of searching the optimal clustering center of the image, the observation bees are used for searching the periphery of the excellent food source for multiple times, so that the population development capability is greatly improved. 2. The genetic clustering algorithm realizes global search through the crossover operator and the mutation operator, however, the crossover operator is only based on the existing gene segments, and the mutation operator is usually only directed at one gene position and has limited mutation amplitude, so the exploitation capability of the genetic clustering algorithm is limited. The genetic clustering algorithm limits the search times by definition, gives up poor food sources in time, and realizes the global roaming capability through column dimension flight, thereby realizing the global search capability in the optimal clustering center solution space and having strong development capability. 3. The employment bees and the observation bees are mutually matched, so that the rapid search of the remote sensing image clustering optimal food source is realized, and the development capability is enhanced; 4. the method has the outstanding advantages that the efficiency of searching the honey source can be improved by the characteristic that the small probability flows to a position, namely the original searching mode conforms to mean distribution and normal distribution, so that the searching range is small, if no honey source conforming to the conditions exists in adjacent areas, the searching efficiency can be greatly reduced, the step length of the column-dimensional flight does not conform to the Gaussian distribution law, the searching efficiency can be improved, the technical bias of technicians in the field is overcome by using the column-dimensional flight model, the technicians in the field only recognize that the global searching and the global roaming can be carried out by using a random algorithm, but the used algorithms are all random algorithms conforming to the Gaussian distribution law, even with increased efficiency, this is not considered from the point of view of non-gaussian distribution algorithms. Therefore, the remote sensing image clustering method based on swarm intelligence is unobvious and has a prominent technical effect.

Drawings

FIG. 1 is a flow chart of the remote sensing image clustering method based on swarm intelligence of the present invention;

fig. 2 is a flowchart of specific steps of intelligent mining of a bee colony in the remote sensing image clustering method based on bee colony intelligence of the present invention.

Detailed Description

The first embodiment is as follows: the remote sensing image clustering method based on the swarm intelligence is characterized by comprising the following steps of:

step 1) determining the classification number of remote sensing images to be classified, and randomly allocating each pixel of the remote sensing images to one classification; each picture element having a predetermined number of features;

step 2) carrying out intelligent bee colony excavation on the pixels according to the characteristics of the pixels, and specifically comprising the following steps:

step 2.1) initializing control parameters; the control parameters comprise the number of bees, the maximum cycle number and the limited search number; bees include hiring bees, observing bees; the number of bees is p, the number of hired bees is p/2, and the number of observed bees is p/2;

step 2.2) establishing a food source; each food source corresponds to a employment bee; the food source is formed by connecting the clustering centers of each category, and the length of the food source is n multiplied by m, wherein n is the wave band number of each pixel in the remote sensing image, and m is the number to be classified; the first n bits of the food source represent the cluster center of the first category, and so on;

step 2.3) calculating the suitability of the food source; the expression of the fitness function f is 1/(M +1), and M is a clustering index; the clustering index M directly reflects the quality of a clustering center, so in the invention, the income of the food source is estimated based on the clustering index M, and the suitability function is made in order to ensure that rich food sources have high income, considering that the larger the clustering index is, the poorer the quality of the food source is.

Step 2.4) searching a new food source; after the suitability of the existing food source is calculated according to the step 2.3), a new food source position is randomly searched around the existing food source;

step 2.5) observing bees according to the random probability P (X)_i) Follow-up of a food source with random probability P (X)_i) The expression of (a) is:

wherein, X_iIs the ith bee food source position, f (X)_i) Is a food source X_iAbundance of pollen, N_eNumber of bees employed; this means that when starting the search, the observing bees randomly become the hiring bees to search the food sources, and the probability of randomly becoming the hiring bees is P (X)_i)，P(X_i) To some extent reflecting the abundance of the food source, the higher the chance of hiring a bee to follow the food source. Therefore, more attention can be paid to abundant foodA source of matter.

Step 2.6) if one food source can not improve f (X) all the time after the limited search times_i) If the value of (a) is less than the predetermined value, the hiring bee is changed into a scout bee, and a new food source is globally searched in the solution space by using column-dimensional flight; if f (X) can be increased_i) If so, jumping to step 2.4);

step 2.7) when all bees finish searching, comparing the current most suitable food source with the optimal food source of the previous cycle, and selecting the food source with a higher numerical value as the current global optimal food source; and when the circulation reaches the maximum circulation times, stopping circulation and outputting an optimal clustering center, thereby completing the task of remote sensing image clustering.

