An integrated remote sensing, petrology, and field geology analyses for Neoproterozoic basement rocks in some parts of the southern Egyptian-Nubian Shield

The main objective of this study was to use deep learning, and convolutional neural networks (CNN), integrated with field geology to identify distinct lithological units. The Samadia-Tunduba region of the South Eastern Desert of Egypt was mapped geologically for the first time thanks to the use of processed developed CNN algorithms using Landsat 9 OLI-2, which were further enhanced by geological fieldwork, spectral measurements of field samples, and petrographic examination. According to previously published papers, a significant difference was observed in the distribution of rocks and their boundaries, as well as the previously published geological maps that were not accurately compatible with the nature of the area. The many lithologic units in the region are refined using principal component analysis, color ratio composites, and false-color composites. These techniques demonstrated the ability to distinguish between various igneous and metamorphic rock types, especially metavolcanics, metasediments, granodiorite, and biotite monzogranite. The Key structural trends, lithological units, and wadis affecting the area under study are improved by the principal component analysis approach (PC 3, 2, 1), (PC 2, 3, 4), (PC 4, 3, 2), (PC 5, 4, 3), and (PC 6, 5, 4) in RGB, respectively. The best band ratios recorded in the area are recorded the good discrimination (6/5, 4/3, and 2/1), (4/2, 6/7, and 5/6), and (3/2, 5/6, and 4/6) for RGB. The classification map achieved an overall accuracy of 95.27%, and these results from Landsat-9 data were validated by field geology and petrographical studies. The results of this survey can make a significant difference to detailed geological studies. A detailed map of the new district has been prepared through a combination of deep learning and fieldwork.


Geologic background and geomorphology
Wadi Samadai-Wadi Um Tunduba district is located in the southern part of the Central Eastern Desert of Egypt (CEDE) just to the north of the major low-angle thrust that marks the boundary between the Central Eastern Desert (CED) and the South Eastern Desert (SED).The SED can be partially defined by its geologic boundaries extending from Marsa Alam to the international border of Egypt at a latitude of 22° N (the Sudan line).In the northern part of the Southeastern desert is the combined tectonic boundary with the CED.The SED seems to also show infrastructure-superstructure relationships like those of the CED, but the units in the SED are somewhat older.The SED in general seems to represent a deeper level of exposure than the CED and is much less affected by Najd shearing.It is situated about 10 km Southwest of Marsa Alam city and 70 km from Abu Ghusun.The present region is approximately 283 km 2 in area and situated between latitudes 24°56'-25°00' N and longitudes 34°46' and 34°52'30" E (Fig. 1).The geologic map of the area (Fig. 2) shows master lineaments of NW and NE trends, with a prominent N-S trend in the western part of the area; the NE trend is the youngest.It was affected by the Marsa Alam Shear Zone trending from NW to SE 32 .Wadi Samadai-Um Tunduba district is covered mainly by Neoproterozoic basement rocks.The studied district is a part of the Cryogenian-Ediacaran belt encountered in the South Eastern Desert of Egypt.According to 33 , the studied district is classified according to their relative age relations into the following rock units, starting with the youngest (Fig. 2).
Dismembered ophiolitic rocks (serpentinites) reach an elevation of 410 m in Gabal Samadai.The dismembered ophiolitic ultramafic rocks represent essentially by serpentinites with minor exposure's dunite, peridotite and pyroxenite which are the oldest rock unit in the study district.
Island arc assemblages are comprising essentially the metavolcanics, arc-related metasediments, and metagabbro-diorite complex.The island arc metavolcanics of are represented essentially by metabasalts, metadolerites, and meta-andesites.It occurs as low to moderate hills at the western side of the mapped area, as well as some scattered small, isolated bodies in the central and southeastern parts of the studied district.Generally, they  www.nature.com/scientificreports/are fine-grained rocks, but sometimes displaying porphyritic texture.The metasediments as a whole tend to be of finer to medium grain.They comprise a well-bedded repeated alternating sequence of both massive and laminated schists.The small, elongated belt of the metagabbro-diorite complex occurs either as boulders, or as medium to high-hilly masses dissected by many joint trends at Wadi Samadai (Fig. 2 14 ) and also exposed at the east of Wadi Um Tunduba.Syn-to Late-tectonic stage comprises the syn-and late-tectonic granitic rocks.However, the syn-tectonic granitic rocks are represented by the tonalite-granodiorite complex.On the other hand, the late-tectonic granite is mainly represented by monzogranites and alkali feldspar granites.It is drained by Wadi Tundeba and Wadi Samadai after which the granitic mass was named.It is intruded into the surrounding metasediments, metavolcanics and metagabbros with irregular sharp contacts.
Younger felsic volcanics extruded within the metavolcanics (older), arc-related metasediments, granodiorites, and monzogranites, essentially represented by rhyolite.Intrusive young gabbros are undeformed post-tectonic plutonic rocks.They are represented by isometric un-layered, small masses intruded into the metavolcanics and the arc-related metasediments.
The relief is generally hilly and desert-like, with two valleys, Umm Tunduba and Umm Samadai, which have been shaped by wind and water erosion.The study district lies between 52 and 550 m in altitude.The highest areas rise steeply to the southwest of the study site.Geomorphologically, this district is characterized by dendritic, rectangular, and parallel drainage systems are also present in the district (Fig. 3a).Due to their arid environment, the valleys are generally devoid of vegetation, except for a few drought-adapted plants, such as thorny bushes and hardy grasses.In addition, relatively steep slopes (25°-35° slope) in the northwest, south-central, north-central, and southwest, and a nearly flat plateau with a gentle slope in the east-north, east-central, and east-south (Fig. 3b) of the study district.Radial drainage can be observed in this district.

