1.1.

Information Fusion

This is a common method for extracting lithology from remotely sensed images, which display the characteristics of different types of rocks and minerals. In recent years, Abarca-Hernandez and Chica-Olmo,⁸ Ninomiya et al.,⁹ Fu et al.,¹⁰ Zhou et al.,¹¹ Fatima et al.,¹² Li et al.,¹³ and other researchers have used a variety of remotely sensed imagery for lithologic classification. For example, Rowan and Mars¹⁴ applied the Radarsat data and ASTER data for effective identification of lithologic information. Chica-Olmo et al.¹⁵ combined the SPOT data and Landsat TM data with GIS space analysis to interpret the lithologic information.

1.2.

Texture Extraction

Texture information is very important in identifying the types of rocks and minerals; therefore, the texture extraction method is widely used in the study of lithologic classification. For example, Arivazhagan and Ganesan¹⁶ used wavelet statistical features for texture analysis. They found that the combination of wavelet statistical features and co-occurrence features could better extract the lithology. Chica-Olmo and Abarca-Hernandez¹⁷ proposed the application of statistical functions representing the texture to improve the classification accuracy of the lithology. In recent years, many new methods have been proposed and widely used for lithologic classification, such as spectral angle mapping¹⁸ and MNF transform and related band absorption depth analysis.¹⁹

1.3.

Pattern Recognition

Methods such as wavelet analysis, neural network, expert knowledge system, pattern recognition, and decision tree classification have been widely applied to lithologic classification. For example, Ishikawa and Gulick²⁰ and Mobasheri and Ghamary-Asl²¹ presented the decision framework of an expert system for lithologic classification using hyperspectral images and achieved a classification accuracy of 96%. In addition, many other researchers such as Wilkinson²² employed the spatial structures in combination with the spectral band to perform the classification.

Despite considerable effort in this field, there remains a shortage of lithologic classification methods. For example, the information fusion method is able to take advantage of different types of data sets, but data collection is expensive. Therefore, it is not suitable for lithologic identification of a large area. The texture analysis method is capable of effectively extracting lithologic information, but it is labor-intensive and has low effectiveness. The pattern recognition method is based on the results of initial classification and identification of rocks. Therefore, this paper takes advantage of the rock spectral information and prior geographical map.

2.1.

Study Area

The study area is located at Guangxi, China, at a longitude and latitude of 104° 26′ through 112° 04′ E and 20° 54′ through 26° 24′ N. The study area covers $\sim 2.38\text{million}{\mathrm{km}}^2$ (Fig. 1). It is a typical karst landform with karst rocky desertification area constituting $\sim 16.0\%$ of the total karst rocky desertification area in China.^23,24 The strata in the study area are well-developed and the outcrop area is large. The three main types of rocks in the study area are sedimentary, magmatite, and marble rocks. The largest coverage is of sedimentary rocks, accounting for 90% of the study area. Because carbonate, shale, basalt, sandstone, granite, and marble rocks approximately account for more than 95% of the study area, this paper only extracts the above-mentioned six types of rocks.

Fig. 1

Schematic diagram of the study area: (a) map of China and (b) Guangxi province.

2.2.

Data Sets

The data sets used in this study contain the following data:

Fig. 2

(a) 1:200,000 geological map of Guangxi Province, China, (b) sandstone, (c) carbonate, (d) marble, (e) shale, (f) granite, and (g) basalt.

Fig. 3

Rock samples: (a) granite porphyry, (b) limestone, (c) sandstone, (d) quartz, (e) quartz porphyry, (f) basalt, and (g) diabase.

The laboratory measurements of seven types of rock samples were conducted as follows. Field Spec^®4 from ASD Inc. with wavelength ranging from 0.350 through $2.500\mu \mathrm{m}$ was used to measure the spectral curves (see Fig. 4). The Field Spec^®4 was first calibrated using a standard whiteboard. After the calibration was completed, the spectral curve was immediately measured for a minimal sample. Each type of sample was measured at least 5 times to ensure the reliability of the data. If the measured spectral curves of the samples showed strong fluctuation, the measurements were repeated until a stable spectral curve was obtained.

