Remote sensing signature analysis of ASTER imagery for geological mapping of Gasus area, central eastern desert, Egypt

Abstract

The using of processed ASTER images and ASTER-GDEM (Advanced Spaceborne Thermal Emission and Reflection Radiometer-global digital elevation model) supplemented by extensive geological field work and spectral measurements of field samples enabled the creating of a new geological map of Gasus area. SVM (Support Vector Machine) method obtained 86% as overall accuracy of the classification and was verified by previous published maps and field verifications. Structure lineaments were automatically extracted and directional analysis of the automatically extracted lineament maps showing that the major trend of the lineaments are NW-SE and NNW-SSE, while NNE-SSW and NE-SW trends can also be recognized but in a less significant order. These trends are compatible and corresponding to the field verification.

Introduction

By the advent of the multispectral sensors, such as Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), several geological and mineralogical studies were carried out through the analysis of absorption features of specific minerals providing a scientific background for the interpretation of remotely sensed spectroscopic data (Crowley, 1986). Each mineral has a specific spectral signature depending on various crystallographic factors. Analyzing the measured spectra of a rock gives information about rock composition, mineral proportions, and crystallinity to provide insight in geological processes, such as magmatic differentiation, regional and contact metamorphism, hydrothermal alteration, and weathering (Pontual et al., 1997). The ASTER data have been chosen because of its relatively good spatial resolution, which make them suitable for mapping geological rock units.

The aim of this study is to establish a geological map for the area under investigation, which is created by mapping the surface structure lineaments using ASTER-GDEM (global digital elevation model) and mapping the main rock units using ASTER data.

Overview of the study area

The present study area is located in the upper central eastern desert between latitude 26°21′22″ to 26°36′ 02″N and longitude 33°39′22″ to 33°57′58″E. The area is about 857 km² and lie at the south of port Safaga (Fig. 1). Geomorphologically, the study area comprises several mountains which grading in height (Fig. 2) such as Gasus (466 m), Wasif (898 m), Um al Huweitat (427), and Weira (621 m) intercalated by valleys of Quaternary undifferentiated deposits such as Safaga, Wasif, Gasus, and Saqi.

Fig. 1

Map of Egypt showing the study area

Fig. 2

ASTER-GDEM (digital elevation model) map of the study area

The Gasus area is highly fractured (Reda et al. 2011) and divided into five disconnected Extensional Sub Basins. Three of them are within the study area namely Rabah, Wasif, and Um al Huweitat. All three were affected by NE-SW trending strike slip faults (Fig. 3). The basement of these basins is comprised by Pre-Cambrian igneous and meta igneous rocks upon which sedimentation took place from Cretaceous since Tertiary (Fig. 3). The area is also characterized by the presence of ore deposits such as phosphate, gold, molybdenum, and polymetallic minerals.

Fig. 3

Compiled geologic map (after Conoco & EGPC 1987)

Data and platforms

The data used for the first part of this study comprises ASTER satellite scenes (Table 1) as well as previous published data such as the geological map of the area (Conoco & Egyptian General Petroleum Corporation 1987).

Table 1 The ASTER scene characteristics of the studied area

Preprocessing of ASTER scenes

Raw data were subjected to some correction to be prepared for processing as shown in Fig. 4. The first step represented by crosstalk correction due to stripping of photons from band 4 to SWIR band system, especially bands 5, 9 which lead to false anomalies. This error was corrected using crosstalk correction method found in Envi software. After that resampled the short-wave infrared (SWIR) to 15 * 15 m pixels by nearest neighbor algorithm to stacked nine bands (VNIR-SWIR bands). Then, applying atmospheric correction with FLAASH module due to the effect of atmosphere on the electromagnetic radiations and after correction need to make normalization to resample all data from 0 to 1 surface reflectance. Whereas, the study area lies between two ASTER scenes hence mosaicking them to one single image, and the final preprocessing step was to resize the image to the size and shape of study area.