Materials and methods
In the present study, a variety of research techniques have been applied, including remote sensing data, fieldwork, laboratory work, which includes petrographical investigations (Fig. 4).

Landsat-9 image pre-processing analysis
The new multispectral remote sensing satellite Landsat-9 (Operational Land Imager 2; OLI-2), which is the most recent of the Landsat satellite heritage, was one of the satellite data sources employed in the current study.The Landsat 9 satellite carries two science instruments, the Operational Land Imager 2 (OLI-2) and the Thermal Infrared Sensor 2 (TIRS-2).The OLI-2 captures observations of the Earth's surface in visible, near-infrared, and shortwave-infrared bands, and TIRS-2 measures thermal infrared radiation, or heat, emitted from the Earth's surface.Both OLI and TIRS have a 5-year mission design life, although the spacecraft has 10+ years of consumables.These two devices collect data in seven multispectral bands.Coastal aerosol (B1), blue (B2), green (B3), red (B4), near-infrared (B5), shortwave infrared (SWIR) (B6), and SWIR 2 are the colors that make them up (B7).

Convolution neural network algorithm
Convolutional neural networks (CNNs), a significant advance in the field of machine learning algorithms, have had a profound impact on geology by revolutionising our understanding of geological landscapes 34,35 .Their ability to overcome the limitations of traditional algorithms, which focus exclusively on pixel-by-pixel classification and neglect spatial features, positions them as major transformation tools in this field.
CNNs are proving to be exceptional in the precise identification of rock formations, mineral deposits, fault lines and other geological structures [36][37][38] .Their hierarchical architecture, incorporating convolutional, pooling and fully connected layers, excels in feature extraction and image classification.Compared to traditional machine learning algorithms, CNNs stand out for their ability to automatically learn more complex feature representations,   39,40 .Additionally, CNNs enable end-to-end learning, eliminating the need for manual feature engineering.This approach enhances the autonomy of the learning process, significantly simplifying the task of geological structure recognition.
For lithological identification, we exploit the well-established AlexNet CNN architecture 41 .As a backbone to learning high-level representations from projected images using a 10-layer structure, inspired by AlexNet with 34 convolutional layers (Conv1 to Conv34).The model architecture starts with an initial convolution layer and incorporates structured blocks to promote residual learning.This allows it to categorize many more objects.In addition, it deals with overfitting by using dropout rather than regularization 41 .
The input image size was set to 96 × 96 pixels, and ten input layers (Landsat 9 OLI-2, 30 m Aster DEM, slope, principal components, and ratio bands) were selected, a decision motivated by the search for a balance between the ability to capture the main spatial features of the target classes and the need to avoid the loss of significant spatial patterns.Experiments were carried out by varying the size of the input image windows, demonstrating that the 96 × 96 dimension proved ideal for achieving the objective.In an exploratory approach, we evaluated the use of the 15 m Landsat pixel size as a base resolution by resampling the Digital Elevation Model (DEM).This initiative reflects a proactive approach to adjusting the resolution of the data to better match the specific characteristics of the geological terrain under study.
The model training process was carried out exclusively with available data.A batch size of 50 was selected, and training was carried out over 100 epochs using a Titan RTX.The learning rate was initially set at 0.01 for the first 20 epochs, then reduced to 0.001 for the next 50 epochs, implementing a strategy for dynamically managing the learning rate as the training progressed.The final layer of the model was structured with a SoftMax classifier, characterized by an output dimension corresponding to the 11 lithological types identified in the Wadi Samadai-Wadi Um Tunduba region.This fully connected layer translates the information extracted by the model into probabilistic predictions for each lithological class, facilitating geological mapping.

