2.1 Geology

The Wahiba Basin of NE Oman, housing the Wahiba Dune Field and surrounding environments, was formed by regional subsidence between the Al Hajar Mountains to the north and the Huqf Arch to the south (Figure 1) (Immenhauser et al., 2000; Shackleton & Ries, 1990). The Al Hajar Mountains contain a 15-km thick slab of oceanic lithosphere obducted onto the continental margin of Arabia (Bernoulli & Weissert, 1987). They have elevations exceeding 2200 m and are flanked by rugged terrains of the emplaced Semail ophiolite, characterised by Late Cretaceous mafic to ultra-mafic oceanic crust (Glennie, 1988; Hanna, 1990) (Figures 1 and 2). A series of imbricated units of continental margin and/or oceanic origin, known as the Hawasina Nappe Complex, are sandwiched between parautochthonous sequences of the Hajar Supergroup (SG) and the Semail ophiolite (Figures 1 and 2) (Bernoulli & Weissert, 1987; Hansman et al., 2021). The Hajar SG consists primarily of Late Permian to Late Cretaceous carbonate strata and minor siliciclastic components (Béchennec et al., 1992). The Hawasina Complex contains extensive units of limestone and radiolarian chert (Figures 1 and 2) (Béchennec et al., 1988, 1992; Glennie, 1988; Hanna, 1990; Kusky, Robinson, & El-Baz, 2005). Post-obductional Tertiary cover (Tertiary Limestones) includes marine slope to shelf deposits of the Hadhramaut/Dhofar/Fars Groups, consisting of limestone, shale and conglomerate (Figures 1 and 2) (Glennie, 1988; Hansman et al., 2021). Late Eocene to Pleistocene uplift and subsequent drainage from the Al Hajar Mountains resulted in the development of extensive, laterally overlapping alluvial megafans forming the Barzaman Formation (Barzamanite) composed of chert and pedogenically altered limestone and ophiolitic sediments (Figures 1 and 2) (Blechschmidt et al., 2009; Maizels, 1988; Rodgers & Gunatilaka, 2002).

The Huqf Arch, located on the southern margin of the Wahiba Basin, contains outcrops of the oldest known sedimentary sequences of Oman, which uncomformably overlie the Neoproterozoic granodioritic basement (Figures 1 and 2) (Searle, 2019). The Huqf Arch is an anticlinal feature that trends SW–NE and hosts continental siliciclastic to marine sediments of the Huqf-Haima SGs (Ediacaran to Early Paleozoic), the Haushi-Akhdar Groups (Carboniferous to Permian) and the Aruma Group (Late Cretaceous) (Figure 2) (Allen, Leather, & Brasier, 2004; Angiolini et al., 2004; Angiolini, Stephenson, & Leven, 2006; Droste, 1997; Searle, 2019). Outcrops of the Huqf SG and overlying Haima SG dominate the structural high and contain extensive continental quartzose sandstones (fluvio-aeolian), marine carbonates/bioclastics and evaporitic deposits (Allen & Leather, 2006; Gorin, Racz, & Walter, 1982; Leather, 2001). The Huashi and Akhdar Groups occur as isolated outcrops containing polymictic conglomerates, quartzose/bioclastic sandstones, siltstones, shales and carbonates (Figure 2) (Angiolini et al., 2003, 2004; Osterloff et al., 2004; Platel et al., 1992). The Aruma Group lies on the western margin of the Huqf Arch and is dominated by quartzose sandstones, bioclastic limestone/marl and extensive chalk deposits (Figure 2) (Dubreuilh et al., 1992; Platel et al., 1992). South of the Huqf Arch, the Jiddat Al Harasis highlands are made up of the Tertiary Limestone deposits of the Hadhramaut/Dhofar/Fars Groups and mark a period of marine transgressive–regressive pulses (Figure 2) (Dubreuilh et al., 1992).