In which the column-dimensional flight is a motion pattern in which most of the cases flow in a small range and a small part of the cases flow to a distant position from the viewpoint of distribution.

The advantage of using the column-dimensional flight is that when a new food source is searched, the search position is communicated in a small range, and the small probability can flow to a far position, so that the cluster of the remote sensing images can be searched globally.

Because the traditional swarm intelligent algorithm cannot perform global search, compared with the traditional algorithm, the method has the advantages that the new honey source can be searched at a random position in a small range by the random flow of the column dimension in the small range; compared with other bee colony algorithms which perform global search, the method has the advantages that the column dimension step size provides a method for generating the long step size with a small probability, and the search efficiency of the honey source is higher.

In step 2.5), f (X) is_i) The same function as the fitness function f in step 2.3), the fitness function f being a function on M, while the expression from MIt can be seen that the value accumulated by twice accumulating the symbols is exactly the position of the ith bee food source. The fitness function f is therefore not directly reflected in the representation in relation to X_iBut in practice the argument of the fitness function f is X_i。

The second embodiment is as follows: the first difference between the present embodiment and the specific embodiment is: the clustering index is calculated by the following formula:

wherein x is_jZ is any one of the image elements in the i (i ═ 1,2, …, k) category_iAs the cluster center of class i, C_iFor the i-th cluster, j is the number of pixels in the category i, k ∈ {1,2, …, p/2} and k ≠ i.

Other steps and parameters are the same as those in the first embodiment.

The third concrete implementation mode: the present embodiment is different from the first to the second embodiments in that:

in step 2.2), the position of the food source is determined by the following formula:

wherein,indicating the location of the ith employment bee at the jth feature,andrespectively, the minimum and maximum values of the jth feature, and rand (0,1) represents a random number varying between 0-1. The jth feature refers to the jth feature in the multi-dimensional features (namely wave bands) of the remote sensing image

Other steps and parameters are the same as those in one of the first to third embodiments.

The fourth concrete implementation mode: the difference between this embodiment mode and one of the first to third embodiment modes is:

the expression of the new food source in step 2.2) is as follows:

wherein,is the new food source location for the ith employed bee at the jth feature (j ═ 1,2, …, n);andrespectively representing the original food source positions of i-th and k-th bees in the j-th characteristic, wherein i, k ∈ {1,2, …, p/2} and k is not equal to i;is a random number that varies between-1 and 1.

This embodiment shows that after the existing food source revenue assessment is completed, the hiring bee starts to randomly search for new food source locations around the existing food sources, and the search mechanism is as shown in the above equation, and if the new food source is more profitable than the original food source, the original food source is replaced by the new food source.

Other steps and parameters are the same as in one of the first to fourth embodiments.

The fifth concrete implementation mode: the difference between this embodiment and one of the first to fourth embodiments is:

the new position in step 2.6) is calculated by the following formula:

wherein,indicating a new location, X, of the food source i that was discarded_iIn the home position, s is the step size generated by the column-dimensional flight.

Other steps and parameters are the same as those in one of the first to fifth embodiments.

The sixth specific implementation mode: the difference between this embodiment and one of the first to fifth embodiments is:

the step length s generated by the column-dimensional flight is calculated by the following formula:

here, μ, ν, and λ are calculated from normal distributions, respectively, that is:

wherein,

where β is a constant varying between 1 and 2, a gamma function.

Other steps and parameters are the same as those in one of the first to sixth embodiments.

Claims (6)

1. A remote sensing image clustering method based on swarm intelligence is characterized by comprising the following steps:

step 1) determining the classification number of remote sensing images to be classified, and randomly allocating each pixel of the remote sensing images to one classification; each picture element having a predetermined number of features;

step 2) carrying out intelligent bee colony excavation on the pixels according to the characteristics of the pixels, and specifically comprising the following steps:

step 2.1) initializing control parameters; the control parameters comprise the number of bees, the maximum cycle number and the limited search number; bees include hiring bees, observing bees; the number of the bees is p, the number of the employed bees is p/2, and the number of the observation bees is p/2;

step 2.2) establishing a food source; each food source corresponds to a employment bee; the food source is formed by connecting the clustering centers of each category, and the length of the food source is n multiplied by m, wherein n is the wave band number of each pixel in the remote sensing image, and m is the number to be classified; the first n bits of the food source represent the cluster center of the first category, and so on;

step 2.3) calculating the suitability of the food source; the expression of the fitness function f is 1/(M +1), and M is a clustering index;

step 2.4) searching a new food source; after the suitability of the existing food source is calculated according to the step 2.3), a new food source position is randomly searched around the existing food source;

step 2.5) observing bees according to the random probability P (X)_i) Follow-up of a food source with random probability P (X)_i) The expression of (a) is:

$P\left(X_i\right)=\frac{f\left(X_i\right)}{\underset{i=1}{\overset{N_e}{\mathrm{\Σ}}}f\left(X_i\right)}$

wherein, X_iIs the ith bee food source position, f (X)_i) Is a food source X_iAbundance of pollen, N_eNumber of bees employed;

step 2.6) if one food source can not improve f (X) all the time after the limited search times_i) If the value of (a) is less than the predetermined value, the hiring bee is changed into a scout bee, and a new food source is globally searched in the solution space by using column-dimensional flight; if f (X) can be increased_i) If so, jumping to step 2.4);

step 2.7) when all bees finish searching, comparing the current most suitable food source with the optimal food source of the previous cycle, and selecting the food source with a higher numerical value as the current global optimal food source; and when the circulation reaches the maximum circulation times, stopping circulation and outputting the optimal clustering center.

2. The method according to claim 1, wherein in step 2.3), the clustering index is calculated by the following formula:

$M=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{x_j\∈C_i}{\mathrm{\Σ}}\left|\right|x_j-z_i\left|\right|$

wherein x is_jC, an image element of any one of the remote sensing images in the i (i ═ 1,2, …, k) category_iFor class i clustering, z_iIs the cluster center of the category i, j is the number of pixels of the category i, k ∈ {1,2, …, p/2} and k ≠ i.

3. The method according to claim 1, wherein in step 2.2), the location of the ith food source is determined by the following formula:

$X_i^j=X_{min}^j+rand(0,1)(X_{max}^j-X_{min}^j)$

wherein,indicating the location of the ith employment bee at the jth feature,andrespectively, the minimum and maximum values of the jth feature, and rand (0,1) represents a random number varying between 0-1.

4. The method of claim 3, wherein the new food source expression in step 2.4) is:

$V_i^j=X_i^j+{\mathrm{\θ}}_i^j(X_i^j-X_k^j)$

wherein, V_i ^jIs the new food source location for the ith employed bee at the jth feature (j ═ 1,2, …, n);andrespectively representing the original food source positions of i-th and k-th bees in the j-th characteristic, wherein i, k ∈ {1,2, …, p/2} and k is not equal to i;is a random number that varies between-1 and 1.

5. Method according to claim 4, characterized in that the new position in step 2.6) is calculated by the following formula:

$X_i^{new}=X_i+s$

wherein,a new location representing food source i that was abandoned; x_iIs the original position; s is the step size generated by the column-dimensional flight, and s is a random value that does not conform to the Gaussian distribution law.

6. The method of claim 5, wherein the step size s generated by the column-dimensional flight is calculated by the following formula:

$s=0.01\frac{\mathrm{\μ}}{{\left|\mathrm{\ν}\right|}^{1/\mathrm{\β}}}\mathrm{\λ}$

mu, ν and λ are calculated from the normal distribution respectively, i.e.:

$\mathrm{\μ}~N(0,{\mathrm{\σ}}_{\mathrm{\μ}}^2),\mathrm{\ν}~N(0,{\mathrm{\σ}}_{\mathrm{\ν}}^2),$ λ～N(0,1),

wherein,

${\mathrm{\σ}}_{\mathrm{\μ}}={\[\frac{\mathrm{\Γ}(1+\mathrm{\β})sin\left(\mathrm{\π}\frac{\mathrm{\β}}{2}\right)}{\mathrm{\Γ}\left(\frac{(1+\mathrm{\β})}{2}\right)\mathrm{\β}{2}^{\frac{(\mathrm{\β}-1)}{2}}}\]}^{\frac{1}{\mathrm{\β}}},{\mathrm{\σ}}_v=1.$

where β is a constant varying between 1 and 2, a gamma function.