Fig. 4

Laboratory equipment for measuring the spectral characteristics of the rock samples.

We screened the measured results for redundant measurements and averaged the data. The spectral curves of the rock samples are shown in Fig. 5.

Fig. 5

Spectral curves of the laboratory measurement: (a) Granite porphyry, (b) limestone, (c) sandstone, (d) quartz porphyry, (e) quartz, (f) basalt, and (g) diabase.

2.3.

Data Preprocessing

2.3.1.

Repair the bad sectors

The Landsat 7 ETM+ scan line corrector suddenly failed to work on May 31, 2003, resulting in the image data being overlapped and partially gapped. Therefore, a gap filling algorithm was developed to fix the problems. The gap filling method improves the images by extracting the “good” image data from other images of the same geographic coordinates but from a different time phase, and then filling the “bad” sectors using the “good” images. The equation is given below

Eq. (1)

$${\mathrm{DN}}_{ti}=\alpha \times {\mathrm{DN}}_{si}+\beta ,$$

where $\alpha$ is the image gain calculated as $\alpha =\frac{\sum _{i=1}^N({\mathrm{DN}}_{ti}-\overline{{\mathrm{DN}}_t})-({\mathrm{DN}}_{si}-\overline{{\mathrm{DN}}_s})}{\sum {({\mathrm{DN}}_{si}-{\mathrm{DN}}_s)}^2}$, $\beta$ is the deviation calculated as $\beta =\overline{{\mathrm{DN}}_t}-a\times \overline{{\mathrm{DN}}_s}$. ${\mathrm{DN}}_t$ represents the digital number (DN) of the $t$ image frame, ${\mathrm{DN}}_{\mathrm{s}}$ represents the DN of the reference image, ${\mathrm{DN}}_{ti}$ is the brightness to be filled at the $i$’th position, and ${\mathrm{DN}}_{si}$ is the DN of the $i$’th position of the reference image. The image with gaps was repaired using the method proposed above, as depicted in Fig. 6.

Fig. 6

Landsat 7 ETM+ images: (a) before repair and (b) after repair of the defective sectors.

2.3.2.

Atmospheric and geometric corrections

After repairing the defective sectors, a second-order polynomial equation was used for geometric correction of the Landsat 7 ETM+ imagery with ENVI 5.0 software. Seventeen ground control points that were well distributed across the study area were selected. During the correction, two points with large errors were eliminated, and the bilinear interpolation method was used for gray resampling. The MODTRAN model was employed for atmospheric correction using the flash unit in the ENVI 5.0 software. MODTRAN (mid spectral resolution atmospheric radiative transfer model) is a widely accepted model for correction of atmospheric radiation such as solar shortwave radiation, as it has high accuracy and flexible radiative transfer in scattering atmosphere. Next, the images were tailored and assembled into a complete image for the entire area of Guangxi, China [see Fig. 7(a)]. The results after the atmospheric corrections are shown in Fig. 7(b).

Fig. 7

(a) The assembled remotely sensed image of the study area and (b) an enlarged window with after the atmospheric correction.

Landsat 7 ETM+ has eight bands. This paper presents the statistical data of each band and studies their relevance. The specific statistical results are shown in Table 1. From Table 1, it can be seen that the correlations between band 2 and band 3 and between band 5 and band 7 are very strong. In order to fully reflect the different bands in the remote sensing image information, this paper applies band 7, band 4, and band 1 to create color-synthesized imagery [see Fig. 7(a)].

Table 1

Relevant statistics of Landsat 7 ETM+ for the study area.

Correlation	Band 1	Band 2	Band 3	Band 4	Band 5	Band 7
Band 1	1.00	0.99	0.98	0.92	0.90	0.88
Band 2	0.99	1.00	0.99	0.92	0.91	0.90
Band 3	0.98	0.99	1.00	0.90	0.93	0.94
Band 4	0.92	0.92	0.90	1.00	0.94	0.87
Band 5	0.90	0.91	0.93	0.94	1.00	0.98
Band 7	0.88	0.90	0.94	0.87	0.98	1.00

3.1.