2.2 Climate and aeolian transport systems

Today, the highland regions of the Al Hajar Mountains host an arid climate (100–200 mm rainfall pa), whereas the low lying Wahiba Basin is classified as hyper-arid, typically receiving less than 100 mm of rainfall annually (Kwarteng, Dorvlo, & Vijaya Kumar, 2009). The east coast of Oman is subject to three seasonal wind modes associated with regional climatic systems (Figure 2) (Fryberger, Hern, & Glennie, 2016a, 2016b; Patlakas et al., 2019): (1) strong and persistent SW winds of the Indian summer monsoon (ISM), (2) Shamal NW winds that persist through the year and (3) NE trade winds of the Indian winter monsoon (IWM). The ISM SW winds are strongest over the Arabian Sea, deflect onshore as southerlies, and diminish toward the interior of Oman. In summer, weak NW Shamal winds form part of the ISM circulation and are relatively consistent. In winter, the Shamal winds occur with periodic frontal systems that originate in the Mediterranean, bringing strong NW winds that sweep across the Arabian Peninsula. Shamal winds generally approach the Wahiba Basin from the W and diminish toward the coast. The moderate IWM NE trade winds rotate inland to approach the Wahiba Basin from the SE (Figure 2).

2.3 Geomorphology and stratigraphy of the Wahiba Dune Field

Various geomorphic provinces characterise the Wahiba Dune Field (Figure 1) (e.g. Fryberger, Hern, & Glennie, 2016a, 2016b; Gardner, 1988; Goudie et al., 1987; Robinson et al., 2007). The northerly High Sands is composed of NNE–SSW trending linear megadunes, separated by interdune corridors, and crossed obliquely by active seif and/or crescentic dunes (Figure 3). The megadunes transition from asymmetrical structures (steep western face) with average heights of ~60 m in the north to more subdued symmetrical structures with average heights of ~30 m in the south. Active reworking is apparent on the eastern margin of the High Sands where infilling of the interdune corridors and mega crescentic features disturb the linear bedforms (Al-Masrahy & Mountney, 2015; Kocurek et al., 2020). At South of the High Sands, the relief lowers and transitions into the Low Sands, dominated by northward migrating crescentic dunes (Figure 1). The deflationary Al Jabin Plateau lies west of the Low Sands and contains variable sandsheet cover separated by aeolianite outcrops that are established as N–S trending yardang landforms (Figure 1). The aeolianite outcrops are dated to glacial and interglacial periods between marine isotope stage (MIS) 7 and 5d (Glennie et al., 2011; Juyal, Singhvi, & Glennie, 1998; Radies et al., 2004). Along the east coast, the Coastal Sands contain NE migrating crescentic dunes with outcrops of aeolianites interspersed within interdune areas (Figure 1). The aeolianites also form beach terraces of up to 20-m high and are dated to the Last Glacial Maximum (LGM) period of MIS 2 (Gardner, 1988; Glennie & Singhvi, 2002; Preusser et al., 2005; Radies et al., 2004, 2005). The Peripheral Sands lies directly north and east of the Coastal Sands and High Sands, respectively, and is composed of barchanoid dunes, linear seif dunes and nabkha (Figure 1).

The Wahiba Dune Field accumulation has occurred over at least the past 229 ka, and has been primarily linked to the dominance of the southerly monsoonal winds systems of the ISM (Glennie et al., 2011; Juyal, Singhvi, & Glennie, 1998; Preusser, Radies, & Matter, 2002; Radies et al., 2004). Continuous core and trench sampling have demonstrated that Milankovitch-controlled cyclicity is the dominant signal across the Wahiba Basin (Glennie et al., 2011; Glennie & Singhvi, 2002; Juyal, Singhvi, & Glennie, 1998; Preusser, 2009; Preusser, Radies, & Matter, 2002; Radies et al., 2004). This cyclicity manifests as periods of dune stabilisation during wet interglacial cycles (e.g. MIS 5), and dune reactivation and accumulation during arid glacial periods (e.g. MIS 2 aged stratigraphy in Hern et al. (2018); Figure 3).

