Lithological Mapping Research Based on Feature Selection Model of ReliefF-RF

Abstract

Multispectral satellite data allow experts to discriminate rock units based on their spectral signature characteristics. Here, Sentinel-2, ASTER and the Landsat-8 Operational Land Imager (OLI) were assessed for lithological mapping by using a random forest (RF) classifier for a study area located in Xitieshan, Northwest China. The classification accuracy of Sentinel-2 was 60.71%, which was 5.24% and 4.77% higher than the accuracies for ASTER and the Landsat-8 OLI, respectively. Three image enhancement techniques, namely, principal component analysis (PCA), independent component analysis (ICA) and minimum noise fraction (MNF), were used with grey-level cooccurrence matrices (GLCMs) to increase the quality of the input datasets. The ICA could discriminate between rock unit datasets better than the other approaches. In contrast, GLCM performed poorly when used independently. The overall classification accuracies were 60.71%, 62.63%, 64.34%, 65.21% and 58.87% for the 10 bands of Sentinel-2, PCA, MNF, ICA and GLCM, respectively. Then, five datasets were combined as a single group and applied in RF classification. Sentinel-2 obtained an overall accuracy of 73.96% and performed better than the other single-dataset approaches used in this study. Furthermore, the classification result of RF was achieved better performance than that of the support vector machine algorithm (SVM). During feature selection processing, ReliefF, the most successful pre-processing algorithm, was employed to preliminarily perform feature screening. Then, the optimal dataset was selected on the basis of the importance ranking of RF. A total of 20 more important predictors were selected from 114 original features using the ReliefF-RF model. These predictors were used in the lithological mapping, and an overall accuracy of 77.63% was reached.

1. Introduction

Lithological mapping is an important application of remote sensing technology in geology. Remote sensing data contain rich spectral and spatial features that can provide various information to produce new geological maps or update existing geological maps, especially inaccessible areas. Due to its lower cost and higher accessibility, moderate-resolution remote sensing multispectral data, such as Landsat-5 Thematic Mapper (TM), Landsat-7 Enhanced Thematic Mapper Plus (ETM+), Landsat-8 Operational Land Imager (OLI) and ASTER, have been successfully used in geological surveys and mineral exploration [1,2,3]. Compared with Landsat and ASTER, Sentinel-2 has rarely been applied in lithological mapping. As one of the most important data continuity and enhancement projects for the earth observation, more case studies are necessary to better demonstrate the capabilities and potential of Sentinel-2 for lithological mapping.

Sentinel-2A/B satellites, an extension of the Earth observation missions of the Landsat and SPOT satellites, were launched with a multispectral imaging instrument by the European Space Agency via the Copernicus Programme in 2015 and 2017, respectively. The mission ensures continuity of the SPOT and Landsat missions and provides operational products such as land cover maps, land change detection maps and geochemical/physical variables [4]. The subtle differences of various ground objects can be distinguished by the 13 narrow bands (4 bands at 10 m, 6 bands at 20 m and 3 bands at 60 m) of Sentinel-2 data, which span from the visible and near-infrared bands to the shortwave infrared band [5]. Furthermore, Sentinel-2 has a short revisit time (every five days for two satellites) and Sentinel-2 scenes in the data archive are made available to all users at no cost. In the existing research, Ge et al. [6] utilised a Sentinel-2A imager and ASTER to discriminate lithological units in the ophiolite complex of Shibanjing in Inner Mongolia, China. The results shown that Sentinel-2 data outperformed ASTER in lithological mapping in the Shibanjing ophiolite complex. In addition, until April 2008, the SWIR subsystem of ASTER had a failure of the detector cooling apparatus (not the Stirling cycle cooler), so no further SWIR data could be captured [7]. Adiri et al. [8] provided a comprehensive review of the use of Landsat-8 and Sentinel-2 in mineral exploration. By comparison and analysis, the Sentinel-2 data have shown potential, especially in mapping iron absorption features, due to the similar or even higher spectral characteristics than the Landsat series and SPOT in the VNIR part. It is shown fully that, based on the spatial and spectral characteristics proposed for Sentinel-2 and its similarity to Landsat ETM, SPOT and to lesser extent ASTER, there seems to be great potential of developing products that can be used for the geology community [9].

Furthermore, some researchers have used commonly processing methods such as principal component analysis (PCA), independent component analysis (ICA), minimum noise fraction (MNF) and multi-source data with Sentinel-2 to improve the lithological classification accuracy and achieved better performance [10,11,12]. For example, Bentahar et al. [13] combined PCA, MNF and ICA to enhance the lithological mapping accuracy. For Sentinel-2A, the lithological mapping accuracy reached 93.93%. Shebl et al. [14] integrated gamma ray spectrometric data with Sentinel-2 data, leading to monitoring minute variations among 13 classes and discrimination of closely related spectral signatures. The results shown that blending the total count band (K + Th + U) with Sentinel 2 data outstandingly boosted the classification accuracy by 7.77%.

Texture is one of the major visual properties of rock units. Spectral features (reflectance) capturing textural features of lithologies that may be unique as a result of differences in physical weathering erosion and drainage, can provide useful supporting information to distinguish rock types [15,16]. Among many textural methods, a grey-level cooccurrence matrix (GLCM) is the most popular technique for the computation of second-order textural features [17]. However, spectral and spatial features derived from Sentinel-2 data are rarely integrated into lithological mapping in current study. These scenarios constitute the motivation for this study, which explores an efficient way to integrate the spectral and spatial data of Sentinel-2 in lithological mapping.

