Thermal infrared hyperspectral feature band selection method and system based on improved particle swarm algorithm
By improving the particle swarm optimization algorithm for feature band selection, the problem of unstable band selection in existing technologies is solved, and efficient and stable band selection is achieved under limited data conditions. This method is suitable for feature band selection in thermal infrared hyperspectral imagers.
Patent Information
- Application Number
- CN202311268859.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-27
- Publication Date
- 2025-12-30
- Estimated Expiration
- 2043-09-27
AI Technical Summary
Existing technologies make it difficult to effectively acquire the characteristic wavebands of materials in data acquisition equipment, resulting in serious information redundancy. Furthermore, in actual production, due to limitations such as time and cost, data acquisition equipment can often only acquire a limited waveband range, causing characteristic peaks to disappear in the spectrum.
An improved particle swarm optimization algorithm is used for feature band selection. Through an adaptive parameter adjustment mechanism and three evaluation indicators, local optima are prevented, global search capability is enhanced, and stability and accuracy of band selection are achieved.
It achieves efficient and stable characteristic band selection under limited data conditions, avoids complex manual parameter tuning, and improves the accuracy and stability of band selection.
Smart Images
Figure CN117253557B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of computer and information service, and particularly relates to a thermal infrared hyperspectral characteristic band selection method and system based on an improved particle swarm algorithm. BACKGROUND
[0002] The emissivity of a substance is largely a reflection of its internal structure, which provides a possibility for estimating the type and content of salt mines. In order to obtain the emissivity spectrum of a substance, a thermal infrared hyperspectral imager and a non-imaging instrument are usually used. However, the number of hyperspectral emissivity spectrum bands obtained in this way is too large, resulting in serious information redundancy. Such redundancy can cause some characteristic peaks to disappear in the spectrum, and in actual production, limited number of band intervals (usually 5-10) can be obtained by the data acquisition equipment due to the limitations of time and cost and other factors.
[0003] In order to solve this problem, it is necessary to mine the characteristic bands of a specific substance in the emissivity spectrum, so as to realize more accurate content analysis. At present, there are various band selection methods, which can be classified into different types, such as ranking-based methods, search-based methods, clustering-based methods, sparsity-based methods, and embedding learning-based methods.
[0004] However, the existing band selection methods have some limitations. On the one hand, a large amount of data may be needed to support the training of the model, which may not meet the data requirements in most actual situations. On the other hand, a large number of iterations may be needed for some methods, such as genetic algorithms. In addition, the results obtained by many methods may not be stable enough, and different output results may be obtained by multiple runs, so that manual parameter adjustment is needed to achieve acceptable accuracy. SUMMARY
[0005] The present application is to overcome the deficiencies of the prior art, and proposes an improved particle swarm algorithm for band selection. The method is based on a conventional particle swarm algorithm and proposes an improved particle swarm algorithm. The improved method introduces an adaptive parameter adjustment mechanism, which can adaptively adjust the parameters of each iteration according to the results of each iteration, and prevent falling into a local optimal solution.
[0006] Based on the above problems, the present application adopts the following technical solutions:
[0007] A thermal infrared hyperspectral characteristic band selection method based on an improved particle swarm algorithm, comprising the following steps:
[0008] Step 1, preparing a mixed salt ore sample, collecting hyperspectral emissivity spectrum of the mixed salt ore sample, measuring the mass percentage of each mixed salt mineral in the mixed salt ore sample as a data set for selecting characteristic waveband and constructing content prediction model;
[0009] Step 2, taking the data set in step 1 as input, establishing a prediction model of characteristic waveband combination of each substance in the mixed salt ore and corresponding content prediction based on the improved particle swarm algorithm;
[0010] Step 3, based on the prediction model in step 2, predicting the characteristic waveband combination of each substance and the corresponding content in the unknown mixed salt ore, collecting the mass percentage of the mixed salt ore and verifying the results.
[0011] Further, the data set acquisition in step 1, the preparation of the mixed salt ore sample and the measurement of the content are all completed in the laboratory, using the Hyper-Cam hyperspectral imager produced by Telops to shoot hyperspectral images of multiple groups of samples, and then using the iterative spectral smoothing temperature-emissivity separation algorithm ISSTES to obtain the emissivity spectrum of the sample.
[0012] Further, the respective content percentage of the mixed salt ore is obtained by the comprehensive analysis method of X-ray diffraction analysis, chemical titration and atomic absorption spectroscopy in step 1.