Methodology
To evaluate the CNN method adopted, we followed the procedure shown in the figure below (Fig. 4).The first sections present the steps involved in the various data and their preparation.These steps were described in the previous sections.
In this study, a semi-supervised approach was used by using clustering of trained network feature maps was used.CNN algorithms are based on complex representations that characterize the features of an input image, known as "feature maps" 42 .These features take into account the distribution of data at different scales and are optimized during training.Once trained, CNN use these feature maps to make predictions based on multiple non-linear operations 36 .
Approximately 8 to 25 training samples were collected for each rock unit in the study district.Selected training samples were evaluated by using histograms, signature averaging plots and signature alarm methods to detect areas of overlap between samples collected in different bands.Supervised classification accuracy assessments were performed by using a confusion or error matrix.The pixel that was classified from the image was compared to the same geological map available and verified on the site.Reference data are selected from samples of each class by using a stratified random sampling strategy.For each class on the classification map, 8 to 25 samples are taken as reference data.

Field work and laboratory experiments
After visual analysis of digitally processed OLI images, field investigation is used to confirm a geological map.Most of the fieldwork is devoted to verifying rock samples, and lithological differentiation.About 75 representative samples were collected from different lithological units, whereas about 59 samples are selected for petrographically examination in an attempt to establish a basis for distinguishing among the main groups.
The petrographic study includes the examination of 59 thin sections of the selected samples representing the different rock varieties in the study district.These thin sections were studied under a polarizing microscope (Optika B-353POL) combined with a digital camera at the Department of Geology, Faculty of Sciences, Al-Azhar University, Egypt.