To effectively derive reliable lithological information from remote sensing data, appropriate classification techniques are essential. Machine learning classifiers could be widely divided into parametric (preferred with normally distributed data) or non-parametric types. The maximum likelihood classifier (MLC) is a typical parametric classifier. Over the decades, it has been widely applied in lithological mapping [18]. MLC labels the pixels in an image according to user-defined classes based on posterior probabilities computed from the statistics of training data. MLC is an optimal classifier if the training data is normally distributed. However, obtaining normally distributed training data is often difficult in practice [19]. In the last decades, various non-parametric classification algorithms have been developed in order to obtain more accurate and more reliable classifications. Among them, RF and SVM achieved better performance and were widely used in lithological mapping [20,21]. The SVM has been increasingly used in the remote-sensing community due to its ability to generate good classification results even with limited and spectral-mixed training data, a common limitation for traditional digital image classification [22]. The RF classifier has been gaining increasing attention because it has fast computation, few parameter requirements, few statistical assumptions on training data, less sensitivity to noise or over-fitting, and the capability to determine variable importance [23,24]. In the supervised classification field survey, numerical and geostatistical analyses with the help of GIS were usually applied to determine the boundary of different rock units, select the training and test datasets, and evaluate the classification accuracy via the producer accuracy, user accuracy, average accuracy (the average of user and producer accuracy), overall accuracy and Kappa coefficient.

Feature selection can be considered as an important pre-processing step for classification, potentially improving performance and efficiency of the classifier [25]. With the increase in feature dimensions, the lithological mapping model will likely become more complex and the classification accuracy will be influenced by data redundancy. Efficiency, accuracy and feature selection can be jointly utilised during classification to balance the complexity of models. The existing feature selection methods can be divided into filters, wrappers and embeddings [26,27,28]. Filter methods such as relief and reliefF, first selects the subset of all features that is independent of the learning processing. These methods have been widely adopted in feature selection owing to their simplicity and efficiency [29]; however, the features selected by this filtering method may not achieve good performance in subsequent classification tasks. Wrapper methods, such as feature elimination [30], simulated annealing [31] and genetic algorithm (GA) [32], directly take the performance of the model as an evaluation criterion of the feature subset, therefore, they are better than the filter methods. However, the wrapper approach has high computational costs due to the process of multiple executions [33]. Embedded methods, such as RF [34], perform feature selection in the process of training and are usually specific to given learning machines. Therefore, the search for an optimal subset of features is built into the classifier construction and can be seen as a search in the combined space of feature subsets and hypotheses. This approach is able to capture dependencies at a lower computational cost than wrappers [35] but is highly affected by the biases of a given learning algorithm [36]. Hence, in this study, a ReliefF-RF combination model was used in feature selection. Firstly, ReliefF was employed to perform feature screening, eliminate low-correlation features and reduce the dimension of the feature space quickly and efficiently. Then, the optimal dataset was selected on the basis of the importance ranking of RF. The method applied in this study combined the advantages of the filtering and embedding approaches.

Above all, the aim of this research is to deep mine the spectral and spatial characteristics of Sentinel-2 and determine its optimal feature datasets for lithological mapping by mean of GIS and remote sensing techniques. Lithological mapping was performed in an area in Xitieshan, which is located in the central segment of North Qaidam; this area is also an ultrahigh pressure (UHP) metamorphic belt of Northwest China. Three moderate-resolution satellite datasets, namely, Landsat-8 OLI, ASTER and Sentinel-2, were used to assess the capability of Sentinel-2 for the lithological mapping of the study area. Furthermore, PCA, MNF, ICA and GLCM were adopted to improve the lithological mapping accuracy of different rock units. During feature selection processing, the ReliefF-RF model was applied to select the optimal features for lithological mapping.

2. Geological Background of Study Area

The northern margin of the Qaidam Basin is located in the northeast of the Qinghai–Tibet Plateau, Northwest China. It consists of the following second-order structural units: (I) the Qaidam platform block, (II) the orogenic belt in South Qilian and (III) the north margin active belt [37]. A prominent UHP metamorphic belt can be observed in the northern margin of Qaidam, extending approximately 400 km in the northwestern–western direction and consisting of UHP metamorphic terranes, such as the Yuka, Lvliangshan, Xitieshan and Dulan terranes, from northwest to southeast [38,39,40]. The UHP metamorphic rocks in these terranes are mainly granitic and argillaceous gneiss, and some of them contain eclogite lenses and various terranes (Yukadi, Xitieshan and Dulan terranes) or garnet peridotite (Lvliangshan terrane).

The study area is located in Xitieshan in the North Qaidam Basin (37°20′–37°40′ N, 95°30′–95°45′ E). It is bounded by the Ulan Yuka Fault in the northeast and the fault zone of the North Qaidam Basin. The study area is adjacent to Oulongbulak Mountain and connected to the Cenozoic depression of the Qaidam Basin (Figure 1). In this area, daytime and nighttime are characterised by high-intensity ultraviolet radiation and large temperature differences. The climate is dry in summer and cold in winter. The study area also has minimal vegetation and bare rock, hence its suitability for lithological mapping research.

The study area includes 14 stratigraphic rock units (Figure 2 and

Table 1. Lithological classification table of study area.

Category Number	Stratigraphic Chronology	Lithology
1	Qpl4	Modern Pvoluvial–alluvial deposit block, gravel and sandy soil
2	Q3pl	Piedmont and terrace Pvoluvial gravel layer
3	Q1q	The upper loess contains gravel sandstone and marl lens, and the lower part is a thick brown conglomerate
4	N2y	Interbedding of yellowish green thick gravelly sandstone and greyish purple medium thick argillaceous siltstone
5	E3g	Yellowish green block bedded rock, lithic feldspathic quartz sandstone intercalated with purplish red siltstone and light grey medium thick marl
6	E1–21	Reddish- and yellowish-brown block layered gravels with purple calcareous feldspar sandstone
7	Kqn	Purplish red thick gravel with pebbly sandstone or greyish feldspathic quartz sandstone
8	Z2dk	Interbedding of biotite plagioclase gneiss and marble
9	Z1bdk	Interbedding of plagioclase amphibole and migmatised biotite plagioclase gneiss
10	Z1adk	Migmatisation, biotite plagioclase gneiss with garnet sillimanite mica schist
11	J3h	Reddish thick fine sandstone and clay rock interbedded with blue-grey siltstone belt
12	J1–2	Interbedding of greyish white gravelly sandstone and black shale intercalated with coal seam
13	Cb3zh	Yellowish grey dolomite and purple sandstone interbedded, with boulder at the bottom
14	Cc3zh	Light grey thick-banded crystalline limestone intercalated with phyllite and calcareous slate