[0013] Further, the improved particle swarm algorithm is as follows:
[0014] Taking the data set as the input of the improved particle swarm algorithm, specifically, first, a certain number S of particles are randomly generated as the initial population, and a speed is randomly initialized for each particle, each particle represents a possible emissivity waveband combination, each value of each combination corresponds to a waveband, and the size of the waveband combination, i.e. the number of characteristic wavebands, is set in advance as an input parameter, and the initial positions of these particles are randomly distributed in the waveband combination space; for each particle, calculate its fitness value, use the inverse of the RMSE value of the result obtained by partial least squares method PLSR regression prediction as the fitness function, in each iteration, the individual can get its own fitness value according to the fitness function.
[0015] Further, in order to enable the individual to adaptively iterate, the following three indicators are set:
[0016] s1=|score current -score pbest |
[0017] s2=|score current -score gbest |
[0018] s3 = |score gbest -score pbest |
[0019] where s1 measures the gap between the current individual and its best result, s2 measures the gap between the current individual and the best result of the group, s3 represents the gap between the best result of the individual and the best result of the group, score is the fitness, and the larger the score means the better the current individual's band combination, and any particle can obtain its current score at any time current , its best score score pbest , and the best score score of the group so far gbest .
[0020] Further,
[0021]
[0022]
[0023] where are the three parameters of the i-th particle in the k-th iteration, where is the inertia weight, are the learning factors of the balanced local search and global search in the k-th iteration, respectively, and σ is a constant value set max and s min The maximum fitness and the minimum fitness obtained when calculating the fitness of all individuals after initialization.
[0024] Further, the indicators s1, s2 and s3 are used to guide the parameter update of each individual, according to the parameter values obtained above, and are brought into the following update equation:
[0025]
[0026]
[0027] where, and are the band labels and updated values of the i-th particle in the group in d dimensions, is the velocity of the i-th particle in the d-dimensional band, and the iteration number is k, and are the individual historical optimal solution and the group optimal solution, respectively, so as to obtain a new round of population, after N iterations, the band combination with the largest score is output, that is, the optimal characteristic band combination.
[0028] In another aspect, the present application also provides a system for selecting characteristic wavebands of thermal infrared hyperspectral based on an improved particle swarm algorithm, comprising:
[0029] A data set collection module, which prepares a mixed salt ore sample, collects the hyperspectral emissivity spectrum of the mixed salt ore sample, measures the mass content percentage of each mixed salt mineral in the mixed salt ore sample as a data set for selecting characteristic wavebands and predicting content;
[0030] A prediction model construction module, which takes the data set in step 1 as input and establishes a prediction model for the characteristic waveband combination of each substance in the mixed salt ore and the corresponding content prediction based on the improved particle swarm algorithm;
[0031] A prediction module, which predicts the characteristic waveband combination of each substance in the unknown mixed salt ore and the corresponding content based on the prediction model in the prediction model construction module, collects the mass content percentage of the mixed salt ore and verifies the results.
[0032] Compared with the prior art, the present application has the following beneficial effects:
[0033] (a) The algorithm can adjust parameters through an adaptive strategy, avoiding complex manual parameter adjustment;
[0034] (b) The three new evaluation indexes set by the algorithm can prevent the algorithm from falling into a local optimal solution, enhancing the global search ability;
[0035] (c) It provides an excellent solution for the characteristic waveband selection task of thermal infrared hyperspectral. BRIEF DESCRIPTION OF DRAWINGS
[0036] Figure 1 The workflow of the improved particle swarm algorithm in the present application.
[0037] Figure 2 The data collection and processing process in the present application.
[0038] Figure 3 The hyperspectral emissivity spectrum of part of the samples in the present application.
[0039] Figure 4 The prediction results obtained by the improved particle swarm algorithm in the present application after selecting different lengths of characteristic waveband combinations.
[0040] Figure 5 The comparison results of the improved particle swarm algorithm in the present application and other waveband selection algorithms in the stability of prediction results. DETAILED DESCRIPTION
[0041] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by the person of ordinary skill in the art without creative labor fall within the protection scope of the present application.
[0042] Embodiment 1
[0043] Step 1, collect the hyperspectral emissivity spectrum of the mixed salt mine sample, and obtain the respective content percentages of the mixed salt mine substances (for example, the mixed potassium salt mine contains langbeinite, potassium chloride, magnesium sulfate, sodium chloride) through the comprehensive analysis method of X-ray diffraction analysis (XRD) and chemical titration and atomic absorption spectroscopy (ASS), as the data set for selecting the characteristic waveband and performing content prediction, that is, the hyperspectral emissivity spectrum of multiple groups of mixed salt mine samples and the respective content percentages of the substances contained therein;
[0044] Step 2, use the data set in step 1 as the input of the improved particle swarm algorithm, set the number of characteristic wavebands to be obtained, the population size S and the maximum number of iterations N, without setting other parameters, and output the characteristic waveband combination of each substance in the mixed salt mine and the corresponding content prediction result after N rounds of iteration processing of the algorithm.