2.4 Fluvial and coastal transport systems

The Wahiba Basin sediments include first cycle wadi sands transported from the local highlands to adjacent coasts or interior regions (Figure 2) (Garzanti et al., 2013). Three major drainage systems surround the Wahiba Dune Field: the Halfayn system to the north and west, the Al Batha to the north and east, and the Al Huqf on the south (El-Baz, 2002). The Halfayn (comprising Wadi Halfayn and Wadi Andam) and Al Batha systems are primarily derived from the Al Hajar Mountains, and the Halfayn additionally erodes Barzamanite megafan sediment (Figures 1 and 2). Collectively, these sediments are identified by mafic grains of the Semail ophiolite, and chert and carbonate grains from the Hawsina Nappe and Tertiary Limestones (Garzanti et al., 2003; Goudie et al., 1987; Maizels, 1988; Pease et al., 1999; Pease & Tchakerian, 2002, 2014; Robinson et al., 2007). Paleozoic to Neoproterozoic detritus of the southern Al Huqf drainage system includes rounded monocrystalline quartz, carbonate and chert (Garzanti et al., 2003). The drainage systems either terminate within the portion of the catchment occupied by the Wahiba Dune Field (Halfayn and Al Batha systems), yield to lowland plains (all systems), or drain onto the coastal shelf (Al Batha and Al Huqf systems) (Blechschmidt et al., 2009; El-Baz, 2002; Garzanti et al., 2003; Robinson et al., 2007). The active littoral zone and shallow coastal shelf hosts a mixture of modern bioclastic sediments of the productive Arabian Sea, and continental terrigenous sediments, which are subject to northward longshore transport (Garzanti et al., 2003, 2013; Nairn et al., 2015; Pease et al., 1999; Pease & Tchakerian, 2014).

3.1 Dataset preparation

3.1.1 Worldview-3 dataset

The Worldview-3 (WV-3) dataset covers an area of 100 km² located on the NW margin of the dune field where the linear megadunes are truncated by Wadi Batha (Figures 1 and 3; Table 2). Coverage includes eight NNE–SSW trending asymmetric megadunes, spaced between interdune sandsheets that are commonly crossed by active linear seif to crescentic dunes and contain partial rock outcrops of the Hawasina Group (Figure 3). A portion of Wadi Batha is exposed on the NE corner of the image where the megadunes terminate.

Cloud free and orthorectified WV-3 scenes were acquired from DigitalGlobe Inc. (Colorado, USA). Owing to commercial restriction, the datasets provided were resampled to 0.5, 2 and 7.5 m per pixel in the panchromatic (PAN), VNIR and SWIR bands, respectively (Table 1). Radiometric calibration was performed in order to convert the 16-bit integer digital numbers (DN) into top of the atmosphere radiance data. To compensate for atmospheric effects, the datasets were processed using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) atmospheric correction module to obtain surface reflectance values (Module, F.L.A.A.S.H., 2009).

Image interpretation and processing benefited greatly from image-sharpening techniques that increased the spatial resolution of the datasets using high-resolution panchromatic images (Figure 5) (e.g. Kim et al., 2016; Kwan et al., 2017; Park et al., 2020; Park & Choi, 2021; Selva et al., 2015). The pansharpening process aims to fuse a multispectral (MS) and a panchromatic image, the result of which contains the spectral resolution of the former with the spatial resolution of the latter (Vivone et al., 2014). Co-registered WV-3 multispectral datasets were pansharpened to 0.5 m per pixel and layer-stacked following the sequential-parallel fusion process of Kwan et al. (2017), utilising the Gram–Schmidt (GS) algorithm of Laben and Brower (2000) (Figure 5).

3.2 Data validation approaches

The validation of the WV-3 dataset was accomplished by means of comparing a band ratio analysis against a supervised classification approach using a Spectral Angle Mapper (SAM) algorithm (Figure 6). The band ratio classification is guided by published reflectance spectra of applicable minerals, and the SAM classification utilises the mean spectra over each georeferenced ground sample. Once validated, the WV-3 dataset served as a reference to evaluate the accuracy of a SAM-classified Sentinel-2 dataset for the same region. A band ratio analysis was not conducted for the Sentinel-2 dataset because the satellite only contains two broad SWIR bands (Figure 4). To understand the limitations of the lower resolution Sentinel-2 dataset, both a visual inspection and a confusion (error) matrix analysis was conducted between the two satellites. Finally, a SAM algorithm was run on the regional Sentinel-2 datasets using the mean spectra of each sample location. The final production of the regional classification maps is guided by the known limitations of the lower resolution Sentinel-2 satellite.