Spectral analysis
Spectral signature analysis is performed to identify rocks based on their spectral signature.The spectral signatures of nine rock samples were carefully selected to represent all the rock units exposed in Wadi Samadai-Wadi Um Tunduba.Analysis of reflectance profile curve results from raw and resampled data were analyzed using the Landsat-9 OLI-2 spectral specification to select the optimum band ratio for identifying and mapping the different lithologies studied.The results of these spectroscopic measurements (absorption and reflection) are mainly based on rock-forming minerals in specific wavelength parts of the electromagnetic spectrum.The resultant profile curve is illustrated in the figure below (Fig. 5).
The spectral profile measured for the gabbro samples (Fig. 5) shows four absorption anomalies at 0.5-0.65 µm, 0.84-0.88µm, and 1.5-1.6 µm.These absorption anomalies are due to the presence of gabbro-forming minerals such as augite and bytownite.Based on the above observations, gabbro shows relatively high absorption in bands 7 and 6, and relatively moderate reflectance in bands 2 and 3. Four main absorption anomalies were recorded in the spectral profile of volcanic felsic samples (Fig. 5).These absorption anomalies result from the mineral composition of this rock unit, as shown in the Fig. 5.The first occurs at 0.85-0.88µm; this wavelength is linked to Landsat 9 band 5.The second anomaly occurs at 1.4 µm, due to the strong absorption of quartz, feldspar, amphibole, and biotite in this wavelength range.The third absorption anomaly is located around 1.5-1.6 µm; this wavelength is linked to Landsat 9 band 6.The fourth strong absorption anomaly appears at 2.2-2.3 µm, as all minerals forming felsic volcanics have strong absorption in this wavelength range.This rock shows high reflectance in bands 6 and 7.
For granite samples, three absorption anomalies were recorded on the spectral profile (Fig. 5) at 0.5-0.65 µm, 1.5-1.6 µm, and 2.2-2.35µm.These absorption anomalies are due to the presence of granite-forming minerals such as quartz, feldspar, and mica.Based on the above observations, granite shows relatively high absorption in bands 3, 5, and 6 and relatively moderate reflectances in all bands.The spectral profiles of biotite monzogranite and granodiorite are similar because their mineral composition is similar.However, biotite monzogranite shows a higher reflectance than granodiorite because the proportions of minerals forming the rock are different.The spectral profile of the biotite monzogranite and granodiorite samples shows four absorption anomalies.The first is strong at 1.6-1.8µm, due to the presence of microcline, oligoclase and albite.The second absorption is around 1.8-1.9µm and is due to the absorption characteristics of anorthite and orthoclase.The third and fourth absorption anomalies are located at around 2.0 µm and 2.1-2.3 µm, respectively, due to the major absorption of microcline, oligoclase, albite, and quartz (Fig. 5).The spectral profile shows that biotite monzogranite and granodiorite have a high reflectance of 0.78 in the 1.6-1.8µm range.In addition, they have high reflectance in the range (1.60-1.65 µm) covering band 6.
Four absorption anomalies were recorded for the metagabbro samples (Fig. 5).The first is located in the 1.65-1.8µm range, due to the presence of actinolite and hornblende (Fig. 5).The second and third are located around 1.85 µm and 2.0 µm respectively, due to the presence of plagioclase, actinolite, and quartz.The fourth absorption anomaly is strong and occurs at 2.2-2.3 µm, as all the minerals forming this rock have strong absorption at this distance.Based on the above observations, metagabbros have a high reflectance in bands 6 and 7.
Several principal absorption anomalies were recorded in the spectral profile of the metasediments samples (Fig. 5).These absorption anomalies result from the mineral composition of this rock unit.The first occurs at 0.85-1.00µm; this wavelength is linked to Landsat-8 band 5.The second anomaly occurs at 1.05-1.25 µm, due to the strong absorption of chlorite schist and mica schist in this wavelength range.The third and fourth absorption anomalies are located around 1.3 and 1.5 µm.The fifth and sixth strong absorption anomalies appear at 1.6-1.8µm, as all minerals forming metasediments have strong absorption in this wavelength range.The others are located between 2.05 and 2.3 µm.This rock shows high reflectance in bands 5, 6, and 7.
Metavolcanics are composed of hornblende, plagioclase, actinolite, augite, hornblende, and chlorite.These minerals show strong absorption anomalies in two regions (Fig. 5): the first absorption at around 1.85 µm and the second at 3-2.5 µm.On the other hand, the basic metavolcanics have a strong absorption that does not exceed 0.13 in all bands.
For serpentinite samples, two main absorption anomalies were recorded on the spectral profile (Fig. 5).The first is located in the 0.85 µm range, due to the presence of chrysotile, lizardite, and antigorite.The second absorption anomaly occurs in the 1.7-1.85µm range, due to the strong absorption of kaolinite, chrysotile, lizardite, and antigorite.Serpentinites have relatively high absorption in near-infrared band 5 (0.85-0.88 µm) and mid-infrared band 7 (1.5-1.66 µm).On the other hand, serpentinites show high reflectance in bands 5 and 7.