[0045] Further, the preparation of the mixed salt mine sample in step 1 and the measurement of the content are both completed in the laboratory, a Hyper-Cam hyperspectral imager produced by Telops is used to shoot the hyperspectral images of multiple groups of samples, and then an iterative spectral smoothing temperature-emissivity separation algorithm ISSTES is used to obtain the emissivity spectrum of the sample, and this process is shown in Figure 2 ; Figure 3 The emissivity spectrum of part of the samples is shown.
[0046] Further, in step 2, the data set is used as the input of the improved particle swarm algorithm, and the algorithm flow chart is shown in Figure 1 . Specifically, first, the algorithm randomly generates a certain number S of particles as the initial population, and randomly initializes a speed for each particle, each particle represents a possible emissivity waveband combination, and each value of each combination corresponds to a waveband. The size of the waveband combination, that is, the number of characteristic wavebands, is set in advance as an input parameter. The initial positions of these particles are randomly distributed in the waveband combination space; for each particle, calculate its fitness value, and in this algorithm, the reciprocal of the RMSE value of the result obtained by using the partial least squares method PLSR regression prediction and the true content in step 1 is used as the fitness function. In each iteration, the individual can get its own fitness value according to the fitness function. In order to enable the individual to adaptively iterate, first, define three indicators:
[0047] s1 = |score current -score pbest |
[0048] s2 = |score current -score gbest |
[0049] s3 = |score gbest -score pbest |
[0050] where score is the fitness, the bigger the score means the better the band combination of the current individual, any particle can get its current score score at any time current , its best score score in the past pbest , and the best score score of the population so far gbest . For the three indicators, s1 measures the gap between the current individual and its own best result, s2 measures the gap between the current individual and the best result of the population, and s3 represents the gap between the best result of the individual and the best result of the population. The three indicators can be used to guide the parameter update of each individual, as follows:
[0051]
[0052]
[0053] where and are the three parameters of the jth particle in the kth iteration, where is the inertia weight; σ is a very small number, which is set to 0.001 here; s max and s min are obtained when the fitness of all individuals is calculated after initialization. According to the parameter values obtained above, the following update equation is brought in:
[0054]
[0055]
[0056] where, and are the band labels and updated values of the ith particle in the population in d dimensions, respectively, and the number of iterations is k. and respectively. Thus, a new population is obtained. After N iterations, the band combination with the maximum score is output, which is the optimal feature band combination. By setting the length of the band combination, such as 2-20, the prediction results corresponding to the selected band combinations of different lengths can be obtained, as shown in Figure 4 .
[0057] To verify the stability of the algorithm proposed in this paper, that is, the prediction results remain small fluctuations under the same parameter settings. For example, for the test of mixed potassium salt, by uniformly setting the band combination length to 10, the number of iterations to 100, and the population size to 200, multiple independent repeated experiments are performed to predict the percentage of soft potassium magnesium sulfate. The standard deviation of the prediction results is calculated, and the comparison results of the improved particle swarm algorithm and other band selection algorithms in the stability of the prediction results are shown in Figure 5 .
[0058] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or replace them with similar ways, without departing from the spirit of the present application or exceeding the scope defined by the appended claims.
[0059] Example 2
[0060] A feature band selection system for thermal infrared hyperspectral based on an improved particle swarm algorithm is provided, comprising:
[0061] A data set collection module prepares a mixed salt ore sample, collects the hyperspectral emissivity spectrum of the mixed salt ore sample, measures the mass percentage of each mixed salt mineral in the mixed salt ore sample, and uses it as a data set for selecting feature bands and predicting content.
[0062] A prediction model construction module uses the data set in step 1 as input and establishes a prediction model for the feature band combination of each substance in the mixed salt ore and the corresponding content prediction based on the improved particle swarm algorithm.
[0063] A prediction module collects a mixed salt ore with unknown mixed salt mineral content percentage, and predicts the feature band combination of each substance in the mixed salt ore and the corresponding content based on the prediction model in step 2.
[0064] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.
[0065] The other parts not described in detail are prior art.