Analysis of Spectral Characteristics

It has been widely accepted that different ions cause different impacts on the spectrum of the rock, whereas rocks with same ion and lithology minerals have similar spectral curves. Therefore, this paper attempts to find an effective classification method for lithologic classes by analyzing the spectral characteristics of each type of rock.

The major components of sandstone are quartz and feldspar. The spectral characteristics of the minerals in sandstone from the USGS spectral library are shown in Fig. 8. The fact that different types of rocks have similar minerals presents a challenge for lithologic classification if their spectral characteristics alone are employed. For example, the spectral characteristics of sandstone in Fig. 9 are very similar to that of muscovite shown in Fig. 8, except for an absorption peak at $1.4\mu \mathrm{m}$.

Fig. 8

Spectral characteristics of minerals in sandstone (from USGS_MIN spectral library retrieved from ENVI 5.0 software).

Fig. 9

Spectral characteristics of different types of sandstone.²⁶

Shale is a type of sedimentary rock composed of mud, which is a mix of flakes of clay with fragments of other minerals such as quartz and feldspar. Shale contains 45% to 80% ${\mathrm{SiO}}_2$.²⁷ The spectral curves of shale from the USGS spectral library are shown in Fig. 10(a). As observed from Fig. 10(a), a prominent absorption peak occurs at $1.4\mu \mathrm{m}$, but no obvious absorption and reflection peaks appear at other spectral ranges. In addition, both shale and sandstone contain quartz, which implies that both should have very similar spectral curves. In order to further understand their spectral characteristics, the spectral curve of shale was measured and the result is shown in Fig. 10(b). It can be found that although the spectral characteristics of both shale and sandstone are similar, the reflectivity of shale is lower than that of sandstone.

Fig. 10

(a) Spectral curve of shale (from USGS_MIN spectral library retrieved from ENVI 5.0 software) and (b) spectral curve of black shale.²⁶

Basalt is a rock with low reflectivity rate.²⁷ This indicates that the remotely sensed image in the basalt region has a dark tone or is black. The spectral curves of the main minerals in basalt, namely, olivine, pyroxene, hornblende, and biotite, from the USGS spectral library are collected and shown in Fig. 11. This figure shows a large absorption peak located at $1.1\mu \mathrm{m}$ and a reflection peak located at $0.7\mu \mathrm{m}$. Therefore, the spectral characteristic/curve of basalt is obviously different from that of other rocks.

Fig. 11

Spectral curves of basalt (from USGS_MIN spectral library retrieved from ENVI 5.0 software).

Granite is classified as an igneous rock that contains $\sim 65\%$ to 75% of silicon dioxide.²⁷ The spectral curves of the constituents of granite rocks, namely, potash feldspar, quartz, muscovite, and biotite, from the USGS spectral library are collected and shown in Fig. 12. It can be found from Fig. 12 that granite and basalt have similar spectral curves, as both of them contain biotite. This implies that an alternate method should be used to classify granite rocks.

Fig. 12

Spectral curves of granite (from USGS_MIN spectral library retrieved from ENVI 5.0 software).

In addition, the spectral curves of marble minerals, namely, tremolite, alusite, serpentine, and potassium feldspar, from the USGS spectral library are collected and shown in Fig. 13. It is found from Fig. 13 that granite and marble rocks have similar spectral curves because both of them contain potash feldspar.

Fig. 13

Spectral curves of marble minerals (from USGS_MIN spectral library retrieved from ENVI 5.0 software).

Carbonate rocks consist of different types of iron ores such as siderite and limonite. The spectral curves of calcite and dolomite have significant absorption peaks, but are relatively flat between 2.075 and $2.351\mu \mathrm{m}$; the spectral curves of siderite and limonite are relatively flat between 1.550 and $1.750\mu \mathrm{m}$ (see Fig. 14).

Fig. 14

Spectral characteristics of minerals in carbonate (from USGS_MIN spectral library retrieved from ENVI 5.0 software).

The following conclusions can be drawn from the above spectral analysis:

3.2.