Landsat-9 image processing analysis
Principal component analysis (PCA) Principal component analysis (PCA), which reduces the variance of the original bands and shows correlations in the raw remote sensing data, which is very useful for identifying the lithological characteristics of rocks and minerals on the basis of their spectral properties [43][44][45][46] .The first PC1 has the highest variance, 96.12% of all variance data and includes an equal mix of all bands; however, band 5 has more influence (higher eigenvalue of 0.325768, followed by SWIR1, SWIR2, and near infrared bands, respectively).It includes the most information in the input bands and can therefore be used to illustrate lithological and structural features.On the other hand, the last PC contains the least information from the input bands, accounting for 0.008% of the total variance.PC2 represents 2.92% of the total variance, perpendicular to PC1, and shows the spectral variations between the visible data, with the highest positive charge (0.18) for band 1 and the highest negative charge (− 0.94) for band 2.
The first three PCAs generally comprise more than 99.5% of the total variance (Table 2); the second and third PCAs can detect surface features that were not initially distinctive due to the high similarity of the original data.In addition, PC4 has a high positive value 0.223510 in band 1 and a low negative value − 0.7959 in band 4; thus, PC4 is useful for discriminating between rock units.However, visual inspection of the upper PCs revealed important information relating to the distribution of lithological units.However, in the current work, lithological discrimination and the delineation of structural features were achieved extremely successfully by using Landsat-9 (OLI-2) pictures in grayscale VNIR-SWIR bands.
Detailed inspection of these ratio combinations shows the particular characteristics of certain rock units.The 6/5 ratio is useful for mapping rock-forming iron minerals, as these minerals have a high reflectance in this ratio.In the band ratio composite image (6/5, 4/3, and 2/1); (Fig. 7a), limestones are cyan, while granites are orange, metagabbro-diorite complex is brown, gabbros is blue, metavolcanics are green, and granodiorites is red.On the other hand, the composite band ratio image (4/2, 6/7, and 5/6; Fig. 7b) was the most discriminating of the four combinations.The 4/2 ratio is used to map iron oxides, the metagabbro-diorite complex was well detected on this combination (light green), metavolcanics in blue, limestones in light purple, metasediments in green, and biotite monzogranites in beige.

Classification results
The results obtained by our methodology enable us to produce a lithological map by forming an evolutionary neural network by using the various detailed data (Fig. 8).The classified map below shows ten lithological units.
To assess the resulting classified map, the accuracy evaluation was calculated for all resampling methodological combinations.For this survey, test categories were formed for the evaluation of classification accuracy.Pixels were selected at random.Two thousand and five hundred sample points were chosen (Table 3).We used overall accuracy, user and producer accuracy, and the Kappa coefficient (standard) for accuracy assessment.
The classified map (Fig. 8) shows good agreement between the thematic map generated from the Landsat-9 image (OLI-2) and the reference data.The accuracy of the lithological map was assessed by using validation  www.nature.com/scientificreports/pixels, and the results summarized by using a confusion matrix (Table 3).According to accuracy assessment techniques, the results indicate an overall accuracy of 95.27%.The Kappa coefficient value of 0.94 indicates very good agreement between image-generated thematic maps and reference data.

Field verification
For more confidence, the above description in the remote sensing part for each exposed rock unit, the findings of a thorough field verification process were used to assess and confirm the distribution data obtained.

Metamorphic assemblages
Serpentinites considers the oldest rock unit in the study district, which are dismembered ophiolitic rock units.These serpentinites are moderate to high mountains, as well as vary in colors from greenish grey to dark green, brownish, and violet.It is distributed in the northern portion of Wadi Samadai and as small mass in the southwestern part at the end of Wadi Tunduba (Fig. 9a).There are extremely few talc-carbonate rocks in the Gabal Samadai region, and they only appear in shearing zones as a result of CO 2 metasomatism of serpentinites (Fig. 9a).Metavolcanics mainly occurred in the western sector of the study district.They are highly sheared, jointed, and composed essentially of metabasalts.Metabasalts are intruded by biotite monzogranites (Fig. 9b).
Arc-related metasediments are commonly exposed in the northern part of Wadi Samadai (Fig. 9c) and southern part of Wadi Um Tunduba.These rocks are mainly dark-colored, massive to strongly foliated, chiefly mafic, graded welded crystal tuffs.They are varies in grain sizes from fine-to medium-grained and characterized by planer structures (Fig. 9d).
The metagabbro-diorite complex represents the younger phase of the island arc assemblages.They are exposed on the northwestern part of the study district.It occurs as boulders, medium to high hilly masses dissected by many joint trends at Wadi Samadai (Fig. 9e).These rocks intruded by granodiorite and alkali feldspar granites (Fig. 9f).