). The rocks belong to the medium- and high-grade metamorphic rock series of the Daken Daban Group with lithologies of migmatite, gneiss, quartz schist, granulite with marble and plagioclase amphibolite schist. The sedimentary rocks in this area belong to the following five formations: (1) Chengqianggou Formation of the Early Carboniferous period, with the lithology of grey dark grey limestone and dolomite; (2) Early–Middle Jurassic Dameigou Formation, which is a set of coal-bearing clastic rock formations (i.e., sedimentary rocks that are mainly composed of gravelly sandstone, fine sandstone, medium coarse quartz sandstone and siliceous slate with coal); (3) the lower members of the Ganchaigou Formation in the Oligocene to Miocene periods, which is mainly composed of gravelly coarse sandstone and a conglomerate with a moderately thin layer of fine sandstone, siltstone and mudstone, belonging to Piedmont accumulation facies, and upper members mainly composed of interbedded gravelly sandstone and fine sandstone, siltstone and mudstone belonging to fluvial facies; (4) the Miocene Youshashan Formation, which is mainly composed of interbedded mudstone, siltstone and thin-layer gravelly sandstone with parallel bedding (i.e., the rock colours are mainly grey, greyish yellow and brown yellow, indicating a lacustrine sedimentary environment); and (5) the Qigequan Formation in the lower Pleistocene, which is composed of grey and greyish purple conglomerate with medium thin sandstone. The five component is a Quaternary system that is mainly distributed in the piedmont plain and intermountain valley basin. Its genetic types mainly include alluvial and alluvial pvoluvial in the Late Pleistocene period and alluvial, pvoluvial–alluvial, aeolian, lacustrine and chemical deposits in the Holocene period.

Most metamorphic rocks of the amphibolite facies are widely spread amongst the abovementioned formations. The rocks were formed by regional dynamic heat flow metamorphism in the Lvliang period and belong to the Paleoproterozoic Daken Osaka Group. Their rock assemblage is mainly dark grey, grey and variegated gneiss, migmatite, marble and granulite, amongst others. The metamorphic rocks of the low greenschist facies formed by regional low-temperature dynamic metamorphism are dominated by Cambrian–Ordovician epimetamorphic rocks with schist, phyllite, slate and metasandstone lithologies. The Triassic and Pre-Triassic strata have undergone metamorphism in different stages and degrees.

3. Material and Methods

In this study, different datasets were designed to evaluate the potential application of Sentinel-2 data and determine the optimal datasets for lithological mapping. Firstly, after pre-processing (radiation correction and atmospheric corrections), georeferencing and co-registering, three multispectral data (Landsat-8 OLI, Sentinel-2 and ASTER) were employed for lithological mapping in order to compare the capability of Sentinel-2 data for lithological classification with other multispectral data. Secondly, image enhancement techniques (PCA, MNF and ICA) and texture analysis techniques (GLCMs) were applied to evaluate the best method for improving the data quality for lithological mapping. Thirdly, with the goal of improving the classification results, five datasets (the Sentinel-2, PCA, MNF, ICA and GLCM datasets) were combined as an input to the RF classifier. Finally, dataset optimisation was performed by the ReliefF-RF model to solve the data redundancy problem (Figure 3).

3.1. Pre-Processing of Data

(1) Sentinel-2

The Sentinel-2A/B sensor can obtain 13 band multispectral images with 4 bands at 10 m, 6 bands at 20 m and 3 bands at 60 m. The 13 spectral bands span from the visible and near-infrared to the shortwave infrared. The downloadable Sentinel-2B data (acquired on 9 September 2018) are distributed at the L1C (TOA reflectance) level with the same projection and the same global geodetic system as Landsat-8 data. The data are only needed to be atmospherically corrected. The sen2cor plug-in developed by the European Space Agency was used for atmospheric correction. Spectral bands 1, 9 and 10 from the Sentinel-2 data with 60 m spatial resolution were disregarded because they were designed for atmospheric correction. Six 20 m bands were resampled to a 10 m spatial resolution by the nearest-neighbour model, whereas the remaining ten bands were layer-stacked as a single file. The radiometric and atmospheric corrections of the Sentinel-2 images were used SNAP 7.0 software then the corrected Sentinel-2 images were converted to ENVI format.

(2) Landsat-8 OLI

Landsat-8 carries two sensors, the OLI and thermal infrared sensors. The former contains nine spectral bands, whereas the latter records thermal data with two bands, as shown in

Table 2. Description of Sentinel-2B, Landsat 8 and ASTER data.

