Lithological Classification Using Sentinel-2 A Data in the Shibanjing Ophiolite Complex in Inner Mongolia , China

As a source of data continuity between Landsat and SPOT, Sentinel-2 is an Earth observation mission developed by the European Space Agency (ESA), which acquires 13 bands in the visible and near-infrared (VNIR) to shortwave infrared (SWIR) range. In this study, a Sentinel-2A imager was utilized to assess its ability to perform lithological classification in the Shibanjing ophiolite complex in Inner Mongolia, China. Five conventional machine learning methods, including artificial neural network (ANN), k-nearest neighbor (k-NN), maximum likelihood classification (MLC), random forest classifier (RFC), and support vector machine (SVM), were compared in order to find an optimal classifier for lithological mapping. The experiment revealed that the MLC method offered the highest overall accuracy. After that, Sentinel-2A image was compared with common multispectral data ASTER and Landsat-8 OLI (operational land imager) for lithological mapping using the MLC method. The comparison results showed that the Sentinel-2A imagery yielded a classification accuracy of 74.5%, which was 2.5% and 5.08% higher than those of the ASTER and OLI imagery, respectively, indicating that Sentinel-2A imagery is adequate for lithological discrimination, due to its high spectral resolution in the VNIR to SWIR range. Moreover, different data combinations of Sentinel-2A + ASTER + DEM (digital elevation model) and OLI + ASTER + DEM data were tested on lithological mapping using the MLC method. The best mapping result was obtained from Sentinel-2A + ASTER + DEM dataset, demonstrating that OLI can be replaced by Sentinel-2A, which, when combined with ASTER, can achieve sufficient bandpasses for lithological classification.


Introduction
There are about 5% of territories in China that belong to Gobi Desert, mainly distributed in the Northwest China.Due to vastness, poor accessibility, and atrocious weather conditions in these areas, regional lithological mapping in the field is costly and challenging.As an economical and efficient technique, remote sensing has become a popular method for regional lithological mapping, especially in arid and semi-arid regions [1][2][3][4].Optical imagery acquired by spaceborne and airborne sensors has been widely applied to mineral and lithological exploration for decades.For instance, the United States Geological Survey (USGS) used ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) data to map the hydrothermally altered rocks in the central and southern parts of the Basin and Range province of the United States [5].
Due to its lower cost and higher accessibility than hyperspectral data, multispectral imagery, such as Landsat-5 TM, Landsat-7 ETM+ (enhanced thematic mapper plus), Landsat-8 OLI (operational land imager), and ASTER, is broadly utilized to extract lithologic, mineral, and structural information in metallogenic provinces [6][7][8].As one of the world's earliest and longest continuously acquired collection of spaceborne moderate-resolution land remote sensing data, the Landsat series has been applied in geology for decades [6,[9][10][11], for applications such as hydrothermal alteration (ferric iron and hydroxides) extraction, lithological discrimination, and geotectonic interpretation.With a higher spectral resolution in the shortwave infrared (SWIR) and thermal infrared (TIR) range than other multispectral data, ASTER provides a higher potential to undertake semi-quantitative mineral mapping.It has become the most popular imagery in geological exploration, especially in hydrothermal alteration and lithological unit discrimination, since the Terra satellite was launched in 1999 [7, [12][13][14].For example, Son et al. [15] utilized ASTER data to map distribution patterns of hydrothermal alteration and igneous rocks in the southwestern Gobi in Mongolia.Yoshiki et al. [16] applied ASTER TIR (thermal infrared) images to map regional lithological rocks in the Tibetan Plateau.However, with Landsat 7 being damaged in 2003 and Landsat 5 retiring in 2013, only Landsat 8, launched in 2013, remains within the Landsat mission for routine operation.Moreover, due to an anomalously high SWIR detector temperature, ASTER SWIR data have been unavailable since April 2008.
Thereafter, the Sentinel-2 mission, launched by the European Space Agency (ESA), acts as important data continuity and enhancement for the Earth observation.The Sentinel-2 Multi-Spectral Imager (MSI) consists of identical Senitnel-2A and Sentinel-2B, which were launched on 23 June 2015 and 7 March 2017, respectively.With a short revisit time (every 5 days with two satellites), the Sentinel-2 missions aimed at global monitoring for environment and security, such as forest monitoring, land cover change detection, and natural disaster management [17,18].In addition, the high-resolution Sentinel-2 multispectral data were confirmed to have potential for mineral mapping in geological investigations in the last several years, especially for the iron absorption feature, due to the similar or even higher spectral setting than Landsat series and SPOT in the VNIR region [19][20][21].However, there is rare research exploring the capability of Sentinel-2 data for complex lithological classification.
Although it is not new to use remote sensing technique for lithological classification in geological investigation [2,22,23], many studies are limited, due to the coarse spatial/spectral resolutions of multispectral data, causing difficulties in accurately classifying rock units [22].As a solution, multiple ancillary data with texture information, such as airborne geophysical data [24], DEM [25], and geomorphic feature [2], can be integrated with multispectral imagery for improved lithological classifications [22].However, the integration of multispectral data with different bandpasses for lithological classification is of little concern in previous literature.
In this study, lithological classification was performed in the Shibanjing ophiolite complex in the Beishan orogenic belt in Inner Mongolia, China.The present research aims at investigating the potential of the Sentinel-2A and the integrations of multispectral imagery for lithological classification in the Shibanjing ophiolite complex.Specifically, (i) five typical machine learning methods, including artificial neural network (ANN), k-nearest neighbor (k-NN), maximum likelihood classification (MLC), random forest classifier (RFC), and support vector machine (SVM), were compared to select the optimal classifier for lithological classification using the data combination of Sentinel-2A and DEM (Sentinel-2A + DEM); (ii) three data combinations, OLI + DEM, ASTER + DEM, and Sentinel-2A + DEM, were employed to classify lithological units, which aimed to evaluate the capability of Sentinel-2A for lithological mapping; and (iii) two data combinations, OLI + ASTER + DEM and Sentinel-2A + ASTER + DEM, were utilized to compare the capability of different combinations of multispectral bands for lithological classification.