Fig. 4

Scheme of the preprocessing steps

Methodology

In the following section, the methods used for mapping and analyzing the lineaments and lithology will be given as follows.

Methodology for Mapping surface structure lineaments

For this purpose, ASTER-GDEM (global digital elevation model) with spatial resolution 30 m was utilized. The main advantage of using GDEM is that it has less distortion in projection. Furthermore, sun can be positioned anywhere over the surface of the area. The processing steps performed in this study for the automatic lineaments analysis were

• 1. Lineament extraction

• 2. Lineament correction

Statistical analysis

Methodology for lithological mapping using multispectral supervised classification (MSC)

The classification processes was carried out using two methods.

SVM method

Support vector machines (SVMs) are widely used in pattern recognition and classification problems and have recently become an area of intense research owing to developments in the techniques and theory coupled with extensions to regression and density estimation (Mercier and Lennon 2003, Vapnik 1998). This supervised method distinguished between different classes by maximizing the margin between them; the surface which maximizes is called the optimal hyperplane, and the samples closest to the hyperplane are called support vectors. The support vectors are the critical elements of the training set (Envi.5.3 n.d.). Geometrically, the SVM modeling algorithm finds an optimal hyperplane with the maximal margin to separate two classes, which requires solving the optimization problem (Soliman et al. 2012):

$$ \operatorname{Maximize}{\sum}_{i=1}^n{\propto}_{i\kern0.5em }-\frac{1}{2}{\sum}_{j=1}^n{\propto}_i{\propto}_j{y}_i{y}_j.K\left({x}_i,{x}_j\right) $$

(1)

$$ \mathrm{Subject}\hbox{--} \mathrm{to}:{\sum}_{i=1}^n{\propto}_{i\kern0.5em }{y}_{i,\kern4em }0\le {\propto}_{\mathrm{i}}\le C $$

(2)

Where

∝_i:

refers to the weight assigned to training data, xi is called support Vector, If xi > 0, C is a regulation parameter, and

K :

is a kernel function, which is used to measure the similarity between two samples.

The RBF is applied in this study because it can classify high-dimensional data:

$$ K\left( Xi, Xj\right)={e}^{\left(-\gamma \left\Vert xi- xj\right\Vert \right)2},\gamma >0. $$

(3)

C parameter represents the penalty cost. The two γ and C depend on the range of data and distribution.

SAM method

Kruse et al. (1993) introduced another type of supervised classification called spectral angle mapper technique since the angle between the spectra resulted from the reference image and that resulted from the spectral library is measured and treating them as vectors in a space in band’s number domain (Fig. 5). Spectral angle mapper (SAM) determines the similarity by applying the following equation:

Fig. 5

2D illustration of two spectral vectors (t and r) and the spectral angle (α) between them (Mahmoud, 2014)

$$ \propto ={\cos}^{-1}\frac{\sum_{i=1}^m{t}_i{r}_i}{{\left({\sum}_{i=1}^m{t}_i^2\right)}^{1/2}{\left({\sum}_{i=1}^m{r}_i^2\right)}^{1/2}} $$

(4)

where m is spectral band number and the value of α is between 0 and π.

Results and discussion

Surface structure analysis

Linear features extracted from eight shaded relief images were created by eight contrasting illumination directions, 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°, at a constant solar elevation of 30 ° (Figs. 6a–h). Then the first four shaded relief images were integrated in one shaded relief image with illumination directions using ArcMap program (Fig. 7a). The procedure was repeated for the remaining four images (Fig. 7b). The resulted two images were used as input data for automatic lineament extraction. Lineaments were detected, depending on curve extraction, edge detection, and thresholding steps, using PCI Geomatica (2013) program. In addition, the lineaments of geological map (Fig. 3) were traced in order to compare their trends with the lineaments trend of two shaded relief images.

