Study on Characteristic Spectrum and Multiple Classifier Fusion With Different Particle Size in Marine Sediments

Marine sediments record much information of the ecological process, which highly correlated with global and local environmental change. Particle size and its distributions of sediments indicate different ecological functions, thus are the key questions in marine ecology. The analysis method is tedious and laborious, which is not conducive for in-situ monitoring. Here, a spectral analysis was explored using surface sediments sampled in the intertidal zone of Dongdayang village, Qingdao, China. These samples were dried and sieved to pass through the mesh size of 0.3 mm, 0.2 mm, 0.1 mm, and 0.075 mm, respectively. Then, four types of subsamples were collected with the particle size of 0.3-0.2 mm, 0.2-0.1 mm, 0.1-0.075 mm, and < 0.075 mm, respectively. The visible and near infrared reflectance spectra (226-975nm) of these subsamples with different particle size were measured. Results showed that there was a negative correlation between the spectral reflectance and the particle size. And the characteristic spectra for particle size classification were 926-975nm. These particle size were classified by the support vector machine algorithm. The classification accuracy for the calibration set and validation set was 100% and 89.06%, respectively. Furthermore, the fusion classifier was compared with the single classifier. Three spectral bands were selected as the single particle size classifier, that is 226-325nm, 826-925nm and 226-975nm. These three single classifiers were fused by voting method, forming multiple classifier fusion. A fusion classifier was recommended whose validation set had the classification accuracy of 93.75%, better than any single classifier. Multiple classifier fusion is a good tool for searching the characteristic spectra of the chemical and physical parameter in sediments. This method provides a solution for the division of particle size in sediment.


I. INTRODUCTION
Marine sediments are the medium and carrier of interaction among hydrosphere, biosphere and lithosphere [1]. They record much information of the ecological process, which highly correlated with global and local environmental change [2], [3]. Particle size is an important structural feature of sediments and the basis of its classification, indicating the source and environment of sediments. It is the most widely used parameter to study the characteristics and environment of sediments [4]. And its distribution in sediments reveals The associate editor coordinating the review of this manuscript and approving it for publication was Alessandro Pozzebon. migration and settlement information of sediments [5]. Particle size analysis is widely used in sediment research, which is to determine the particle size and distribution of sediment components. The granularity analysis of marine sediments is not only the basis of dividing the types of seabed sediments and compiling the type map of seabed sediments, but also the material source and sedimentation of seabed sediments [6], [7]. Therefore, the analysis of different particle sizes of marine sediments is of great significance to the study of sediments. Traditional techniques of the particle size measurement generally include direct measurement, sieve analysis, binocular microscopy, sedimentation method and electron VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ microscopy [8], [9]. The most commonly used methods are sieve analysis and sedimentation. The sieve analysis method is to sieve the samples from coarse to fine through different sieve sizes [10]. The basic principle of sedimentation method is to use particle sedimentation velocity to divide particle distribution [11]. With the development of analytical technology, laser particle analyzer has been developed and widely used [12], [13]. The analysis speed and measurement range have been improved greatly, but it still needs tedious and laborious pretreatment. Furthermore, results from laser particle analyzer are often different from the traditional sedimentation method and sieve analysis method. Cuven et al. [14] used laser particle size analysis to study the relationship between chemical profile and particle size of sediments at macro and micro scales. Menéndez-Aguado et al. [15] carried out power spectrum analysis using laser diffraction spectrum to accurately represent the mathematical model of sediment particle size distribution. Image method is also one of the current particle size analysis methods, which can express the size of each particle and contain particle shape information. However, the data processing is complex and their interpretation is difficult [16]. Yates et al. [17] used the landsat5 TM satellite data to classify the surface sediments.
Results showed a more accurate classification for the muddy areas. Kim et al. [18] studied the classification of sediments in xx sites using the UAV Orthophoto Image combined with the field survey (232 samples). Results showed that the overall accuracy was 72.8%. Visible near infrared spectroscopy (VNIRS) is a fast, simple, nondestructive and green analysis technology [19], [20], thus has been widely used in many fields [21], [22]. The absorbance and reflectance spectra contain a lot of material information, which mainly reflects the types and quantities of O-H, N-H, C-H and other organic functional groups. Varied achievements have been made in the rapid determining many chemical and physical parameters in soils and sediments by VNIRS [23], [24]. Sediment size classification was also explored by VNIRS in limited studies but needs a further research [25].