Study Area Description
The study area is located at 41  1a).The Beishan orogenic belt, located in the southern part of the Central Asian Orogenic Belt (CAOB), is composed of ophiolites, arcs, and blocks formed by subduction-accretion between the Tarim craton and the Kazakhstan plate [26] (Figure 1b).The Beishan orogenic belt represents a major part of the southern CAOB, which makes it a critical area for unraveling the evolution and accretionary orogenic history of the CAOB [27].There are several complete ophiolites in the Beishan orogenic collage, which possibly formed in the early Paleozoic era in the southern CAOB.From north to south, three main ophiolitic belts are distributed along the Beishan orogenic belt: the Hongshishan ophiolitic belt, the Mingshui-Shibanjing-Xiaohuangshan ophiolitic belt, and the Hongliuhe-Niujuanzi-Xichangjing ophiolitic belt, as shown in Figure 1b [28].

Study Area Description
The study area is located at 41°31′24″-41°36′26″N and 98°19′15″-98°29′53″E.It consists of the Shibanjing ophiolite in Beishan orogenic belt in Inner Mongolia, China (Figure 1a).The Beishan orogenic belt, located in the southern part of the Central Asian Orogenic Belt (CAOB), is composed of ophiolites, arcs, and blocks formed by subduction-accretion between the Tarim craton and the Kazakhstan plate [26] (Figure 1b).The Beishan orogenic belt represents a major part of the southern CAOB, which makes it a critical area for unraveling the evolution and accretionary orogenic history of the CAOB [27].There are several complete ophiolites in the Beishan orogenic collage, which possibly formed in the early Paleozoic era in the southern CAOB.From north to south, three main ophiolitic belts are distributed along the Beishan orogenic belt: the Hongshishan ophiolitic belt, the Mingshui-Shibanjing-Xiaohuangshan ophiolitic belt, and the Hongliuhe-Niujuanzi-Xichangjing ophiolitic belt, as shown in Figure 1b [28].The Shibanjing ophiolite complex is in the Mingshui-Shibanjing-Xiaohuangshan ophiolite belt, which is approximately 80 km away from the Xiaohuangshan ophiolite (Figure 1b).Structure activities in the complex area are intense and characterized by NW strike-slip faults, which are the main faults controlling the distribution of formation and magmatic rocks.The ophiolite complex is characterized by tectonic slices of ophiolitic rocks, including meta-ultramafic rocks, mylonitic gabbros, metabasalts, and clastic rocks in a matrix of turbidites, and the mélange contains blocks of argillaceous matrix and marble, where many ophiolitic rocks in the complex experienced amphibolite facies metamorphism [31][32][33], as shown in Figure 1c.The emergent strata are Cambrian, Ordovician, Silurian, Devonian, and Quaternary.The magmatic rocks are dominated by the felsic intrusion, and the mafic and ultrabasic rock unit is one of the important rock units that make up the ophiolitic complex.The basic and ultrabasic rocks mainly consist of pyroxenite, olivinite, and augite peridotite (Figure 1c).The geological map of the Shibanjing ophiolite complex displayed in Figure 1c was generated via visual interpretation of remote sensing data, and then verified and modified in the field in June 2016.The arid study area with sparse vegetation cover and well-exposed rocks is an ideal region to test lithological mapping by remote sensing data.

Data-Preprocessing
Three types of multispectral imagery, Sentinel-2A, ASTER and OLI, were downloaded from the USGS website (https://glovis.usgs.gov/)and employed in this study.In addition, one scene of digital elevation model (DEM) data of ALOS/PALSAR with 12.5 m spatial resolution was obtained from the NASA website (https://vertex.daac.asf.alaska.edu/#).The Sentinel-2A image contains 13 spectral bands in the VNIR and SWIR spectral range, with four bands at 10 m, six bands at 20 m, and three atmospheric correction bands at 60 m spatial resolution [34], as shown in Table 1.The cloud free image was automatically atmospherically corrected and orthorectified using the Sentinel Application Platform (SNAP) software package provided by ESA.After atmospheric correction, all the image bands were cubically resampled to 20 m spatial resolution and layer stacked to one file.Since bands 1, 9, and 10 were designed for atmospheric correction, they were excluded from lithological classification in this study.As an advanced multispectral sensor launched onboard Terra spacecraft in December 1999, ASTER covers a broad range of spectral region with 14 spectral bands, including three VNIR bands with 15 m spatial resolution, six SWIR bands with 30 m spatial resolution, and five TIR bands with 90 m spatial resolution.In addition, one more telescope is used to see backward in the near infrared spectral band (Band 3B) for stereoscopic capability [35] (Table 1).The level-1B ASTER data employed in this research were acquired on 14 August 2002.The data without cloud were atmospherically corrected and calibrated using the ENVI 5.4 software package as follows.The VNIR and SWIR bands of the ASTER data were cubically resampled to 30 m spatial resolution, layer stacked to construct a single 9-band file, and atmospherically corrected using the FLAASH (Fast Line of Sight Atmospherics Analysis of Hypercubes) module, to eliminate the effects caused by water vapor and cloud and to convert the digital counts to surface reflectance.A cross-track illumination correction was applied to the ASTER images to remove the effects of energy overspill from band 4 into bands 5 to 9.
As the latest mission in Landsat series launched by NASA on 11 February 2013, Landsat 8 consists of a two-sensor payload, the OLI, and the TIRS (thermal infrared sensor).The OLI contains two more bands than Landsat 7 (ETM+), including a new deep blue band for coastal/aerosol observation and an SWIR band for cirrus detection (Table 1).In this study, one OLI scene acquired on 5 December 2017, was selected for lithological discrimination.The VNIR and SWIR bands of OLI were layer stacked to form a single file with 30 m spatial resolution.A radiation correction and an atmospherical correction by FLAASH were successively applied to the OLI/Landat-8 scene to eliminate the effects caused by atmosphere.
The atmospherically corrected Sentinel-2A, OLI and ASTER imagery was georeferenced to the UTM (Universal Transverse Mercator projection) coordinate system in Zone 47 north, using ground control points chosen from the DEM and high-resolution orthorectified Google Earth imagery.The information about the three multispectral images employed in this study is listed in Table 2.Although these images were acquired in different seasons, maximum normalized different vegetation index (NDVI) of each image is lower than 0.07, indicating that the study area is an ideal exposed region with almost no vegetation cover for lithological classification.In addition, there are barely snow or cloud covered regions in the study area, as shown in Table 2. Therefore, there are no vegetation, cloud or snow effects in this study.