Fig. 6

Shaded relief images (a–h) generated using eight illumination directions 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° respectively, at a constant solar elevation 30°

Fig. 7

Two shaded relief images generated from a integration of four shaded relief images at solar angles 0°, 45°, 90°, and 135°. b Integration of four shaded relief images at solar angles 180°, 255°, 270°, and 315°

During the extraction of lineaments, errors may occur as a result of overlapping and/or discontinuity of reflectance (Fig. 8). These errors were corrected automatically, by the use of ArcMap program, thus enabling the creation of corrected lineaments and line density maps.

Fig. 8

Example of incorrect and correct lineament

The directional analysis of the automatically extracted lineament maps (rose diagrams besides histogram frequency distributions of both lineament directions and lengths) has been performed by using Rockworks.16 program to determine the major trends of the surface lineaments in the area. As it is shown in Figs. 9, 10, and 11, the major trend of the lineaments are NW-SE and NNW-SSE, while NNE-SSW and NE-SW trends can also be recognized, but in a less significant order. Field observations and collected data (Fig. 12) are in agreement with the trend analysis results indicating that the above methodology could be an efficient tool for the process of mapping lineament.

Fig. 9

a Lineaments extracted from the first shaded relief which created from four shaded relief images at solar angles 0°, 45°, 90°, and 135°. b Line density map. c Rose diagrams show the trend analysis of the interpreted structures by number and length

Fig. 10

a Lineaments extracted from the second shaded relief created from four shaded relief images at solar angles 180°, 255°, 270°, and 315°. b Line density map. c Rose diagrams show the trend analysis of the interpreted structures by number and length

Fig. 11

a Lineaments extracted from the geological map (field Observation). b Line density map. c Rose diagrams show the trend analysis of the interpreted structures by number and length

Fig. 12

Some features of observed fractures, faults, and dykes in the study area. a Major faults dissected Jabal Gasus. b Dyke affected by fault (N20 W) in Safaga area. c Boundary between dokhan volcanics and younger granite show highly deformed shear zone. d Fracture (E-W trend) along younger granite cutting in dokhan volcanics. e Horizontal mafic dike dissected the younger granite (N60 W) in Safaga area

Supervised classification analysis

The ASTER imagery was used with an SVM and SAM classifiers to provide an accurate and rapid mapping of the rock type’s distribution with high accuracy levels. Quaternary deposits were masked before starting the classification process to avoid the wrong classification due to weathering of neighboring rocks deposited in a valley.

The selection of the kernel is considered the most important task in the implementation and the performance of the SVM classifier (Li and Liu, 2010). In this study, A radial basis kernel function with penalty parameter = 100 and gamma in kernel function = 0.11 shows the highest overall classification accuracy. By using regions of interest and the spectral library of resampled spectral reflectance extracted from the image (Fig. 13), SVM and SAM methods provided efficient classification results which show discrimination in lithology units according to spectral signatures (Figs. 14 and 15).

Fig. 13

Plotting of spectral library extracted from ASTER image

Fig. 14

The classified geological map using Support vector machine

Fig. 15

Supervised Classification using spectral angle mapping by Spectral Library extracted from ASTER image

The performance of these methods were evaluated using overall accuracy using confusion matrix. The obtained overall accuracy of SVM was presented as follows.

• 1. Creating AOIs

All patterns were related to known features by using the available geological map. There were 200 regions of interest outlined. They were represented by 11 patterns in order to collect a set of statistics that describe the spectral response pattern of each rock class (Table 2). Areas of interest (AOIs) of each pattern was divided into 50% as training data and 50% testing data.

Table 2 Number of AOIs for each pattern name

• 2. Contingency Matrix Classification

Each sampled pixel only weights the statistics that determine the classes. Then, the contingency matrix (usually known as a confusion matrix or error matrix) was calculated. The matrix contains numbers and percentages of pixels that were classified as expected (Lecia Geosystems 2003). So, making the correlation between the results of classification and the reference image.