Classification of Rocks

Based on the above analysis, a flowchart for the classification of six types of lithologies is proposed in Fig. 15. The details are described as follows.

Fig. 15

Flowchart of the lithologic classification.

3.2.2.

Classification of shale and marble rocks

After carbonate is extracted, the remaining five types of rocks, namely, shale, marble, sandstone, granite, and basalt, will be classified using five training areas representing five types of lithologies. The five types of training areas are selected from the 1:200,000 geological map. The operations are as follows:

• 1. Identify the five types of lithology distributions in the 1:200,000 geological map.

• 2. Use the coordinate of each type of lithology in the 1:200,000 geological map to find the corresponding position/area in the Landsat 7 ETM+ imagery. The corresponding features and spectral features are recorded. For example, the coverage of five types of lithologies and their coordinates can be determined by interpreting the 1:200,000 geological map [see Fig. 16(a)], and then the corresponding positions in the Landsat 7 ETM+ imagery can be located [see Fig. 16(b)].

• 3. Repeat the second step to find each type of lithology in the study area and record their features and spectral curves.

Fig. 16

Selection of training area from 1:200,000 geographic map: (a) 1:200,000 geographic map and (b) training area.

The training data sets of the five types of lithologies were selected from the 1:200,000 geographic map and are shown in Fig. 17. As observed from Fig. 17, the texture features of shale and marble rocks are obviously different. Figures 10 and 13 show that the spectral curves of shale and marble rocks are significantly different. Therefore, the minimum distance method is used to classify the shale and the marble rocks. To avoid misclassification of shale and sandstone, it is proposed to first classify shale and then extract sandstone. Therefore, the training data sampled in the analysis in this section should be very accurate and reliable. The classification results are depicted in Fig. 20(a).

Fig. 17

Image features of the six types of lithologies in the study area.

With the 50 selected checkpoints, the classification accuracies of shale and marble rocks reach $\sim 67.4\%$ and 65.7%, respectively.

3.2.3.

Classification of sandstone and granite rocks

In order to further classify these rocks, ENVI 5.0 was used to collect the spectral curves measured in the laboratory, as described in Sec. 3.1. We further resampled these spectral curves at the spectral range of 0.4 through $2.4\mu \mathrm{m}$. Finally, the spectral curves of seven rock samples were created, as shown in Fig. 18.

Fig. 18

Spectral curves of laboratory measurements of seven rock samples.

Because the content of acidic igneous rocks is generally more than 65%,²⁷ this paper uses 65% quartz and 35% quartz porphyry and granite porphyry as the rock sample below. Based on the components of granite rocks, the spectral curves can be calculated using Spectral Math Suit in ENVI 5.0 as

Eq. (3)

$$s1\times 0.65+(s2+s3)\times 0.35,$$

where $s1$ represents the spectral curve of quartz, and $s2$ and $s3$ represent the spectral curves of quartz porphyry and granite porphyry, respectively. Similarly, the spectral curves of sandstone can be expressed as

Eq. (4)

$$(s1+s4+s5)/3,$$

where $s1$, $s4$, and $s5$ represent the spectral curves of quartz, dolomite, and sandstone, respectively. In addition, the spectral curves of basalt can be expressed as

Eq. (5)

$$s1\times 0.48+(s6+s7)\times 0.52,$$

where $s1$ and $s6$ represent the spectral curves of quartz and diabase, respectively, and $s7$ represents that for basalt. Because the component of quartz in basalt accounts for $\sim 48\%$, the proportion of quartz is set to 48%.

The calculated spectral curves for granite rocks, sandstone, and basalt calculated using Eqs. (3)–(5) are shown in Fig. 19.

Fig. 19

Calculated spectral curves of granite, sandstone, and basalt.