Lithological Mapping
Sentinel-2A data contains higher spectral and spatial resolution in the VNIR to SWIR range than common multispectral data, such as OLI and ASTER data.To find the optimal classifier for the classification assessment of Sentinel-2A data in the Shibanjing ophiolite, five typical machine learning techniques, ANN, k-NN, MLC, RFC, and SVM, were applied for lithological mapping.
Different data combinations were designed for the assessment of multispectral data from Sentinel-2A for lithological classification, as listed in Table 3.Firstly, considering DEM data can provide geomorphological and subsurface geological information [36], data integration of Sentinel-2A and a DEM (Sentinel-2A + DEM) were compared with Sentinel-2A data alone for lithological classification, for the purpose of evaluating the importance of DEM data for lithological classification in this study.Secondly, three datasets (OLI + DEM, Sentinel-2A + DEM, and ASTER + DEM) were employed in turn for lithological discrimination, in order to compare the capability of Sentinel-2A for lithological classification with other data.Thirdly, with the band configurations of OLI, Sentinel-2A, and ASTER in the VNIR and SWIR range, the data combinations OLI + ASTER or Sentinel-2A + ASTER are expected to provide more diagnostic spectral features of lithological units than the OLI, ASTER, or Sentinel-2A data alone.Therefore, the OLI + ASTER + DEM and Sentinel-2A + ASTER + DEM datasets were compared for lithological classification.After that, two classification results were compared to evaluate whether OLI could be substituted by Sentinel-2A.Figure 2 illustrates the above process in a simplified sequence flow diagram.

Training and Testing Samples
The geological map depicted in Figure 1c was used as a reference for the selection of the training and testing samples.According to the texture characteristics and distributions of the fifteen lithological units (except the quartz, because of its small area) shown in Figure 1c, the training dataset was carefully selected manually by visual interpretation.A total of 12,000 pixels of ASTER data for the fifteen lithological units were selected, which represents approximately 1.9% of the entire study area.The training datasets were directly utilized by the ANN, k-NN, MLC, RFC, and SVM classifiers.Moreover, 1200 ground truth points randomly distributed in the study area were selected as a testing dataset according to the geological map (Figure 1c).The areas, training samples, and testing samples of lithological units are listed in Table 4.

Training and Testing Samples
The geological map depicted in Figure 1c was used as a reference for the selection of the training and testing samples.According to the texture characteristics and distributions of the fifteen lithological units (except the quartz, because of its small area) shown in Figure 1c, the training dataset was carefully selected manually by visual interpretation.A total of 12,000 pixels of ASTER data for the fifteen lithological units were selected, which represents approximately 1.9% of the entire study area.The training datasets were directly utilized by the ANN, k-NN, MLC, RFC, and SVM classifiers.Moreover, 1200 ground truth points randomly distributed in the study area were selected as a testing dataset according to the geological map (Figure 1c).The areas, training samples, and testing samples of lithological units are listed in Table 4.

Machine Learning Methods
Lithological classification was performed using the five machine learning methods ANN, k-NN, MLC, RFC, and SVM.In the following subsections, a brief explanation of the five algorithms is provided.

Artificial Neural Network
As a typical machine learning method, the artificial neural network classifier, also known as neural network (NN), is widely employed in pattern recognition and the classification of image data.The NN classifier is an artificial intelligence technique, which attempts to simulate the way in which humans classify patterns, learn tasks, and solve problems [37].It is made up of a number of simple processing units called nodes or neurons [38].The nodes are linked by weighted connections according to a specified architecture.There are three layers in the ANN classifier, an input layer, a middle layer (i.e., hidden layer), and an output layer.Each layer of the ANN classifier consists of one or more nodes, which are adjusted through the iterative experiment to obtain the most reasonable output [37].
In this study, a multi-layer feed-forward ANN method was employed with the S2A_DEM dataset for lithological classification in the ENVI 5.4 software.The ANN classifier used a logistic activation function, the training threshold contribution value was 0.9, the training rate was 0.2, the training momentum was 0.9, and the training root mean square (RMS) exit criterion was 0.1.

k-Nearest Neighbors
The nonlinear classifier k-nearest neighbors (k-NN) is one of the most common machine learning techniques, and a favorite classification algorithm in statistical applications [39].As a non-parametric algorithm, k-NN employs an instance-based leaning algorithm, or a "lazy learning," to find a group of K samples nearest to unknown samples [40].K is a key parameter and plays a significant role in the performance of the k-NN classifier [40,41].In this study, the optimal value of k was 5.The k-NN classifier was utilized for lithological classification using the S2A_DEM dataset based on C++ language.