Sentinel-2B			Landsat-8			ASTER
Band	Central Wavelength (nm)	Spatial Resolution (m)	Band	Central Wavelength (nm)	Spatial Resolution (m)	Band	Central Wavelength (nm)	Spatial Resolution (m)
B1	442.3	60	B1	443	30	B1	556	15
B2	492.1	10	B2	482.5		B2	661
B3	559		B3	562.5		B3N	807
B4	665		B4	655		B3B	807
B5	703.8	20	B5	865		B4	1656	30
B6	739.1		B6	1610		B5	2167
B7	779.7		B7	2200		B6	2209
B8	833	10	B8	640	15	B7	2262
B8a	864	20	B9	1375		B8	2336
B9	943.2	60	B10	10,900		B9	2400
B10	1376.9		B11	12,000		B10	8291	90
B11	1610.4	20				B11	8634
B12	2185.7					B12	9075
						B13	10,657
						B14	11,318

. In this study, one Landsat-8 OLI scene acquired on 5 December 2017 was selected for lithological discrimination with cloud-free cover in the study area. These data are distributed at the L1T (corrected terrain) level, with a Universal Transverse Mercator projection and a World Geodetic System WGS 84 datum. In geological research, bands 1 and 9 are not used because the first band is intended for recovering the properties of atmospheric aerosols, whereas the ninth band is specific to the study of cirrus [41]. Here, the VNIR and SWIR bands of Landsat-8 OLI were layer-stacked to a single file with a 30 m spatial resolution. Radiation correction and atmospheric corrections by FLAASH were applied to Landsat-8 OLI scene to eliminate the effects caused by the atmosphere using ENVI 5.3 software.

(3) ASTER

ASTER covers a broad range of spectral regions. It has 14 spectral bands, namely, 3 VINR bands with a 15 m spatial resolution, 6 SWIR bands with a 30 m spatial resolution and 5 TIR bands with a 90 m spatial resolution. Furthermore, another telescope is used for backward viewing in the near-infrared spectral band (Band 3B). The ASTER data employed in this study consisted of a cloud-free level 1T scene acquired on 3 October 2006. The image was pre-georeferenced to the UTM zone at a 46-north projection with the WGS-84 datum. Radiation and atmospheric corrections were applied using fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH) in the ENVI 5.3 software. Furthermore, cross-tracking correction was applied to ASTER to remove the effects of energy overspill from band 4 into bands 5 to 9 [42]. Six 30 m bands were resampled to a 15 m spatial resolution by the nearest-neighbour model, whereas the remaining fifteen bands were layer-stacked as a single file using ENVI 5.3 software.

Then, the pre-processing images of Landsat-8 OLI and ASTER were georeferenced and co-registered using Sentinel-2 as a spatial reference and ground control points were chosen from the 30 m DEM.

3.2. Feature Extraction

PCA is a convenient and effective multivariate statistical method for removing the redundancy and irradiance effects of satellite images [10,43]. It has been widely used with multispectral remote sensing data to identify lithologic units. Here, the first three components contained 98.69% information (

Table 3. Percentage of information contained in each PCA, MNF, ICA components.

Eigenvalue		Percentage	Eigenvalue		Percentage	Eigenvalue		Percentage
pc1	4,485,100.755	93.42%	MNF1	49.297124	34.19%	IC1	1	10.00%
pc2	149,135.8372	3.11%	MNF2	43.139567	29.71%	IC2	1	10.00%
pc3	103,718.3366	2.16%	MNF3	17.33235	11.86%	IC3	1	10.00%
pc4	31,448.10436	0.66%	MNF4	12.900324	8.76%	IC4	1	10.00%
pc5	14,635.23885	0.30%	MNF5	9.778669	6.60%	IC5	1	10.00%
pc6	10,136.33866	0.21%	MNF6	3.930291	2.63%	IC6	1	10.00%
pc7	2356.034608	0.05%	MNF7	2.299615	1.53%	IC7	1	10.00%
pc8	1785.780308	0.04%	MNF8	2.100346	1.39%	IC8	1	10.00%
pc9	1383.225922	0.03%	MNF9	1.75929	1.16%	IC9	1	10.00%
pc10	1199.32786	0.02%	MNF10	1.649774	1.08%	IC10	1	10.00%

). The higher PC components revealed substantial information related to the distribution of lithological units. Thus, the first seven PCs were selected in this study.

The MNF proposed by Green et al. [44] is a well-known data dimension and noise reduction technique. It consists of two cascaded PCA rotations. The first PC rotation performs noise whitening and results in transformed data in which the noise has a unit variance but no band-to-band correlation. The second rotation uses PCs derived from the original image data after they have been whitened by the first rotation and rescaled by noise standard deviation. Similar to the PCA output bands, the first MNF band contains the highest variance, and the second MNF band contains the second-highest variance [45,46]. As the last three components that were included are strongly affected by noise, only the first seven MNF components were selected for lithological mapping (Table 3).

ICA is used on multispectral or hyperspectral datasets to transform a set of mixed, random signals into mutually independent components. It is a statistical method for transforming an observed multidimensional random vector into components that are statistically as independent from each other as possible. Compared with PCA and MNF, the ICA transform uses higher-order statistics for signal separation and feature extraction [47,48]. Dimension reduction is not recommended for anomaly or target detection analyses, as those signals occupy only a small portion of the image and could be buried in noisy bands. Moreover, a feature of interest occupies only a small portion of all pixels, which makes an insignificant contribution to the covariance matrix [12]. Therefore, all 10 ICA components were reserved (Table 3).

Texture is the comprehensive product of image feature shape, size, pattern and hue. It plays an important role in lithological mapping and spectral reflectance and emission properties. In recent research [49,50,51], GLCM is considered the most popular technique for extracting texture features. Here, eight texture measures (mean, variance, homogeneity, contrast, difference, entropy, second moment and correlation) extracted by GLCM, with five window sizes of 9 × 9, were applied in this study.

3.3. Sampling

The visually interpreted lithological map (DD20190536) was used for the training and validation sample selection shown in Figure 2b. The lithological mapping was updated based on the previous geological map at a scale of 1:200,000 and the field investigation in 2018 (Figure 4).

The study area is an exposed bedrock area, which means that vegetation has a negligible effect on the processing of sample selection. The former geological maps were used to carefully select the training samples corresponding to 14 lithological units and avoid the boundaries of different lithological units, as well as test samples. The test samples could differ considerably from the training samples. In order to avoid underfitting, the number of samples in each category should be larger than the total number of classifications. The sample data featured 33,000 pixels and represented approximately 3.6% of the study area image from Sentinel-2 (

Table 4. The training samples and testing samples of each lithological unit.