In this paper, VNIRS was applied to analyze the particle size in the surface sediments of the intertidal zone in Qingdao, China. Two questions would be resolved. Firstly, how would the reflectance spectra of the sediments vary with different particle size? We hypothesized that the reflectance of the sediments increased with the decreased particle size because smaller size particle could produce larger reflectance. Secondly, how could the particle size of the sediments be expressed by the spectra? As other chemical or physical parameters, the particle size of sediments could be modelled by stoichiometry, such as multiple classifier fusion combined the support vector machine algorithm. Here, through the introduction of spectrum technology, we can find the characteristic wavelength to characterize the particle size characteristics, and establish the classification model of different particle sizes, so as to achieve the rapid division among different particle sizes of sediment. It provides a new method for rapid particle size analysis of sediments.
The outline of the paper is as follows. The model method and multiple classifiers fusion explained in Section II. Section III is the acquisition of sediments samples and data. Section IV and V presents and discusses the experimental studies that demonstrate the performances of the method. Finally, conclusion is given in Section VI.

II. THEORETICAL ANALYSIS A. MODELING METHOD
In this paper, support vector machine algorithm (SVM) was used as the modeling method used. The core idea of SVM is to construct an optimal classification hyperplane by mapping vectors into higher dimensional space through kernel function [26], [27].Two parallel hyperplanes are found on both sides of the hyperplane, which maximizes the distance between two hyperplanes. By constructing a hyperplane f (x) = ωx + b = 0, where ω is the normal vector of the classification plane and B is the offset of the classification plane. The classification function constructed is f (x) = ωx + b. The larger the distance between the parallel hyperplanes is, the smaller the error of the classifier is. The classification function of SVM is as follows: where x i is the support vector, x is the unknown vector, sign is the sign function. SVM algorithm is usually divided into the following steps. Firstly, the kernel function is used to transform the sample space into a space that can be linearly separable. Then, the partition line with the largest interval is obtained by maximizing the interval, and the support vector is obtained. Finally, the new samples can be classified and predicted by using the segmentation lines and support vectors.
Radial basis function (RBF) is one of the kernel function of SVM algorithm. In the algorithm, C and g are two important parameters [28], [29]. When RBF function is selected as the kernel function, the parameter g determines the distribution of the data mapped a new feature space implicitly. The function relationship is as follows: The larger the g is, the finer the classification will be. It attempts to distinguish each sample from other samples. If the g is too large, it will lead to over fitting. The smaller the g is, the rougher the classification will be. If the g is too small, it is easy to cause the data undistinguished and appear under fitting.
The C is the penalty factor. From the perspective of risk, C weighs the empirical risk (the fitting ability to the sample) and the structural risk (the prediction ability to the test sample). After the introduction of C, the objective functions and constraints of SVM optimization are as follows: ξ i is the relaxation factor, which is the classification loss for the sample i. If the classification is correct, it is 0. If the classification has some deviation, it corresponds to a linear value. n i=1 ξ i is the total error. The goal of optimization is that the smaller the total error is, the better the effect will be. It means that training set classification is more accurate. The optimization direction of w 2 minimization is to make the interval size 2 w maximum. The value of penalty factor C weighs the empirical risk and structural risk. The greater the C is, the less empirical risk is, the greater structural risk is, and over fitting is likely to occur. The smaller the C is, the lower the model complexity is, and under fitting is likely to occur.

B. MULTIPLE CLASSIFIERS FUSION
The process of multiple classifier fusion is shown in Figure 1. Different models are established by using different classifiers for samples. Then appropriate fusion methods are selected according to the output data of different models. Finally, the end decision of fusion methods are obtained by rules [30]- [32]. During the process of multi classifier fusion, the input samples of each single classifier are the same, but the feature information of the samples can be different. Different feature information can be used for multi classifier fusion, and the same type of classifier can be used. And also, different classifiers are used for both the same and different feature information. The output of each classifier represents the decision for the same sample. Different types of single classifier determine different classification results. Single classifier includes decision tree, support vector machine, neural network, and other classification algorithms.