Maximum Likelihood Classifier
The maximum likelihood classification (MLC) is one of the most common supervised classifiers in remote sensing [42][43][44], and was proposed firstly by German mathematician C.F. Gauss in 1821 for normal distribution.This method is based on the hypothesis that the probability density function for each class is multivariate, and an unknown pixel is assigned to a class with the highest probability of belonging [42,45].As a common multivariate statistical classification method, MLC, is embedded in many image processing software packages, such as PCI, ERDAS, and ENVI.In this study, this classifier was applied to varied datasets for lithological classification using ENVI 5.4 software provided by Exelis Visual Information Solutions in USA.

Support Vector Machine
The support vector machine (SVM) is one of the most widely used supervised machine learning algorithms; it is based on statistical learning theory [25,43,46,47], which was proposed by Vapnik and Chervonenkis in 1963 [48].The classification mechanism of SVM is to find a hyperplane to classify the given dataset in an n-dimensional space.One reasonable choice as the best hyperplane is the one that represents the largest separation between two classes.Target detection using SVM method can be simplified as a dichotomy.In a higher dimensional space, the SVM method needs an optimal linear hyperplane with the maximum margin for separating the given dataset.As the given data in a larger dimensional space can be complex, the kernel function is introduced to address this issue [49,50]: As proposed by previous research, there are several common kernel functions for the SVM method, such as linear, polynomial, radial basis function, and sigmoid [41].
Being one of the widely applied multivariate statistical learning methods, SVM has been embedded in many software packages.In this study, the one against one support vector machine (OAO-SVM) classifier accomplished in C++ was employed for lithological mapping.The radial basis function was selected as the kernel type, the penalty parameter was set to 100, and the gamma in kernel function was the inverse of the band number of the S2A_DEM dataset, namely, 0.091 [51].

Random Forest Classifier
The random forest classifier (RFC) has become a favorable and efficient classifier for scientists in many fields, such as geology and ecology [52][53][54].This method is a supervised machine learning approach that contains multiple decision trees and combines the predictions from all trees [52].Each decision tree returns a classification, and a random forest decides which class each observation is attributed to, based on majority rule, that the class that has the most votes across all trees is determined as the final class [55].
To implement the RFC method, two parameters need to be set up: the number of trees, and the number of features in each split [56].In the present work, the number of trees is 1000, and the number of features in each split is 30.The RFC method was performed in ArcGIS 10.4 software.

Results
The classification accuracy of the lithological map obtained from each classifier and dataset described above was assessed using the overall accuracy, average accuracy (the average of user's and producer's accuracies), and Kappa coefficient, obtained from the confusion matrix.The overall accuracy is calculated as the total number of correctly classified pixels (diagonal elements of the confusion matrix) divided by the total number of test pixels.The producer's accuracy is the probability that the classifier has correctly labeled an image pixel, and the user's accuracy indicates the probability that the classifier has correctly labeled a pixel into its pre-given class.The Kappa coefficient is a measure of agreement between the classified map and the reference data.The Kappa coefficient is different from the overall accuracy and takes into account the entire contingency matrix, which is designed to measure the consistency of the results [57].The assessment was to estimate the classification results obtained from different methods and datasets, evaluating the capability of classification techniques for lithological mapping in the study area.

Lithological Discrimination Using Machine Learning Methods
Lithological classification of the S2A_DEM dataset using the different machine learning methods of ANN, k-NN, MLC, RFC, and SVM are displayed in Figure 3.The lithological results obtained using k-NN show more noise than other results, as illustrated in Figure 3.The mapping results of granodiorite and quartz diorite using k-NN are largely misclassified into surrounding lithological units.In addition, Remote Sens. 2018, 10, 638 9 of 22 the dolomite is completely misclassified into basalt by the ANN method, as shown in Figure 3c.The lithological units extracted using the methods of SVM and MLC are more precise than the results obtained from the ANN and SVM techniques, according to the geological map (Figure 1c).However, due to their similar compositions and adjacent positions, it is difficult to distinguish carbonatite with sandstone from quartz sandstone and siltite using these five machine learning methods.