The classification of granite rocks, sandstone, and basalt based on the results in Fig. 19 is performed as follows. First, the difference in the spectral information for sandstone and granite rocks at $1.4\mu \mathrm{m}$ is used as the basis for classification, as the absorption peaks of granite rocks are obvious at 1.4 and $1.9\mu \mathrm{m}$, whereas the absorption peaks for both sandstone and basalt are not obvious at 1.4 and $1.9\mu \mathrm{m}$. With the interpreted results in Sec. 2.2, we can roughly determine the range of interest of sandstone and acidic igneous rocks on the images. After binarization processing, we conduct image segmentation by selecting thresholds for classification of sandstone and acidic igneous rocks. Finally, we can segment the areas of sandstone and granite rocks manually.

After the above processing, sandstone and granite rock areas were extracted for analysis. The areas of sandstone and granite rocks were segmented by the decision tree. The classification results were evaluated using 25 checkpoints selected in the study area. The 25 checkpoints were superimposed on the geologic map for accuracy assessment [Fig. 20(a)]. The results of the classification accuracy are shown in Table 2. As seen from Table 2, an average classification accuracy of 65.2% $[(60.3+70)/2]$ was achieved.

Fig. 20

(a) Classification results for six types of rocks, (b) and (c) field verification.

Table 2

Lithologic classification accuracy.

		Reference data (%)
		Carbonate	Sandstone	Granite	Shale	Marble	Basalt
Remote sensing image classification data (%)	Carbonate	64.5
	Sandstone		60.3
	Granite			70
	Shale				67.4
	Marble					65.7
	Basalt						67.8

4.1.

Field Verification of Classification Accuracy

In order to verify the classification accuracy, the city of Guanyang located at the longitude of 110°43′16″ E through 111°20′13″ E and latitude of 25°10′32″ N through 25°45′37″ N was selected for the accuracy assessment of the field verification [see Figs. 20(b) and 20(c)]. Six types of rocks were sampled and their coverage areas were investigated. Based on the field verification, the classification accuracies of carbonate, sandstone, granite, shale, marble, and basalt rocks were $\sim 64.5\%$, 60.3%, 70%, 67.4%, 65.7%, and 67.8%, respectively. The average accuracy was $\sim 66\%$ [see Table 2 and Fig. 20(a)]. The classification accuracy with 66% can meet the requirement of 1:100,000 geological mapping, which is referenced to the Standard of China Geological Map Remote Sensing Interpretation Technology.²⁹

In addition, the areas containing the six types of rocks were calculated in accordance with the number of pixels and the corresponding lithologies. The results are shown in Fig. 21.

Fig. 21

Lithologic statistics of the study area.

From Figs. 20 and 21, it can be noted that the area occupied by carbonate in the study area is the largest. The carbonate rocks mainly cover the northern part of the study area; they also appear in the southwest part of the study area. The area occupied by sandstone is the second largest. The sandstone is widely distributed in the northwest part of the study area; they also appear in the central and eastern regions. Shale is mainly distributed in the eastern region. The granite distribution is concentrated in the south and northeast parts of the study area. The area covered by marble rocks is mainly located in the northern and southeastern regions. Basalt covers the smallest area and is scattered throughout the study area.

4.2.

Discussions and Remarks

As analyzed above, the classification accuracies of carbonate rocks and sandstone are not ideal. This is probably because their minimal components are similar. For example, carbonate rock components consist of quartz and calcite, and over 52% of sandstone is comprised of quartz. As shown in Fig. 22, the spectral curves of quartz and calcite differ widely around the bands of 0.6 and $1.4\mu \mathrm{m}$. Based on the results of the field verification, the classification results of limestone are correct, as limestone is a type of carbonate rock with calcite as the main mineral and only a small amount of other components. The classification accuracy for carbonate rocks is high because carbonate rocks in the city of Guanyang are mainly composed of limestone, i.e., without the components of sandstone and shale.

Fig. 22

Spectral analysis of quartz and calcite (the major components of carbonate and sandstone).

On the other hand, $s1$, $s2$, and $s3$ in Eq. (4) are based on laboratory measurements, in which the sample numbers are limited. Therefore, the classification accuracy of sandstone is low. From Table 2, the classification accuracies for both shale and carbonate rocks are close. In addition, the band ratio method affects the spectral characteristics of both clay and green mudstone. However, clay is the main component of shale. Therefore, the band ratio method affects the classification accuracies of both carbonate and shale.