The fusion method can be divided into soft decision and hard decision. For the soft decision, the output of each single classifier is a probability value or range value, of which the probability value is processed to determine the sample category. Soft decision has many common fusion algorithms, such as product method, sum method, and other algorithms. The product method multiplies the posterior probability values of all single classifiers corresponding to the same category, and takes the largest one as the final classification result. The sum method is to add the posterior probability of the corresponding category to get the maximum value, which is the final classification result.
For the hard decision, the output of each single classifier is a unique category number. The common fusion method is voting method (MV). For each single classifier, the decision is taken as one vote, and then the most voted class is counted as the final output class. The formula of voting method is as follows: where T x j is the single classifier classification result of the jth sample. The weighted voting method is based on the voting method. According to the different classification effect of each single classifier, the weight parameters of each classifier are determined by the classification accuracy of the calibration set. And then the multiple classifiers are fused according to the weight parameters. In addition, according to the empirical value, each single classifier is weighted based on the prediction probability of classification. It makes that the single classifier with good classification effect has more decision-making power.

C. EVALUATION STANDARD
Accuracy is a common evaluation standard in classification. The correct number of samples divided by all the samples is the accuracy of classification. Generally, the higher the accuracy is, the better the classifier is.

III. MATERIALS AND EXPERIMENT A. SAMPLES ACQUISITION
The sampling site is located in the intertidal zone of Dongdayang village, Qingdao, Shandong Province, China. In August 2019, 59 sediment samples were collected with the help of bamboo rafts. The collected sediment samples were put in the laboratory for air drying, crushing, low-temperature drying and grinding. All the samples were passed the 0.3mm sieve. And then these samples were screened through three layers of sieves of 0.2mm, 0.1mm and 0.075mm, respectively (Sieve method). Subsamples were obtained with different particle size of 0.3-0.2mm, 0.2-0.1mm, 0.1-0.075mm and < 0.075mm, respectively (Table 1). These subsamples were classified as four particle size type according to the classification of Udden and Wentworth [33], [34]. In these types, only the fine sand type (0.2-0.1mm) was pure, others were all mixed (Table 1). Our classification was rough, but increased the interpretation difficulties.

B. DATA ACQUISITION
The reflection spectra of the sediment samples was collected by QE65000 spectrometer and DH-2000 light source (Ocean Optics, USA). The spectral sampling interval was 1nm, the integration time was 600ms, and the spectral range was from 200nm to 1100nm. The spectrometer and the light source were connected by the Y-type optical fiber (QR400-7-UV-VIS). The Y-type optical fiber probe was fixed by the bracket (RPH-1) with the angle of 45 • . The samples were placed in a self-made sample box. The size and length of the self-made sample box were the same as that of the bracket. There were two spherical sample chambers on the sample box. The position and diameter of the sample chamber coincided with the probe hole of the bracket. More details see Li et al. [37]. Before collecting the spectrum, turned on the light source and the spectrometer for 0.5-1 hour to make them work normally. After calibration by the white board, the sample was put into the self-made sample box and gently levelled off. Then the bracket were put on the sample box tightly and the reflectance of the sample was collected. Each sample was measured 5 times and the average of its reflectance was used. The schematic diagram of sediment samples spectral measurement was shown in Figure 2. Finally, 128 reflectance spectra were collected. Since the front and back segments of the spectrum were affected by noise, only the spectral band of 226-975 nm was taken.
The concentrations of total carbon (TC) and total nitrogen (TN) in sediments with different particle size was determined using carbon and nitrogen analyzer (PE 2400II, Perkin Elmer, USA). Results shown in Table 2, the concentrations of TN and TC in sediments with different particle size was different.
C. DATA ANALYSIS SVM algorithm was used to establish the spectral model using the RBF as the kernel function. The model based on different spectral band was used as the single classifier.  The voting method was used to fuse different single classifier to make the final decision of classification. If the classification of each single classifier is not always consistent, the best classification for a single classifier model is the decision value.