Lithological Discrimination Using Machine Learning Methods
Lithological classification of the S2A_DEM dataset using the different machine learning methods of ANN, k-NN, MLC, RFC, and SVM are displayed in Figure 3.The lithological results obtained using k-NN show more noise than other results, as illustrated in Figure 3.The mapping results of granodiorite and quartz diorite using k-NN are largely misclassified into surrounding lithological units.In addition, the dolomite is completely misclassified into basalt by the ANN method, as shown in Figure 3c.The lithological units extracted using the methods of SVM and MLC are more precise than the results obtained from the ANN and SVM techniques, according to the geological map (Figure 1c).However, due to their similar compositions and adjacent positions, it is difficult to distinguish carbonatite with sandstone from quartz sandstone and siltite using these five machine learning methods.The classification accuracies and Kappa coefficients of mapping results of S2A_DEM using different machine learning methods are displayed in Figure 4.The classification using the k-NN method shows the lowest accuracies and Kappa coefficient in Figure 4, in keeping with Figure 3.The average accuracies of most lithological units obtained by k-NN are lower than 70%.The ANN and RFC methods show similar overall accuracies and Kappa coefficients, which are lower than those of the MLC and SVM techniques.Moreover, the classification of dolomite obtained by the ANN method is 0%, and the accuracies of more than 10 lithological units extracted by ANN and RFC were lower than 70%, indicating that these two methods (ANN and RFC) are not suitable for lithological The classification accuracies and Kappa coefficients of mapping results of S2A_DEM using different machine learning methods are displayed in Figure 4.The classification using the k-NN method shows the lowest accuracies and Kappa coefficient in Figure 4, in keeping with Figure 3.The average accuracies of most lithological units obtained by k-NN are lower than 70%.The ANN and RFC methods show similar overall accuracies and Kappa coefficients, which are lower than those of the MLC and SVM techniques.Moreover, the classification of dolomite obtained by the ANN method is 0%, and the accuracies of more than 10 lithological units extracted by ANN and RFC were lower than 70%, indicating that these two methods (ANN and RFC) are not suitable for lithological discrimination in the Shibanjing ophiolite complex.As shown in Figures 3 and 4, the two statistical learning methods of SVM and MLC yielded quite good performances in lithological classification.Due to its slightly higher overall accuracy, the MLC approach was employed to evaluate the potential of Sentinel-2A imagery to discriminate lithological units in the following research.The classification accuracies and Kappa coefficients of mapping results of S2A_DEM using different machine learning methods are displayed in Figure 4.The classification using the k-NN method shows the lowest accuracies and Kappa coefficient in Figure 4, in keeping with Figure 3.The average accuracies of most lithological units obtained by k-NN are lower than 70%.The ANN and RFC methods show similar overall accuracies and Kappa coefficients, which are lower than those of the MLC and SVM techniques.Moreover, the classification of dolomite obtained by the ANN method is 0%, and the accuracies of more than 10 lithological units extracted by ANN and RFC were lower than 70%, indicating that these two methods (ANN and RFC) are not suitable for lithological discrimination in the Shibanjing ophiolite complex.As shown in Figures 3 and 4, the two statistical learning methods of SVM and MLC yielded quite good performances in lithological classification.Due to its slightly higher overall accuracy, the MLC approach was employed to evaluate the potential of Sentinel-2A imagery to discriminate lithological units in the following research.

Lithological Discrimination Using S2A and S2A_DEM Datasets
The classification results of two datasets, S2A and S2A_DEM, using MLC technique, are displayed in Figure 5.As shown in Figure 5a, there is more noise for the S2A dataset in the lithological classification.Affected by the extensive weathering and complex terrain, the alluvium deposits were largely misclassified into surrounding lithological units.Moreover, with the influence of topography, other lithological units are also misclassified into the alluvium deposits, as displayed in Figure 5a. Figure 5b is the partial magnification of the classification of Sentinel-2A, which clearly shows the misclassified region.By combining DEM data with the Sentinel-2A image, the mapping results are dramatically improved, as shown in Figure 5c.The alluvium deposits are effectively differentiated from surrounding lithological units, and most noise is eliminated.

Lithological Discrimination Using S2A and S2A_DEM Datasets
The classification results of two datasets, S2A and S2A_DEM, using MLC technique, are displayed in Figure 5.As shown in Figure 5a, there is more noise for the S2A dataset in the lithological classification.Affected by the extensive weathering and complex terrain, the alluvium deposits were largely misclassified into surrounding lithological units.Moreover, with the influence of topography, other lithological units are also misclassified into the alluvium deposits, as displayed in Figure 5a. Figure 5b is the partial magnification of the classification of Sentinel-2A, which clearly shows the misclassified region.By combining DEM data with the Sentinel-2A image, the mapping results are dramatically improved, as shown in Figure 5c.The alluvium deposits are effectively differentiated from surrounding lithological units, and most noise is eliminated.
largely misclassified into surrounding lithological units.Moreover, with the influence of topography, other lithological units are also misclassified into the alluvium deposits, as displayed in Figure 5a. Figure 5b is the partial magnification of the classification of Sentinel-2A, which clearly shows the misclassified region.By combining DEM data with the Sentinel-2A image, the mapping results are dramatically improved, as shown in Figure 5c.The alluvium deposits are effectively differentiated from surrounding lithological units, and most noise is eliminated.Figure 6 shows the average accuracies of each lithological unit based on the S2A and S2A_DEM datasets using the MLC technique.As with Figure 5, the classification accuracies of the S2A_DEM dataset are all higher than those of the single Sentinel-2A data, demonstrating the significance of DEM data for lithological mapping.The overall accuracy of the classification using the S2A_DEM dataset is 74.5%, which is 8.25% higher than that of the S2A dataset, as shown in Figure 7.   Figure 6 shows the average accuracies of each lithological unit based on the S2A and S2A_DEM datasets using the MLC technique.As with Figure 5, the classification accuracies of the S2A_DEM dataset are all higher than those of the single Sentinel-2A data, demonstrating the significance of DEM data for lithological mapping.The overall accuracy of the classification using the S2A_DEM dataset is 74.5%, which is 8.25% higher than that of the S2A dataset, as shown in Figure 7. Figure 6 shows the average accuracies of each lithological unit based on the S2A and S2A_DEM datasets using the MLC technique.As with Figure 5, the classification accuracies of the S2A_DEM dataset are all higher than those of the single Sentinel-2A data, demonstrating the significance of DEM data for lithological mapping.The overall accuracy of the classification using the S2A_DEM dataset is 74.5%, which is 8.25% higher than that of the S2A dataset, as shown in Figure 7.