Stratigraphic Chronology	Area (km2)	Training Samples	Test Samples
Qpl4	3.79	1711	567
Q3pl	24.14	2945	949
Q1q	4.63	1239	357
N2y	1.72	1447	175
E3g	19.56	3227	1337
E1–21	6.73	1718	620
Kqn	7.84	2348	876
Z2dk	1.68	1099	384
Z1bdk	4.95	2331	778
Z1adk	5.48	1078	376
J3h	4.05	1360	519
J1–2	6.80	1480	629
Cb3zh	0.97	1212	478
Cc3zh	6.78	1345	415

).

3.4. Feature Selection and Lithological Mapping Method

(1) ReliefF

Relief algorithms are efficient filtering feature selection algorithms; they were proposed by Kira and Rendell [52] for binary classification problems in 1992. Furthermore, relief algorithms are widely used because of their relative simplicity, high operation efficiency and satisfactory results. However, only two classes of data can be processed. Kononeill [53] analysed and extended the Relief algorithm to the ReliefF algorithm in 1994 to address noise, incompletion and multiclass problems. Instead of using a single nearest neighbour, ReliefF averages the contribution of k-nearest neighbours from the same and opposite class of each sampled instance to smooth the influence of noise in the data. The selection of the number of nearest neighbours is the primary difference between Relief and ReliefF; this process ensures greater robustness of the algorithm with regard to noise. A user-defined parameter k controls the locality of the estimates. This parameter is set heuristically; for most purposes, k can be safely set to 10 [25]. The ReliefF method was performed in Matlab 2021 for preliminary feature selection.

(2) Random Forest

RF, proposed by Beriman in 2001 [54], is a popular supervised classification technique. It uses a bagging algorithm to generate numerous new training datasets and a decision tree algorithm to identify the probability of each class within each decision tree. RF obtains a class vote from each tree and classifies them according to a majority vote [23]. In order to produce forest trees, two parameters need to be set up: the number of decision trees (Ntree) and the number of variables to be selected and tested for the best split when growing the trees (Mtry). In this study, the number of trees was 500; the Mtry parameter is usually set to the square root of the number of input variables [55,56].

In RF, the ranking of feature importance is implemented as follows: a classification and a regression tree are initially built to generate the out-of-bag (OOB) sample data. On the basis of the OOB data, the RF verifies the importance of the input data and obtains the importance score of each feature, which is expressed by the mean decrease accuracy (MDA). The principle is to turn the value of a feature parameter into a random number, calculate its impact on the accuracy of the model and measure the importance of this parameter based on the MDA obtained from multiple calculations. The higher the value is, the more important the variable [57,58]. The RF classifier was utilised for lithological mapping and feature selection based on Matlab 2021.

The advantages offered by the two algorithms were combined in this study. Firstly, ReliefF was used to remove part of the noise and redundant features. This step helped to improve the quality of the datasets and reduce the complexity of the classification model. Then, the former dataset was further screened to obtain the final dataset by using the importance sorting function of the RF method.

(3) Support Vector Machine

SVM is a supervised machine learning algorithm developed by Vapnik [59]. It has the ability to define non-linear decision boundaries in high-dimensional variable space by solving a quadratic optimisation problem [60]. This method is embedded in ENVI 5.3 software. The optimal procedure aims to maximise the margin between the closest training data and the hyperplane, therefore minimising the confusion between classes. In practice, SVMs use a subset of training data (called a support vector) to partition the feature space without assumptions on the statistical distribution of the training data.

When referring to classification based on remote sensing data using SVMs, it does not usually mean a linear SVM but a nonlinear separable SVM, which requires a nonlinear kernel function to map the classified data set to a higher dimensional feature space for linear hyperplane [17]. The usual kernel types are linear, polynomial, radial basis function and sigmoidal. In the current work, the radial basis function kernel was yielded. The radial basis function is the most widely used kernel type and is versatile when the penalty parameter is at 100 [61]. Hence, a penalty parameter of 100 and gamma in kernel function of 0.05 were used in this study. In this study, the SVM was utilised as a comparison model with Relief-RF for lithological mapping based on ENVI 5.3 software.

For the purpose of this research, ArcGIS 10.2, ENVI 5.3 and Matlab 2021 were used. These three types of software were mainly used for all the calculations, spatial analysis and feature extraction, such as lithological map modification by visual interpretation and field surveys, training and test dataset selection and lithological mapping research, etc. The confusion matrix was used to compare different datasets for lithological mapping, and the producer accuracy, user accuracy, average accuracy (the average of user and producer accuracy), overall accuracy and Kappa coefficient were determined.

4. Results

4.1. Comparison between Lithological Maps of Different Multispectral Data

Figure 5 shows the classification maps of Sentinel-2, Landsat-8 OLI and ASTER and their comparison with the reference maps. The details on the lithological map of Sentinel-2 data were more pronounced than those of the Landsat-8 OLI and ASTER maps, especially for discrimination J_1–22. The continuity and integrity of this rock unit is the best of all the three types of multispectral data.

The classification accuracies of different multispectral data are shown in

Table 5. The classification accuracy and kappa coefficient of Sentinel-2, ASTER and Landsat-OIL.