Because of the complexity of sediment composition, it was difficult to find its characteristic wavelength. Therefore, the full spectral band was roughly divided. The spectral bands of the sediments were divided into 8 groups per 100nm. That is, 226-325nm, 326-425nm, 426-525nm, 526-625nm, 626-725nm, 726-825nm, 826-925nm and 926-975nm. The group 9 was full spectrum band, 226-976nm. The spectral data of these 9 different bands were normalized first, and then SVM was used to build the classification models for the four particles size types. 32 sediment samples were randomly selected as calibration set and validation set, which were used to build and validate the classification model. And other 27 sediment samples were test set, which were used to predict the classification of particle size. The proportion of the calibration set and validation set was 1:1 (64 calibration set samples, 64 validation set samples) and 2:1 (86 calibration set samples, 42 validation set samples), respectively. The calibration samples were selected by equal distance and randomly. The sampling order among particle size was sequentially 0.3-0.2mm, 0.2-0.1mm, 0.1-0.075mm, and < 0.075mm. There were 108 samples in the test set. All data analysis were completed in MATLAB R2017a. The SVM algorithm is implemented by the libsvm toolbox on MATLAB platform.

A. THE VISIBLE AND NEAR INFRARED SPECTRA OF SEDIMENTS WITH DIFFERENT PARTICLE SIZE
The average spectral reflection of 32 sediment samples with different particle size of 0.3-0.2mm, 0.2-0.1mm, 0.1-0.075mm and < 0.075mm were calculated. Four average spectra of different particle size are obtained, as shown in Figure 3. The visible and near infrared spectral reflectance was significantly different among the four particle size types of sediments. But all types had the same spectral characteristics. The reflectance increased with the wave length for all types. All types had a peak reflectance at about 580nm.
The reflectance of sediments increased with the decreased particle size (Figure 2). The reflectance was largest for the particle size <0.075mm, followed by the particle size of 0.1-0.075mm and 0.2-0.1mm, and smallest for the particle size of 0.3-0.2mm.

B. CHARACTERISTIC BAND OF REFLECTION SPECTRUM OF SEDIMENT WITH DIFFERENT PARTICLE SIZE
The proportion of calibration set and validation set was 1:1 and 2:1 by equal distance, respectively. The classification results of 9 groups' spectral band with different particle size were shown in Table 3 and Table 4. According to the Table 3, under the condition of 1:1 ratio of calibration set and validation set, the classification results under different wavebands were quite different. The calibration set and validation set accuracy of full spectrum band 226-976nm with different particle size were 95.31% and 70.31%. The classification accuracy of 326-425nm, 426-525nm, 526-625nm, 626-725nm and 726-825nm spectral band were lower than that of the full band. While the classification accuracy of 226-325nm, 826-925nm and 926-975nm spectral band were higher than that of the full band. The classification accuracy of 226-325nm and 826-925nm spectral band were the same, the accuracy of the calibration set and the validation set were both 100.0% and 73.44%. 926-975nm spectral band had the best classification result, and the classification accuracy of calibration set and validation set were 100% and 89.06%, respectively. Table 4 under the condition of 2:1 ratio of calibration set and validation set also showed that the validation set accuracy of 226-325nm, 826-925nm and 926-975nm spectral band were better than that of the full band. The calibration set accuracy of 226-325nm and 826-925nm were slightly lower than that of the full band. The classification accuracy of 926-975nm spectral band was the highest, and the accuracy of calibration set and validation set were 100% and 88.10%. Therefore, 926-975nm spectral band could best represent the reflection spectrum of different particle size sediments. Choosing 926-975nm spectral band for different particle size classification could not only improve the overall classification results, but also improved the classification accuracy of each particle size. In the 526-625nm band, the calibration set R 2 was basically the same under the different proportion of the model set and validation set. The R 2 of the 2:1 proportion increased from 53.13% to 71.43%. It showed that the 580nm characteristic band could improve the final classification results when the proportion of calibration sets was increased. However, the classification results were still not ideal, and the 580nm band did not have the ability to classify different particle size sediments.