Claims
1. A thermal infrared hyperspectral characteristic band selection method based on an improved particle swarm algorithm, characterized in that, The method comprises the following steps: Step 1, preparing a mixed salt ore sample, collecting a hyperspectral emissivity spectrum of the mixed salt ore sample, and measuring the mass content percentage of each mixed salt mineral in the mixed salt ore sample as a data set for selecting characteristic wave bands and performing content prediction modeling; Step 2, taking the data set in step 1 as input, and establishing a prediction model of the characteristic wave band combination of each substance in the mixed salt ore and the corresponding content prediction based on an improved particle swarm algorithm; the improved particle swarm algorithm is specifically as follows: Taking the data set as the input of the improved particle swarm algorithm, specifically, first, a certain number S of particles are randomly generated as an initial population, and a speed is randomly initialized for each particle, each particle represents a possible emissivity wave band combination, and each value of each combination corresponds to a wave band; the size of the wave band combination, that is, the number of characteristic wave bands, is set in advance as an input parameter, and the initial positions of these particles are randomly distributed in the wave band combination space; for each particle, the fitness value thereof is calculated, and the reciprocal of the RMSE value of the result obtained by using a partial least squares method PLSR regression prediction and the true content in step 1 is taken as the fitness function; in each iteration, the individual can obtain its own fitness value according to the fitness function; In order to enable the individual to adaptively iterate, the following three indexes are set: wherein, measures the gap between the current individual and the individual's best result, measures the gap between the current individual and the group's best result, represents the gap between the individual's best result and the group's best result, score is the fitness, score the greater the better the current individual's band combination, and any particle can obtain its current score at any time its best score ever and the group's best score so far ; Step 3, based on the prediction model in step 2, the characteristic wave band combination of each substance in the unknown mixed salt ore and the corresponding content are predicted, the mass content percentage of the mixed salt ore is collected, and the result is verified.
2. The method according to claim 1, wherein the method is characterized in that, The data set acquisition in step 1, the preparation of the mixed salt ore sample and the measurement of the content are all completed in the laboratory; a Hyper-Cam hyperspectral imager produced by Telops is used to shoot hyperspectral images of multiple samples, and then an iterative spectral smoothing temperature-emissivity separation algorithm ISSTES is used to obtain the emissivity spectrum of the sample.
3. The method according to claim 1, wherein the method is characterized in that, The content percentage of each mixed salt mineral is obtained by a comprehensive analysis method of X-ray diffraction analysis, chemical titration and atomic absorption spectroscopy.
4. The thermal infrared hyperspectral characteristic wave band selection method based on the improved particle swarm algorithm according to claim 1, characterized in that, in , , yes k In the nth iteration i The three parameters of a particle, among which It is inertial weight. , They are k The learning factor that balances local and global search in each iteration. This is a constant value that is set. and The maximum and minimum fitness values are obtained when calculating the fitness of all individuals after initialization.
5. The method according to claim 4, wherein the method is characterized in that, Indicators , , For guiding the parameter update of each individual, according to the parameter values obtained above, the following update equation is brought in: wherein, and are the wavelength band label and update value of the i th particle in the group in the d th dimension, is the velocity of the i th particle in the wavelength band in the d th dimension, and the iteration number is k , and are the individual historical optimal solution and the group optimal solution, respectively, so as to obtain a new round of population. After N iterations, the score largest wavelength band combination, i.e. the optimal characteristic wavelength band combination, is output.
6. A system for selecting characteristic wavebands for thermal infrared hyperspectral based on an improved particle swarm optimization algorithm, characterized in that, The method comprises: a data set collection module, which prepares a mixed salt ore sample, collects a hyperspectral emissivity spectrum of the mixed salt ore sample, and measures the mass content percentage of each mixed salt mineral in the mixed salt ore sample as a data set for selecting characteristic wave bands and performing content prediction; a prediction model construction module, which takes the data set in step 1 as input, and establishes a prediction model of the characteristic wave band combination of each substance in the mixed salt ore and the corresponding content prediction based on an improved particle swarm algorithm; the improved particle swarm algorithm is specifically as follows: The dataset is inputted into the improved particle swarm algorithm. Specifically, firstly, a certain number S of particles are randomly generated as an initial population, and a speed is randomly initialized for each particle, each particle representing a possible emissivity band combination, each value of each combination corresponding to a band, the size of the band combination, i.e., the number of characteristic bands, being set in advance as an input parameter, and the initial positions of the particles are randomly distributed in the band combination space; for each particle, the fitness value thereof is calculated, the reciprocal of the RMSE value of the result obtained by using the partial least squares method PLSR regression prediction is taken as the fitness function, and in each iteration, the individual can obtain its own fitness value according to the fitness function; In order to enable the individual to adaptively iterate, the following three indexes are set: wherein, measures the gap between the current individual and the individual's best result, measures the gap between the current individual and the group's best result, represents the gap between the individual's best result and the group's best result, score is the fitness, score the greater the better the current individual's band combination, and any particle can obtain its current score at any time its best score ever and the group's best score so far ; A prediction module is configured to predict the characteristic band combination of each substance in the unknown mixed salt mine and the corresponding content based on the prediction model in the prediction model construction module, collect the substance content percentage of the mixed salt mine, and verify the result.