Table 3 shows the percentage contingency matrix and clarifies that the accuracy assessment is ranging between 75% for Sp class and 100% for Kutq class. The accuracy was obtained by the following equation (Moawad, 2008):

Table 3 Percentage classification contingency matrix of collected of AOIs

$$ X\mathbf{1}+\mathrm{X}2+\dots .\mathrm{Xn}/\mathrm{reference}\ \mathrm{pixels}\ \mathrm{total}\ \mathrm{number}\times 100 $$

(5)

where X1, X2 … Xn is the absolute number of pixels of each class within the contingency matrix, and the reference pixel total number is the row and/or column total in the contingency matrix. Accordingly, the overall accuracy of the classification was about 86%. In the same manner, the producer’s accuracy and user’s accuracy were calculated. Lillesand & Kiefer (2000) state that producer’s accuracy gives an indication about the accuracy of training data in the classification process, while user’s accuracy gives an indication about the probability that a pixel has been classified into a given category actually. Producer’s accuracy and user’s accuracy were obtained by the following equations (Moawad, 2008):

$$ \mathrm{Producer}'\mathrm{s}\ \mathrm{accuracy}=\mathrm{X}/\mathrm{col}.\mathrm{total}\times 100 $$

(6)

$$ \mathrm{User}'\mathrm{s}\ \mathrm{accuracy}=\mathrm{X}/\mathrm{Row}\ \mathrm{total}\times 100 $$

(7)

where X is as mentioned above, col., and row total represent the total columns and rows of each class within the matrix. Therefore, producer’s and user’s accuracies are generally high indicating high credibility of the classification process (Table 4). The classified map (Fig. 15) within the present study that contains 11 patterns represented the main rock units in the study area.

Table 4 Individual class accuracy

Verification of methodology with field data

Twenty-seven rock samples from different rock types in Gasus area such as younger, older granite (gβ& gα), dokhan volcanic (Vd), gabbro (gb), hammamat sediments (ha), intermediate and ophiolitic metavolcanics (mv& mvb), serpentine (Sp), black shale (kutq), phosphate (kuwd), limestone (Tpe), and Nubian sandstone were prepared for reflectance spectroscopic measurements. Using an ASD FieldSpec 3 Spectroradiometer, the spectral reflectance data were obtained with the wavelength range of 0.35 μm to 2.5 μm. Each rock sample is measured four times for both fresh and weathered samples then taking the average reflectance value to enhance the amplitude of measured spectra.

Before the classification process, the measured spectral data were subjected to preprocessing due to the effect of the atmosphere which absorbed the sun’s ultraviolet radiation. Spectral signature of rocks was recorded in a ASD format, which was converted into ASCII format using View Spec Pro Version 6 software. Subsequently, they were reprocessed with Microsoft Excel Software. During reprocessing, data from atmospheric water absorption bands (1.35–1.46 μm and 1.79–1.96 μm) were removed, since they were considered as noisy data in the shortwave infrared region (Fig. 16). Then, the corrected spectral data were plotted for averages of spectral reflectance to be measured (Fig. 17). The spectral signature of each rock type was classified using spectral, and a classified image was produced (Fig. 18).

Fig. 16

A reflectance spectra of rock slab showing the sites of atmospheric water absorption features (cyan rectangles) (Mahmoud, 2014)

Fig. 17

Averages of measured spectral reflectance of the rock samples without noise

Fig. 18

Supervised classification (SAM) using spectral library which extracted from rock Samples

Finally, all the output classified images and lineaments maps were integrated with previous geological map to construct a new geological map (Fig. 19).

Fig. 19

Interpreted geologic map of the study area

Conclusion

This study investigated the suitability of the ASTER-GDEM imagery analysis for surface lineaments and the rapid classification of surface lithology using the SVM and SAM methods. According to data collected from satellite images, previous geological map and field observations, a detailed geologic map was constructed. The results showed the efficient role of multispectral supervised classification as one of the best image processing methods for distinguishing between the main rock units within the study area. The interpreted geological map derived from automatic and manual interpretation considered a good base map within the present study that will be helpful for further analysis.