Furthermore, whether 926-975 nm was the characteristics spectral band was tested by the randomly sampling method (Table 5-6). Table 5 was the classification results of randomly selected calibration set and validation set in the scale of 1:1, the calibration set accuracy of the two spectral band was not much VOLUME 8, 2020  different. The accuracy of 926-975nm spectral band was slightly better than that of 226-975nm, and the average accuracy of random 10 times was 98.28% and 93.13%, respectively. There was a big difference between the validation set accuracy of the two spectral band. The classification accuracy of 926-975nm spectral band was significantly higher than that of 226-975nm spectral band, and the average accuracy of random 10 times was 86.72% and 66.25%, respectively. The validation set accuracy of 226-975nm spectral band was between 60% and 70%. While the validation set accuracy of 926-975nm spectral band was higher than 80%, and the best classification accuracy was 90.63%. Table 6 showed the same regularity in the classification results of randomly selected calibration set and validation set in the scale of 2:1. The calibration set accuracy of 926-975nm spectral band was slightly better than that of 226-975nm spectral band. The validation set accuracy of 926-975nm spectral band was significantly higher than that of 226-975nm spectral band. The average accuracy of validation set were 91.40% and 73.02% respectively. The validation set accuracy of 926-975nm spectral band in random 10 times was about 90%, while that of 226-975nm spectral band was about 70%. Therefore, the random division of calibration set and validation set could also verify that 926-975nm spectral band was the most representative reflection spectrum of different particle size sediments. 926-975nm spectral band could be used as the characteristic spectral band to divide four particle sizes of 0.3-0.2mm, 0.2-0.1mm, 0.1-0.075mm and < 0.075mm.

C. MULTIPLE CLASSIFIERS FUSION WITH SINGLE CLASSIFIER IN EACH BAND
It can be seen from the table 2 and table 3 that the calibration set and validation set accuracy of 226-325nm, 826-925nm and 926-975nm spectral band were all higher than that of the full spectral band 226-975nm. The first four spectral bands with classification accuracy ranging from large to small were 926-975nm, 226-325nm, 826-925nm and 226-975nm spectral band. The classification results of 226-325nm and 826-925nm spectral band were consistent. Therefore, the four band classification model were selected as single classifier. Using the four single classifiers of 926-975nm, 226-325nm, 826-925nm and 226-975nm, five recommendation results were got. Recommendation  Table 7. The overall classification accuracy were shown in Figure 4.  As the table 7 shown, the classification accuracy of each four particle size in the 926-975nm spectral band was <0.075mm, 0.2-0.1mm, 0.3-0.2mm and 0.1-0.075mm particle size in order from large to small. The accuracy of 0.3-0.2mm particle size was the same as that of 0.1-0.075mm particle size. The classification accuracy of each four particle sizes was more than 80%, and the accuracy of <0.075mm particle size was 100%. The accuracy of each four particle size in 826-925nm spectral band was < 0.075mm, 0.3-0.2mm, 0.2-0.1mm and 0.1-0.075 mm particle size in order. The accuracy of 0.2-0.1mm particle size was the same as that of 0.1-0.075mm particle size. The accuracy of two particle size was more than 80%, while that of the other two particle size was less than 70%. The accuracy of each four particle size in 226-325nm spectral band was 0.3-0.2mm, < 0.075mm, 0.2-0.1mm, 0.1-0.075mm particle size in order. The accuracy of 0.2-0.1mm particle size and 0.1-0.075mm particle size were the same. The classification results of each four particle sizes were uneven, the best classification accuracy was 96.75%, while the worst was only 56.25%. The accuracy of each four particle size classifications in 226-945nm spectral band was < 0.075mm, 0.2-0.1mm, 0.1-0.075mm, 0.3-0.2mm particle size in order. The accuracy of 0.2-0.1mm particle size was the same as that of 0.1-0.075mm particle size. Only one of the four particle size classification results reached more than 80%, and the other three particle size classification results were less than 70%.
When the particle size was 0.3-0.2mm, the classification accuracy of 226-325nm spectral band was the highest, only one sample was misclassified, while the lowest was 226-945nm spectral band. Under the three particle sizes of 0.2-0.1mm, 0.1-0.075mm and < 0.075mm, the highest classification accuracy was 926-975nm spectral band, and the lowest was still 226-945nm spectral band. This result was basically consistent with the overall classification results. The best spectral band was 926-975nm, while the worst spectral band was 226-945nm ( Figure 3).