Lithological Classes	Sentinel-2			ASTER			Landsat-OIL
	Overall Accuracy: 60.71% Kappa Coefficient: 0.57			Overall Accuracy: 55.47% Kappa Coefficient: 0.52			Overall Accuracy: 55.94% Kappa Coefficient: 0.52
	Producer’s	User’s	Average	Producer’s	User’s	Average	Producer’s	User’s	Average
Qpl4	69.24	70.92	70.08	60.92	63.4	70.08	43.97	43.48	43.73
Q3pl	76.36	85.77	81.07	81.42	83.83	81.07	32.94	24.5	28.72
Q1q	45.12	40.53	42.83	31.34	47.81	42.83	57.75	50.41	54.08
N2y	21.9	100	60.95	18.05	77.08	60.95	48.39	73.33	60.86
E3g	87.97	53.98	70.98	92.33	46.16	70.98	49.75	11.3	30.53
E1–21	48.87	31.76	40.32	47.33	41.4	40.32	55.32	54.35	54.84
Kqn	56.84	39.45	48.15	57.86	34.81	48.15	55.49	35.71	45.60
Z2dk	42.17	80	61.09	58.05	47.27	61.09	37.7	58.24	47.97
Z1bdk	79.76	69.57	74.67	61.7	58.39	74.67	35.42	30.08	32.75
Z1adk	58.39	69.12	63.76	36.75	57.25	63.76	48.74	51.97	50.36
J3h	49.76	74.32	62.04	53.59	72.26	62.04	37.04	58.2	47.62
J1–2	70.99	89.53	80.26	47.83	72.79	80.26	13.92	46.03	29.98
Cc3zh	66.78	64.16	65.47	32.94	73.4	65.47	28.42	40.87	34.65
Cb3zh	42.38	73.55	57.97	76.68	55.68	57.97	52.46	65.88	59.17

. Sentinel-2 data are better than the two multispectral data in terms of lithological mapping. The overall accuracy of Sentinel-2 (60.71%) is higher than that of ASTER (55.47%) and Landsat-8 OLI (55.94%). Amongst the lithological classes, the classification effect and average accuracy of Sentinel-2 presented better performance in nine lithological units, namely, Q^pl₄, N_2y, E_3g, K_qn, Z²_dk, Z^1b_dk, Z^1a_dk, J_1–2 and C^c_3zh. The classification effect and average accuracy of the Q_1q and E_1–21 of Landsat-8 OLI were better than those of Sentinel-2. The classification average accuracies of Q₃^pl, J_3h and C^b_3zh of ASTER were better than those of Sentinel-2.

4.2. Comparison between Lithological Maps of Different Optimisation Methods

The lithological mapping results of the PCA, ICA, MNF and GLCM datasets are shown in Figure 6. Only the PCA, ICA and MNF datasets yielded good results and correctly matched the visual interpretation results of the study area. Compared with the results of Sentinel-2, less noise was involved in the classification results of PCA, ICA and MNF. The classification results of the GLCM datasets had the worst performance. For the classification results of each rock unit, the four datasets had low degrees of recognition of $Q_{1q}$ and $E_{1-21}.$ Given the similar lithological composition, $Q_{1q}$ was easily misclassified with $Q_3^{pl}$ and $Q_4^{pl}$.

The classification overall accuracies and kappa coefficients of different datasets are shown in

Table 6. The classification accuracies and kappa coefficients of PCA, MNF, ICA and GLCM.

Lithological Classes	PCA			MNF			ICA			GLCM
	Overall Accuracy: 62.63% Kappa Coefficient: 0.60			Overall Accuracy: 64.34% Kappa Coefficient: 0.61			Overall Accuracy: 65.21% Kappa Coefficient: 0.62			Overall Accuracy: 58.87% Kappa Coefficient: 0.56
	Producer’s	User’s	Average	Producer’s	User’s	Average	Producer’s	User’s	Average	Producer’s	User’s	Average
Qpl4	59.26	69.89	64.58	63.90	76.20	70.05	70.78	74.87	72.83	57.60	54.37	55.99
Q3pl	75.14	85.12	80.13	83.02	85.01	84.02	78.52	85.06	81.79	83.11	86.44	84.78
Q1q	49.29	37.23	43.26	46.31	40.19	43.25	48.45	41.07	44.76	35.00	53.36	44.18
N2y	35.40	69.37	52.39	41.15	89.50	65.33	35.63	95.09	65.36	56.78	93.21	75.00
E3g	86.33	58.23	72.28	86.80	59.49	73.15	86.33	56.63	71.48	87.85	50.64	69.25
E1–21	38.05	38.37	38.21	39.48	40.59	40.04	43.88	45.39	44.64	36.03	21.34	28.69
Kqn	67.84	47.97	57.91	71.11	49.88	60.50	67.84	49.11	58.48	80.23	55.32	67.78
Z2dk	80.47	71.00	75.74	81.09	68.28	74.69	82.79	71.77	77.28	50.08	47.85	48.97
Z1bdk	60.94	70.50	65.72	55.67	63.21	59.44	66.67	70.89	68.78	64.68	86.00	75.34
Z1adk	67.24	63.24	65.24	70.23	60.32	65.28	67.13	63.83	65.48	50.92	64.30	57.61
J3h	61.38	67.33	64.36	60.77	78.07	69.42	59.44	66.08	62.76	29.06	69.77	49.42
J1–2	70.29	95.55	82.92	70.88	88.60	79.74	69.01	92.62	80.82	60.47	87.93	74.20
Cc3zh	51.70	77.82	64.76	51.23	76.44	63.84	56.73	76.02	66.38	41.75	63.30	52.53
Cb3zh	59.84	60.32	60.08	66.98	66.14	66.56	65.56	74.28	69.92	87.14	61.69	74.42

. Compared with the four optimal datasets, ICA attained the highest overall accuracy and kappa coefficient, at 65.21% and 0.62, respectively. GLCM obtained the lowest overall accuracy and kappa coefficient, with values of 58.87% and 0.56, respectively. The overall accuracies and kappa coefficients of PCA, ICA and MNF were higher than those of Sentinel-2. $Q_4^{pl}$ and $Z_{dk}^2$ had the highest average classification accuracy in the PCA datasets; $J_{1-2}$ and $C_{3zh}^c$ had the highest average classification accuracy in the MNF datasets; and $C_{3zh}^b$ and $Q_3^{pl}$ had the highest classification accuracy in the ICA datasets.