Lithological Discrimination Using OLI_DEM, AST_DEM and S2A_DEM Datasets
Lithological classifications of the OLI_DEM and AST_DEM datasets using the MLC method are shown in Figure 8a,b, respectively.In comparison with the lithological results of the S2A_DEM dataset (Figure 3e) and the geological map (Figure 1c), Sentinel-2A shows a better capability for lithological discrimination than ASTER and OLI, especially in identifying basaltic andesite and tuff, and argillaceous matrix.

Lithological Discrimination Using OLI_DEM, AST_DEM and S2A_DEM Datasets
Lithological classifications of the OLI_DEM and AST_DEM datasets using the MLC method are shown in Figure 8a,b, respectively.In comparison with the lithological results of the S2A_DEM dataset (Figure 3e) and the geological map (Figure 1c), Sentinel-2A shows a better capability for lithological discrimination than ASTER and OLI, especially in identifying basaltic andesite and tuff, and argillaceous matrix.As displayed in Figure 9, the OLI_DEM dataset shows the worst capability for lithological classification among the three datasets.The overall accuracies of the OLI_DEM dataset (69.42%) is also lower than those of the AST_DEM (72.0%) and S2A_DEM (74.5%), as shown in Figure 7.As illustrated in Figures 6-8, the multispectral Sentinel-2A shows better capability for lithological classification than OLI and ASTER in the Shibanjing ophiolite complex.As displayed in Figure 9, the OLI_DEM dataset shows the worst capability for lithological classification among the three datasets.The overall accuracies of the OLI_DEM dataset (69.42%) is also lower than those of the AST_DEM (72.0%) and S2A_DEM (74.5%), as shown in Figure 7.As illustrated in Figures 6-8, the multispectral Sentinel-2A shows better capability for lithological classification than OLI and ASTER in the Shibanjing ophiolite complex.
As displayed in Figure 9, the OLI_DEM dataset shows the worst capability for lithological classification among the three datasets.The overall accuracies of the OLI_DEM dataset (69.42%) is also lower than those of the AST_DEM (72.0%) and S2A_DEM (74.5%), as shown in Figure 7.As illustrated in Figures 6-8, the multispectral Sentinel-2A shows better capability for lithological classification than OLI and ASTER in the Shibanjing ophiolite complex.

Lithological Discrimination Using OLI_AST_DEM and S2A_AST_DEM Datasets
The OLI_AST_DEM and S2A_AST_DEM datasets were used for classification using the MLC method, and the results are shown in Figure 10.The results using the S2A_AST_DEM dataset, which classified the granodiorite, quartz diorite (Figure 10c-e), and basalt (Figure 10f-h) better, show a higher consistency with the geological map than the results obtained from the OLI_AST_DEM dataset.In addition, the classification results obtained from dataset S2A_AST_DEM show clearer boundaries between different rocks because of the higher spatial resolution of dataset S2A_AST_DEM.Moreover, there are fewer misclassified cases in the classification map of the S2A_AST_DEM dataset.The alluvium deposits were ideally recognized in Figure 10b.
The average accuracies of lithological classification by the data combinations of OLI + ASTER + DEM and Sentinel-2A + ASTER + DEM using the MLC method are displayed in Figure 11.Most of the lithological units obtained from the data combination of Sentinel-2A + ASTER + DEM have higher accuracies than those extracted from the OLI + ASTER + DEM dataset.The overall accuracy of lithological classification using the S2A_AST_DEM dataset is 77.83%, which is 1.83% higher than that of the OLI_AST_DEM dataset, and higher than the accuracies without incorporating ASTER dataset, as shown in Figure 7.The comparison between different datasets using the MLC method indicates that the data combination of Sentinel-2A + ASTER + DEM shows the greatest potential to classify the lithological units in the Shibanjing ophiolite complex.method, and the results are shown in Figure 10.The results using the S2A_AST_DEM dataset, which classified the granodiorite, quartz diorite (Figure 10c-e), and basalt (Figure 10f-h) better, show a higher consistency with the geological map than the results obtained from the OLI_AST_DEM dataset.In addition, the classification results obtained from dataset S2A_AST_DEM show clearer boundaries between different rocks because of the higher spatial resolution of dataset S2A_AST_DEM.Moreover, there are fewer misclassified cases in the classification map of the S2A_AST_DEM dataset.The alluvium deposits were ideally recognized in Figure 10b.The average accuracies of lithological classification by the data combinations of OLI + ASTER + DEM and Sentinel-2A + ASTER + DEM using the MLC method are displayed in Figure 11.Most of the lithological units obtained from the data combination of Sentinel-2A + ASTER + DEM have higher accuracies than those extracted from the OLI + ASTER + DEM dataset.The overall accuracy of lithological classification using the S2A_AST_DEM dataset is 77.83%, which is 1.83% higher than that of the OLI_AST_DEM dataset, and higher than the accuracies without incorporating ASTER dataset, as shown in Figure 7.The comparison between different datasets using the MLC method indicates that the data combination of Sentinel-2A + ASTER + DEM shows the greatest potential to classify the lithological units in the Shibanjing ophiolite complex.