In the recommendation result 1, the classification results of 0.3-0.2mm, 0.2-0.1mm and < 0.075mm particle size were all up to the optimal classification accuracy of the three single classifiers (926-975nm, 226-325nm and 826-925nm spectral band), which was 93.75%, 93.75% and 100% respectively. The classification accuracy of 0.1-0.075mm particle size was higher than that of two single classifiers, but lower than the optimal classification accuracy of a single classifier. Compared with the optimal classification accuracy, one more sample was misjudged. Because the classification accuracy of 0.1-0.075mm particle size at 226-325nm spectral band was too low, only 56.25%, which affected the final recommended classification results. Among the four particle size classification results, the overall classification accuracy of recommendation result 1 was 93.75% (Figure 3), which was higher than that of every single classifier. The number of correct samples was increased by one sample compared with the optimal single classifier in 926-975nm spectral band.
In the recommendation result 2, only the classification results of < 0.075mm particle size reached the optimal classification accuracy of four single classifier in 926-975nm, 226-325nm, 826-925nm and 226-975nm spectral bands, which was 100%. The other three particle sizes did not reach the optimal classification accuracy of four single classifier. The overall classification accuracy was better than the single classifier of 226-325nm, 826-925nm and 226-975nm, but lower than that of 926-975nm.
Recommendation result 3, recommendation result 4 and recommendation result 5 were all lower than the recommendation result 1. This three recommendation results did not choose the best three single classifier, which could not cause the better results. Therefore, the single classifiers chosen were very important, which decided the final recommendation results.
108 sediment samples of test set were predicted in the classification model based on 926-975nm, 226-325nm, 826-925nm and 226-975nm spectral band. The classification accuracy and the number of correct samples of four spectral bands and recommendation 1-5 with overall classification and each four particle sizes were shown in table 8.
According to table 8, in the classification results of four spectral bands, the overall classification results of 926-975nm spectral band were still the best, followed by 826-925nm, 226-975nm and 226-325nm spectral band. The overall accuracy of the first three spectral bands were more than 80%. The classification results of 0.3-0.2mm and < 0.075mm were better, which were more than 90%. Among the classification results of recommendation 1-5 based on multiple classifiers fusion, recommendation 2 (the classification results fused by four single classifiers) had the highest accuracy of overall classification and each particle sizes, and the overall accuracy was 90.74%. The second was recommendation 1, recommendation 3 and recommendation 4, and the overall accuracy was 89.81%. Their common feature was that they had the single classifier with highest accuracy, which was 926-975nm spectral band. The overall accuracy of the five recommendation results was higher than that of each single classifier.

V. DISCUSSION
As hypothesis 1, the reflectance of the sediments increased with the decreased particle size because smaller size particle could produce larger reflectance ( Figure 2). The pattern had been rarely reported, but was the basis of spectral analysis. And 926-975 nm was the characteristic spectra of the particle size of sediments. As hypothesis 2, the particle size of sediments could be spectrally analyzed by stoichiometry. In this study, we found that the particle size of sediments could be quantitatively analyzed by the multi spectral classifier fusion. We found that the classification accuracy of fusion classifier was determined by the compositions of the single classifiers. Therefore, getting the characteristic spectra was the most important step.
Whether for chemical or physical parameters, the characteristic wavelength is the key to spectral analysis. Li et al. [38] extracted the characteristic wavelengths of soil total nitrogen (TN) by non-information variable elimination (UVE) and successive projection algorithm (SPA), and obtained better prediction results of total nitrogen content. Jiang et al. [39] selected two characteristic bands of PCs by principal component analysis (PCA), and obtained higher accuracy for detection of pear bruise. Zhou et al. [40] used wavelet transform to extract the characteristic wavelength and established the water content model of lettuce samples, which improved the accuracy of water content detection. In this paper, the characteristic wavelengths of sediment particle size were selected manually, and the classification results were improved to a certain extent. By analyzing the classification results with different particle size in each band, the optimal characteristic band was 926-975nm spectral band, followed by 226-325nm and 826-925nm spectral band. Compared with the full band 226-975nm spectral band, the classification accuracy had been improved. It showed that the 226-975nm contained the spectral band which was not related to the particle size characteristics, that is useless information. 426-525nm, 626-725nm, 726-825nm and other bands had low classification accuracy, which affected the full band classification results. The correct extraction of characteristic wavelengths is very important in spectral analysis, which can characterize the key information of sediment particle size. In this paper, the characteristic wavelength is a rough band range, furthermore we will study the band range which can better characterize its characteristics. In the future, the algorithm that can automatically select the characteristic wavelength of sediment will be studied, to further improve the classification result of sediment particle size.