4.3. Dataset Combination and Optimisation

For each single data point, the highest classification accuracy (65.33%) was achieved using the MNF of the Sentinel-2 data (Table 6). After combining the five datasets, the classification accuracies of the SVM and RF were improved significantly (Figure 7 and

Table 7. The classification accuracies and kappa coefficients of combination and optimal datasets.

Lithological Classes	Combination Five Datasets (SVM)			Combination Five Datasets (RF)			Optimal Dataset
	Overall Accuracy: 74.68% Kappa Coefficient: 0.73			Overall Accuracy: 75.75% Kappa Coefficient: 0.74			Overall Accuracy: 77.63% Kappa Coefficient: 0.76
	Producer’s	User’s	Average	Producer’s	User’s	Average	Producer’s	User’s	Average
Qpl4	59.26	77.97	68.62	66.39	88.03	77.21	72.33	88.52	80.43
Q3pl	89.96	84.57	87.27	95.31	97.6	96.46	96.53	98.09	97.31
Q1q	58.57	64.31	61.44	71.31	69.17	70.24	63.1	59.68	61.39
N2y	68.05	87.32	77.69	55.86	98.78	77.32	59.31	87.46	73.39
E3g	84.93	65.79	75.36	86.21	60.64	73.43	85.98	64.62	75.30
E1–21	53.75	54.46	54.11	68.25	51.57	59.91	68.85	63.00	65.93
Kqn	68.42	55.77	62.10	84.44	56.99	70.72	81.05	59.74	70.40
Z2dk	87.13	73.95	80.54	55.66	91.12	73.39	80.78	82.96	81.87
Z1bdk	76.96	89.16	83.06	88.54	88.75	88.65	88.07	90.94	89.51
Z1adk	81.84	79.46	80.65	75.98	75.72	75.85	85.98	71.72	78.85
J3h	75.18	83.47	79.33	51.94	78.43	65.19	62.23	87.86	75.05
J1–2	74.97	94.82	84.90	69.47	96.27	82.87	75.67	97.59	86.63
Cc3zh	72.75	84.51	78.63	80.35	77.19	78.77	75.67	89.24	82.46
Cb3zh	93.81	68.8	81.31	95.24	84.27	89.76	77.14	73.52	75.33

). In this study, the classification result using RF achieved better performance than the that using SVM. The overall accuracy and kappa coefficient of RF were 75.75% and 0.74 respectively, 1.07 and 0.1 higher, respectively, than those of the SVM classifier. The Q₃^pl, Z^1b_dk, J_1–2 and C^b_3zh classes had the best average accuracy at 80%. The lowest average accuracy was 59.51% for E_1–21.

Figure 8 presents the weights of 114 features calculated by the ReliefF algorithm. Here, b1–b10 corresponds to the bands of Sentinel-2, b11–b17 corresponds to the PC1–PC7 components, b18–b27 corresponds to the ICA1–ICA10 components, b28–b34 corresponds to the MNF1–MNF7 components and b35–b114 corresponds to features of the mean, variance, homogeneity, contrast, difference, entropy, second moment and correlation of each spectral band. In the primary selection stage, 64 features with weight values of >0.005 were selected as the primary optimised datasets. In particular, the datasets included 1 spectral feature, 6 PC components, 8 IC components, 7 MNF components and 42 texture components.

Then, the average accuracy reduction of each feature selected by ReliefF was calculated using RF (Figure 9). The higher the MDA, the more important the variable. The MNF2 and MNF3 components were the most important features, with average precision reductions of 1212.28 and 905.12, respectively, followed by the mean of band2, the ICA3 component and the ICA2 component.

After sorting by importance, the features were added one by one to the RF model. The OOB data error decreased when features were added to the datasets. When the number of features increased to 20, the OOB error stabilised (Figure 10). Consequently, the 20 features were selected as the optimal dataset.

After optimisation, the datasets included 2 PCA components, 3 MNF components, 5 ICA components and 11 texture features. Amongst them, MNF2 and MNF3 showed the highest impact on the stability of the lithological classification model (

Table 8. The optimal datasets selected by Relief-RF model.

	Selected Features
Optimal datasets	(1) MNF2, (2) MNF3, (3) mean of band 1, (4) ICA3, (5) ICA2, (6) mean of band 2, (7) MNF4, (8) IC5, (9) mean of band 9, (10) mean of band3, (11) mean of band 4, (12) mean of band 8, (13) PCA5, (14) mean of band 6, (15) mean of band 7, (16) PC6, (17) mean of band 12, (18) mean of band 5, (19) ICA1, (20) homogeneity of band 7

).

Lithological classification was applied to only the top 20 most important features (Figure 7 and Table 7). The overall accuracy and kappa coefficient values were 77.63% and 0.76, respectively. In contrast, with pre- and post-optimisation, the average accuracies of Q^pl₄, Q₃^pl, E_3g, E_1–21, Z²_dk, Z^1b_dk, Z^1a_dk, J_3h, J_1–2 and C^c_3zh increased.

5. Discussion

This research demonstrated the performance of Sentinel-2 data for the lithological mapping of a study area in Xitieshan, Northwest China. Sentinel-2 performed better than Landsat-8 OLI and ASTER. These results are consistent with the findings in previous research [6,12]. Given the similar or even higher spatial and spectral characteristics in the VNIR part, Sentinel-2 has a greater potential for lithological mapping than Landsat-8 OLI, especially for ferric rocks [9]. In this study, the classification accuracies of Sentinel-2 for the 11 rock units were better than those of Landsat-8 OLI, especially for Q₃^pl, E_3g, Z²_dk and J_1–2. Sentinel-2 performed better in mapping, whereas ASTER performed better in classifying rocks containing carbonate, hydrate and hydroxyl-bearing sulphate, silicate and other minerals with SWIR absorption features [61], such as Q₃^pl, E_3g, K_qn, Z²_dk and C^b_3zh (Table 1).