3.2.1 Band ratio analysis

Band ratio analysis utilises spectral indices to discriminate and map surface rocks/sediments. Considering that each object has a unique spectral reflectance curve in each wavelength of the electromagnetic spectrum (Figure 7), spectral indices can be used to emphasise the anomaly of a target object by determining the band at which the reflectance is high or at a point of highest absorption (Gupta, 2018; Lillesand, Kiefer, & Chipman, 2015; Mars, 2018); Pour & Hashim, 2015; Prost, 2013; Rockwell & Hofstra, 2008.

With a total of 16 narrow bands, the WV-3 dataset is ideal for reflectance and absorption that distinguish certain minerals (Figure 7a and Table 1). Ferric iron (Fe3+) leads to strong absorption in the UV to blue wavelength region, while ferrous iron (Fe2+) is known to exhibit some absorption features around 600–2050 nm. Calcium carbonates have absorption features at 2320 and 2510 nm in the SWIR region. Clay minerals also have a characteristic absorption feature in SWIR spectrum of 2100–2400 nm and shifts slightly depending on the clay type. Caution should therefore be exercised in interpreting clays versus calcium carbonate abundance, as both minerals have similar absorption features and common reflectance features around 1550–1750 nm. Quartz consists of a simple silicate tetrahedra, and the absorption bands only occur in the TIR around 8500–9500 nm. However, as WV-3 lacks TIR bands, quartz cannot be directly identified and required the used of the ASTER dataset. (Figure 7b).

To select suitable band ratios, reflectance spectra for the appropriate minerals were obtained from the NASA Jet Propulsion Laboratory (JPL) spectral library (https://speclib.jpl.nasa.gov/library) and the USGS Spectral Library v.7 (Kokaly et al., 2017) (Figure 7). The band ratio application produces a single-channel greyscale image, in which pixels with elevated reflectance values correspond to characteristic minerals. A ‘bright pixel slice’ classification approach was used to delineate the high-reflectance targets.

4.1 Surface samples

Quantitative data of XRD and thin section analyses indicate that quartz, undifferentiated carbonates and mafic material typically make up more than 80% of aeolian dune samples, with increased quantities of clay and/or evaporite minerals present in fluvial, lacustrine and sabkha samples. For simplification and with the aim of spectral characterisation of mixed sediment, the samples in this paper are reported normalised by their total dominant mineralogy (Figure 8).

A total of 14 megadunes with variable heights were sampled across the High Sands. On average, the low stoss, mid stoss and crestal sampling occurred approximately 18, 31 and 50 m above local interdune elevations, respectively (e.g. Figure 3c). Mineralogically, mean values indicate that carbonate and mafic minerals increase upslope, from low stoss to mid stoss and crest, inversely with decreasing quartz (Figure 8a). Samples of the Halfayn system indicate that southward drainage is associated with a decrease in carbonates, mafic material and clay, and increasing quartz content (Figure 8b). Mafic and clay content decrease downstream in Wadi Al Batha drainage, inversely with quartz, until reaching the coastal delta that contains low levels of quartz and high levels of clay and evaporite minerals (Figures 2 and 8b). Carbonate content varies in the Al Batha System with an interdune pond containing the highest levels of carbonate and evaporite minerals. At the Coastal Sands, beach samples are depleted in mafic content, however, contain nearly double the amount of carbonate than the inland coastal dunes that are relatively rich in quartz and mafic content (Figures 2 and 8b). The coastal aeolianite deposit contains similar mineralogy to the carbonate-rich beach sand with a mafic content similar to that of an adjacent active dune (Figure 8b). The coastal sabkha contains high levels of evaporite minerals and relatively high mafic material. In the Low Sands, quartz content increases southwards inversely with carbonate and mafic material. A sampled dune on the proximal Al Jabin Plateau contains a similar carbonate concentration to Low Sands samples but significantly more mafic content (Figures 2 and 8b). At the Huqf Arch, in both wadi and aeolian dune deposits, higher quartz content is inverse with carbonate content in more northerly located regions. The beach sediment of Huqf contains significant amount of carbonate grains and includes evaporite minerals (Figures 2 and 8b).