In the spectral model of marine sediment with different particle sizes, different proportion of model set and validation set, different division methods of model set and validation set and different characteristic bands of marine sediment will lead to different models. In particular, the selection of the characteristic spectral band which can best characterize the sediment particles has the greatest impact on the model results. Different spectral bands contain different information, and different information can represent different characteristics of marine sediment. In order to improve the accuracy of the model, the characteristic spectral band which can best represent the sediment particles is selected to build the model. Therefore, it is very important to choose the appropriate model set method and characteristic spectral band for the establishment of stable and accurate model.
In the research of the fusion with the single classifier, the classification accuracy in order from large to small was recommendation result 1 > the optimal single classifier > recommendation result 2, recommendation result 3, recommendation result 4 and recommendation result 5. The reason was that the classification results of four single classifiers were uneven. Poor single classifier had certain influence on final particle size classification. The useless information or noise information would be introduced into the fusion results, which leads to unsatisfactory final classification result. The classification accuracy of single classifier (926-975nm spectral band) was much higher than the other three single classifiers (226-325nm, 826-925nm and 226-975nm spectral band). The classification accuracy of the three single classifiers would affect the results of the single classifier (926-975nm spectral band). In the test set, the four single classifiers have similar classification accuracy, which can complement other's classification results, so that the classification accuracy of all recommended results is higher than that of each single classifier.
The physical properties of the sediment can be estimated with the spectra baseline that is affected by the grain size in the visible and near-infrared ranges [35]. Jacq et al. [36] used visible and near-infrared hyperspectral imaging to predict particle size fractions and distribution at a resolution of 200 µm on a previously well-studied sediment core taken from Lake Bourge. In this paper, different particle sizes of intertidal sediments were classified by visible near infrared spectroscopy. Compared to using the UAV red, green and blue (RGB) orthophotometer, the accuracy assessment of the surface sediment classification based on these six types indicated an overall accuracy of 72.8% [18]. The best classification accuracy of this paper is up to 93.75%. The former was based on the sediment in the field, while this paper was aimed at the simple treated sediment samples. And this method will be applied to the analysis of the sediment in the field later. The result of the sieve method was used as the criteria. Compared with the sieve method, the classification accuracy was higher and the classification speed was much faster than that of sieve method. Visible -near infrared spectroscopy will be a new technique for rapid particle size analysis of Marine sediments.

VI. CONCLUSION
The particle size of sediment has a close relation to the identification of sediment source, the influence of organic matter in sediments, and the formation of sediment hydrate. Therefore, it is of great significance to analyze the different particle size of marine sediments. We studied the classification of particle size of marine sediments in Qingdao, China by the visible and near infrared spectroscopy. The sediments were divided into four particle sizes of 0.3-0.2mm, 0.2-0.1mm, 0.1-0.075mm and <0.075mm by sieve method. With decreased particle size, the reflectance increased greatly. A negative correlation was between the spectral reflectance and the particle size. The characteristic spectrum located at 900-975nm; therefore, the particle size could be classified well by the characteristic spectra. Using the single classifier from the spectral band of 226-325nm, 826-925nm and 926-975nm, multi-classifier fusion was got and better than all single classifiers. The classification accuracy of the validation set could be as high as 93.75%. In this paper, visible near-infrared spectroscopy combined with multi-classifier fusion technology is creatively applied to the classification of marine sediment particle size, so as to quickly and accurately classify different particle sizes of marine sediment. This study provides a new method for locating the characteristic spectra of sediments and the rapid determination of particle distribution.