Image enhancement techniques, such as PCA, ICA and MNF, could highlight the lithological units and improve the quality of input datasets. After MNF transformation, the geological information was mainly concentrated into two to five components. The false colour composite image with R(MNF4)–G(MNF3)–B(MNF2) allowed the boundary information of $Z_{dk}^2$, $Z_{dk}^{1b}$ and $Z_{dk}^{1a}$ to become more obvious than the true colour image of Sentinel-2 (Figure 11).

In lithological mapping, image enhancement techniques can increase the quality of the input dataset and improve classification accuracy. Amongst the techniques used in this study, the ICA datasets obtained the highest lithological mapping accuracy owing to their higher-order statistics for signal separation and feature extraction and all ICA components were preserved. For PCA and MNF, which are well-known data dimension and noise reduction techniques, the higher-order components contained more information, but the lower-order components also influenced the accuracy of lithological classification (

Table 9. Overall accuracies of different PCA and MNF component datasets classified by RF.

Image Enhancements	PCA	MNF
The first three components	35.68%	44.57%
The first five components	53.54%	59.60%
The first seven components	62.63%	64.34%
All components	64.72%	65.33%

). The lithological mapping results indicate that the first three components of PCA and MNF, which contained 98.69% and 75.76% of the information, respectively, performed poorly in the lithological classification task. After adding the first seven components, the classification accuracies of the PCA and MNF datasets significantly increased to 26.95% and 19.77%, respectively. The highest classification accuracy was achieved when all components were used in PCA and MNF. This result indicates that the rankings of PCA and MNF are insufficient in determining their importance in lithology classification. The lower-order components also contained only a few useful pieces of information for lithological mapping.

The textural dataset, when used independently, obtained the lowest classification accuracy (overall accuracy of 58.87%); however, the lithological mapping accuracy improved when other datasets were added. In a previous study [16,62], the band with the highest standard deviation was consistently selected to extract textural features. In this study, all bands from Sentinel-2 data were used to extract textural features. This approach has rarely been mentioned in previous studies. The scheme adopted in this study performed better than when combining only the textural features extracted from the highest deviation band (67.92%).

A suitable texture window size is important in lithological mapping. Figure 12 presents the change in the rock boundary, in which the window size condition was set from 3 × 3 to 11 × 11. After adjusting the texture window from 3 × 3 to 9 × 9, the boundary between $Q_{pl}^3$ and $z_{dk}^{1a}$ was enhanced. However, when the window size was increased from 9 × 9 to 11 × 11, the changes in the boundary were negligible. In terms of integrating with other datasets, the 9 × 9 datasets presented the highest overall accuracy and kappa coefficient.

The classification accuracy increased to 75.75% when the five datasets of Sentinel-2, PCA, ICA, MNF and GLCM were combined, and the improvement was approximately 15% higher than the results obtained by using Sentinel-2 only. Therefore, the more useful the information extracted from the satellite data is, the better the performance of lithological mapping. Then, the same approach was applied to Landsat-8 OLI and ASTER, and their classification accuracies improved to 70.92% and 63.02%, respectively.

In the processing, the threshold of weight values needs to be set carefully. In this study, if the threshold was larger than 0.01, then some useful information was removed from the original datasets. If the threshold was lower than 0, the redundant information could not be eliminated. Thus, the purpose of improving the feature selection efficiency could not be achieved.

Feature selection and optimisation are the basis of accurate classification. The hybrid ReliefF-RF model can be used to optimise datasets, effectively reduce redundancy and improve operation efficiency and classification accuracy. The numbers of the optimisation datasets (only 20 bands) were only one-sixth of the combination datasets, whereas the overall accuracy and kappa coefficient values were better than those of all feature combination datasets. Therefore, in the processing stage of feature selection, large amounts of information can be integrated into optimisation datasets, subsequently enhancing the total accuracy.

Some natural and human factors, such as erosion and buildings, affect the accuracy of classification, as shown in Figure 13. Currently, no effective approach can automatically eliminate this impact. The datasets can only be modified by human–computer interactive interpretation.

6. Conclusions

This study evaluated the capability of Sentinel-2 data for lithological mapping by producing and testing various integrations of the raw and derived Sentinel-2 data in terms of their role in improving the lithological mapping accuracy. By means of advanced GIS and remote sensing techniques, the distribution boundaries of rock units were determined. The spectral and spatial information of different rock units were achieved from remote sensing data as training and test datasets for lithological mapping. The ReliefF-RF model was employed to determine the optimal dataset for lithological classification. The lithological mapping accuracy of Sentinel-2 had been significantly improved by the optimising model in this study. The results were as follows:

(1) The classification accuracy of Sentinel-2 (60.71%) was 4.77% higher than that of ASTER and 5.24% higher than that of Landsat-8 OLI. Sentinel-2 performed better than ASTER and Landsat-8 OLI in lithological mapping.

(2) The selected image enhancement methods, namely, PCA, MNF and ICA, could increase the quality of input datasets and improve the classification accuracy. The highest classification accuracy of 65.21% was achieved using ICA. Textural features, when used independently, failed to improve the classification accuracy (i.e., lowest classification accuracy of 58.87%). After integrating the raw data of Sentinel-2, the overall accuracy improved significantly (75.75%). In addition, the classification result of RF achieved better performance than that of SVM.

(3) Feature selection is an important pre-processing step for dimensionality reduction in lithological mapping. In this study, the ReliefF-RF model was used to optimise datasets; it effectively reduced redundancy and improved operation efficiency and classification accuracy. The use of 20 variables selected by the ReliefF-RF model resulted in an overall accuracy of 77.63%, which was 1.88% higher than the combination dataset of 114 variables.

The problem of data redundancy caused by the increase in feature dimensions still needs to be explored. In future work, multisource remote sensing data can be used in lithological mapping to explore a more efficient and accurate lithological method, especially in vegetation covered areas.