Discussion
The machine learning methods are evaluated for lithological classification in the barely vegetation covered and well-exposed ophiolite complex using the lithological mapping results.It is shown that the MLC method yielded similar results to the SVM technique, better than the supervised k-NN, ANN, and RFC methods, which matches the geological map well in most of the lithological units in the study area, except for the quartz vein.However, although the supervised technique k-

Discussion
The machine learning methods are evaluated for lithological classification in the barely vegetation covered and well-exposed ophiolite complex using the lithological mapping results.It is shown that the MLC method yielded similar results to the SVM technique, better than the supervised k-NN, ANN, and RFC methods, which matches the geological map well in most of the lithological units in the study area, except for the quartz vein.However, although the supervised technique k-NN has been one of the foremost techniques for classification in many fields [41,58], the lithological classification using the k-NN method in the Shibanjing ophiolite complex shows somewhat lower consistency with the geological map, especially in quartz diorite, basic, and ultrabasic rock, ultramafic rock, and limestone.The classified results using the k-NN method were greatly affected by the terrain, making it difficult to differentiate alluvium deposits from other lithological units.SVM has a lower requirement for the distribution of data than MLC, whereas the input data of MLC need to satisfy the assumption of normality.With the normally distributed samples in the Shibanjing ophiolite complex, the classification using MLC shows a higher overall accuracy than that of SVM.
In addition, an analysis of variance was conducted to compare the effect of machine learning methods for lithological classification using the non-parametric McNemar test.The McNemar test is a chi-squared statistic with 1 degree of freedom, which is used on nominal data to determine whether the row and column marginal frequencies are equal [59].The McNemar test is compared to the critical value at a particular significance level, e.g., 5%, to test if there are significant differences in the expected values of two results [60].The MLC method shows statistically significant differences with respect to k-NN, ANN, and RFC techniques at the 95% confidence interval, as shown in Table 5. MLC and SVM have the similar performance for lithological classification, and there was no substantial difference in their performances at the 95% confidence interval.Therefore, the MLC and SVM methods were proven to be more capable for lithological mapping in the Shibanjing ophiolite complex than the other three techniques.However, the optimal classifier for image classification may be different in other study areas.Different rock units have different susceptibilities to weathering, because of their diverse mineralogical composition, texture, age and rate of erosion, which lead to different topographic expressions in the field [22].For example, alluvium deposits generally form flatlands, whereas granitic plutons form high hills, because of their high resistance to erosion.The DEM data, obtained from satellite, could be used to quantify the topographical expressions of rock units.In this research, the overall accuracy of mapping results obtained from the S2A_DEM dataset is 74.5%, approximately 8% higher than the results obtained from the Sentinel-2A image alone.The dramatic improvements of classification obtained by incorporating DEM data are mainly focused on schist and metasandstone with marble (18.11%), quartz diorite (18.95%), dolomite (25%), and argillite (16.94%), as displayed in Figures 4 and 5.
Figure 12 reveals the terrain characteristics of the study area.From a geomorphological point of view, the alluvium deposits show a significant difference from the surrounding lithological units, which makes a great improvement in classification, as displayed in Figures 4 and 5. Similarly, the significantly increased accuracies of quartz diorite, basalt and marble are all due to the input of DEM data.Because of the small outcrops of dolomite, argillite and limestone in the study area, a small improvement in classification could lead to a significant increase in accuracy.Therefore, the DEM data play an essential role in lithological classification as it quantifies topographic features, which is a good indicator of rock types.
Figure 12 reveals the terrain characteristics of the study area.From a geomorphological point of view, the alluvium deposits show a significant difference from the surrounding lithological units, which makes a great improvement in classification, as displayed in Figures 4 and 5. Similarly, the significantly increased accuracies of quartz diorite, basalt and marble are all due to the input of DEM data.Because of the small outcrops of dolomite, argillite and limestone in the study area, a small improvement in classification could lead to a significant increase in accuracy.Therefore, the DEM data play an essential role in lithological classification as it quantifies topographic features, which is a good indicator of rock types.By comparing the three different datasets OLI_DEM, AST_DEM, and S2A_DEM for lithological classification, the capability of Sentinel-2A for lithological mapping in the semi-arid and arid regions, such as the Shibanjing ophiolite complex, was evaluated.As part of the data continuity of Landsat and SPOT, Sentinel-2A has higher spatial and spectral resolution in VNIR range than ASTER and OLI, especially in the ferric iron (Fe 3+ ) diagnostic absorption features, approximately 0.65 and 0.87 By comparing the three different datasets OLI_DEM, AST_DEM, and S2A_DEM for lithological classification, the capability of Sentinel-2A for lithological mapping in the semi-arid and arid regions, such as the Shibanjing ophiolite complex, was evaluated.As part of the data continuity of Landsat and SPOT, Sentinel-2A has higher spatial and spectral resolution in VNIR range than ASTER and OLI, especially in the ferric iron (Fe 3+ ) diagnostic absorption features, approximately 0.65 and 0.87 µm [61], which makes it more sensitive for ultramafic rocks and andesite in the present study.As shown in the classification accuracies in Figure 8, the accuracy of ultramafic rocks in the ophiolite complex obtained from the S2A_DEM dataset is approximately 7.9% and 8.5% higher than those of OLI and ASTER, respectively.Moreover, due to the lower spectral resolution of OLI compared to Sentinel-2A in the VNIR and SWIR range, the fifteen lithological accuracies of the OLI_DEM dataset are all lower than those of the S2A_DEM dataset.
However, with the higher spatial resolution of ASTER compared to Sentinel-2A in the SWIR range, some of the classification accuracies of lithological units obtained from the AST_DEM dataset are higher than those extracted from the S2A_DEM dataset.For example, with the diagnostic spectral features of calcite in the ASTER band 8 [62,63], the classified marble using ASTER shows a higher accuracy than that of Sentinel-2A.In addition, with the significant SWIR absorption features of silicate minerals in ASTER bandpasses [64], ASTER has a better capability for classifying sandstones and quartz diorite than Sentinel-2A, as illustrated in Figures 4 and 5.
Through the comparison of lithological classification using different multispectral data, it can be concluded that, due to its higher bandpasses in the VNIR and SWIR range, the multispectral Sentinel-2A has a greater potential for lithological mapping than OLI.In addition, because of their respective higher spectral resolutions in VNIR and SWIR, Sentinel-2A shows a better capability for mapping ferric rocks, whereas ASTER has a better ability to classify rocks that contain carbonate, hydrate, and hydroxyl-bearing sulfate, silicate, and other minerals that possess SWIR absorption features.In the present research, due to the large outcrop of alluvium deposits and ophiolitic rocks, the overall accuracy of the S2A_DEM dataset is 74.5%, 2.5% higher than that of the AST_DEM dataset.