Magmatic assemblages
Syn-to late-tectonic granites: the syn-tectonic granitic rocks are represented by the granodiorites, while the late-tectonic granite includes biotite monzogranites and alkali feldspar granites.Samadai granitic mass is situated in central, southeast of the mapped area.These granitic masses belonged to syn to late-tectonic stages of pan-African movement.Granodiorites often occur as massive rocks, grey color (Fig. 9g).Sometimes, they are highly weathered and deformed especially near the contact with the older rock units.Um Tunduba granodiorite intruded by alkali feldspar granites (Fig. 9g).The biotite monzogranite is represented by a high body, pinkish-grey and pinkish red in color, medium-to coarse-grained, highly weathered, and jointed (Fig. 10a).Alkali feldspar granites are pink or red in color, medium-to coarse-grained, and are extensively weathered along the joints.They are exposed in the central and eastern parts of the mapped area.These rocks are extensively fractured and jointed (Fig. 2) to form huge blocks.Alkali feldspar granites enclose basic enclaves and xenoliths of metagabbros and granodiorites, which are irregularly distributed throughout the granite pluton.The alkali feldspar granites intruded within metagabbros, granodiorites, and intrusive gabbro, with sharp contact (Fig. 10b).
Younger felsic volcanics are widespread in the study district, especially in the northeastern part of Wadi Samadai and in the western part of the mapped area (Fig. 2).They are represent by rhyolite.Rhyolite is occurring as small exposure encountered in the western part at the end of Wadi Um Tunduba (Fig. 10c).It is characterized by light grey color, fine-to very fine-grained.
Intrusive young gabbros are highly weathered, characterized by abundant content of pyroxene, calcic plagioclase, and hornblende minerals.Wadi Samadai alkali feldspar granites are intruded directly by small plug-like masses of gabbro, which simulate the young intrusive gabbros (Fig. 10d, e).
Dykes and veins: Wadi Samadai-Wadi Um Tunduba district is traversed by numerous dyke swarms, which consequently crossed all the pre-existing rock units.Most of these dykes have NE-SW trend, less often they have NW-SE.The thickness ranges from centimeters to more than 1m.These dykes are represented by acidic and basic dykes.Acidic dykes are represented by granitic and rhyolitic dykes (Fig. 10f).The granitic dykes are pink to red in color, fine-grained with thickness ranging from 25 cm to 1.5 m.On the other hand, the rhyolitic dykes are light grey and red in color, fine-grained with thickness up to 4 m.Basic dykes are represented by dolerites and basalts.They are characterized by the less weathered and dark green in color with length ranges from 1 to 4 m (Fig. 10g).

Petrography
Serpentinites are mainly composed of serpentine minerals (antigorite-chrysotile) with secondary talc carbonates as well as accessory opaques (Fig. 11a).Opaque minerals are mainly magnetite and chromite, which occurs as euhedral crystals of bloody color or as dark grains.It characterized by mesh and ribbon textures.Metavolcanics are represented mainly by metabasalts.These rocks are composed of plagioclase and pyroxene as well as some secondary minerals such as actinolite, chlorite, and carbonate minerals.Sometimes, they exhibit ophitoblastic and diabasic texture (Fig. 11b).
Arc-related metasediments are represented by schists.They are characterized by medium-grade metamorphic rocks with sheet-like grains in a preferred orientation.According to the mineral composition and grade of metamorphism, the schists are divided into three types: tremolite schists, staurolite-kyanite-biotite schists, and mylonitic schists.
Tremolite schists are composed essentially of amphiboles, plagioclase, and quartz as well as veinlets of quartz and plagioclase crosscutting the foliation.Amphiboles represented by tremolite and actinolite (Fig. 11c).Staurolite-kyanite-biotite schists are dominated by staurolite, kyanite, and biotite interlayered with quartz and feldspar (Fig. 11d).It is a fine-to medium-grained schistose texture.Prismatic staurolite and kyanite are preferred orientations, which show nematoblastic texture (Fig. 11e).Mylonitize schists formed by ductile deformation during intense shearing encountered during folding and faulting, a process termed (cataclastic or dynamic

Figure 3 .
Figure 3. (a) Drainage system and (b) Slope model of the study district.

Figure 4 .
Figure 4. Flowchart of the study district.