Conclusions
Through a comprehensive comparison, the machine learning method MLC was chosen as the classifier for the assessment of multispectral Sentinel-2A data for lithological mapping in the Shibanjing ophiolite complex in the Beishan orogenic belt, Inner Mongolia, China.Various data combinations of OLI, ASTER, Sentinel-2A, and DEM data were compared to evaluate the capability of Sentinel-2A for lithological classification using the MLC method.According to the comparisons and discussions, the following points are concluded: (1) The MLC and SVM machine learning methods are equally applicable for lithological classification in the Shibanjing ophiolite complex and better than the techniques of k-NN, ANN, and RFC using Sentinel-2A data.(2) Multispectral Sentinel-2A data have greater potential for lithological classification than ASTER and OLI in this research, and the DEM data also play a significant role in lithological mapping.(3) OLI could be substituted by Sentinel-2A, which when combined with ASTER, exhibits better performance in lithological classification in semi-arid and arid regions, such as the Shibanjing ophiolite complex.
Lithological classification is an important application in geological remote sensing, and the multispectral data Sentinel-2A is demonstrated to be efficient for lithological discrimination.In our future research, we plan to use image processing methods (e.g., PCA, MNF) and morphometric variables (e.g., slope, curvature, roughness) to improve the lithological classification.In addition, to expand the applicability of classification in geological investigation, we plan to study the requirements for accurate lithological classification using optical imagery in covered regions (e.g., vegetation or regolith covers).

Figure 2 .
Figure 2. Flowchart of the lithological classification process.

Figure 4 .
Figure 4. (a) The lithological classification accuracies of each class; (b) overall accuracies; and (c) Kappa coefficient of the S2A_DEM dataset using different machine learning methods.

Figure 4 .
Figure 4. (a) The lithological classification accuracies of each class; (b) overall accuracies; and (c) Kappa coefficient of the S2A_DEM dataset using different machine learning methods.

Figure 6 .
Figure 6.The lithological classification accuracies of the S2A and S2A_DEM datasets using the MLC method.

Figure 6 .
Figure 6.The lithological classification accuracies of the S2A and S2A_DEM datasets using the MLC method.

QFigure 6 .
Figure 6.The lithological classification accuracies of the S2A and S2A_DEM datasets using the MLC method.

Figure 9 .
Figure 9.The classification accuracies of datasets of OLI + DEM and AST + DEM using the MLC method.

Figure 9 .
Figure 9.The classification accuracies of datasets of OLI + DEM and AST + DEM using the MLC method.

Figure 10 .
Figure 10.Lithological classification of three different datasets using MLC.(a) The classification of OLI_AST_DEM; (b) the classification of S2A_AST_DEM; (c) partial magnification of quartz diorite in (a); (d) partial magnification of quartz diorite in (b); (e) partial magnification of quartz diorite in geological map; (f) partial magnification of basalt in (a); (g) partial magnification of basalt in (b); (h) partial magnification of basalt in geological map.

Figure 10 .
Figure 10.Lithological classification of three different datasets using MLC.(a) The classification of OLI_AST_DEM; (b) the classification of S2A_AST_DEM; (c) partial magnification of quartz diorite in (a); (d) partial magnification of quartz diorite in (b); (e) partial magnification of quartz diorite in geological map; (f) partial magnification of basalt in (a); (g) partial magnification of basalt in (b); (h) partial magnification of basalt in geological map.Remote Sens. 2018, 10, x FOR PEER REVIEW 15 of 22

Figure 11 .
Figure 11.The classification accuracies of the OLI + AST + DEM and S2A + AST + DEM datasets using the MLC method.

Figure 11 .
Figure 11.The classification accuracies of the OLI + AST + DEM and S2A + AST + DEM datasets using the MLC method.

Figure 12 .
Figure 12.The topographic map of the Shibanjing ophiolite obtained from DEM.

Figure 12 .
Figure 12.The topographic map of the Shibanjing ophiolite obtained from DEM.

Table 2 .
Multispectral image used in this study.

Table 3 .
Datasets for lithological mapping in this study.

Table 3 .
Datasets for lithological mapping in this study.
Figure 2. Flowchart of the lithological classification process.

Table 4 .
The lithological unit areas, training and testing samples for lithological units.

Table 7 .
Results from lithological classification using different remote sensing datasets and classifiers from the last few years.Grey-level co-occurrence matrix; 2 Multispectral; 3 Knowledge-based systems; 4 Hyperspectral;5Airborne Thematic Mapper;6Object-based image analysis; 7 Jeffries-Matusita; 8 Hyperspectral and Infrared Imager; 9 Genetic algorithm coupled with the Spectral Angle Mapper. 1