Thin sections analysis reveals characteristic grain varieties present across environments (Figure 9). Ferric-coated quartz grains (Q) are ubiquitous in all environments and vary from sub-angular very fine sand to silt, to well-rounded fine sand. Carbonate grains are classified into four categories, namely rounded limestone grains (Lr), carbonate aggregate grains (Ag), caliche fragments (Cf) and biogenic fragments (Bf). Lr are found in all environments across the study region and are typically well-rounded fine sand composed of micrite or cement filled tests interpreted as reworked fossil foraminifera. Ag are medium-to-coarse sand-sized grains consisting of an aggregate of micrite cemented Lr grains. Ag range from sub-angular, typical of wadi samples, to more rounded grains identified in the High Sands. Cf are typically sub-rounded, medium to coarse sand and are only identified on megadune low to mid stoss regions. The grains consist of a micrite cement encasing sub-angular very fine sand to silt grains of quartz, biogenic/limestone fragments and minor mafic material. Cf may contain calcite filled cellular clusters indicative of plant roots. Bf include skeletal debris and shallow water benthic foraminifera that may be fairly pristine and are interpreted as modern foraminifera. These fragments are identified on Huqf's beach, the Low Sands including Al Jabin dunes and the Coastal Sands including aeolianite deposits. Ophiolite fragments (Of) consist of olivine and pyroxene, typically in grains of serpentinised harzburgite. Of range from fine sand, generally concentrated on megadune crests along the W and NE margin of the High Sands, to coarse sand through pebble size concentrated in wadis of the Al Batha and Halfayn drainage systems. Varying amounts of pyroxene (Px), plagioclase feldspar (F) and chert (H) grains are identified in most samples.

Megadune grain size analysis indicates that upslope variation is present across many metrics that include mean grain size (Mz), sorting (σ₁), skewness (sk₁) and kurtosis (k_g) (Figures 10 and 11) (formulas of Folk & Ward, 1957). On average, both the low- and mid-stoss regions contain a mean grain size of medium sand (low = 0.21 mm; mid = 0.25 mm), are moderately sorted and contain a distribution that is coarse skewed. This is in contrast with crestal sands that consist of fine sand (0.19 mm), are moderately well sorted and have a near symmetrical distribution. All areas have a mesokurtic kurtosis, although the low and mid stoss tend towards flatter distributions (i.e. k_g < 1). This is evident in the considerably higher coarse sand fraction on the stoss, with an average of 10% on the low stoss and 17% on the mid stoss, compared to only 1.3% on the crest. The coarse sand fraction is above the typical aeolian load (0.1–0.3 mm) and forms lag deposits on the elevated stoss slopes (e.g. Figure 11a). Stoss samples with a well-developed lag contain increased mean grain size (up to 0.33 mm), are platykurtic (bimodal) and contain variable skewed distributions with a coarse sand fraction up to 32.5% (Figure 11a). Microscope analysis (Figure 11b) reveals that the fraction of lagged grains is typically sub-rounded and dominated by caliche fragments (Cf) (e.g. Figure 9c, d).

4.2 Image classification

4.2.1 WorldView-3 classification

Three distinct band ratios (spectral indices) were utilised in the mineralogical classification of the WV-3 scene (Figure 12): (1) a ferric index to represent ferric-coated quartz minerals that are ubiquitous in all environments; (2) a ferrous index specific to mafic minerals of the Semail ophiolite; and (3) a carbonate index to represent undifferentiated carbonate minerals. Because quartz minerals only contain diagnostic features in the TIR spectrum (Figure 7b), an additional quartz index of the ASTER satellite was utilised to assess the coincidence of quartz with observations of the WV-3 ferric index (Appendix S1). Results from the band ratio analysis were subsequently compared against SAM pixels for locations that contained relatively high proportions of each mineral type (Figure 12) (Appendix S4).

Quartz

Ferric-coated quartz minerals contain a broad set of absorption features centred around 500 nm and 850 to 910 nm, which correspond to WV-3 bands V2–V4 and bands V6–V7 (Figure 7a). The laboratory spectra of both goethite and ferric-coated quartz show a high absorption and reflection feature in band V3 and band V5, respectively (Figure 7a). In order to reveal the ferric-coated quartz distribution, the V5/V3 index was applied. The resulting images (Figure 12a i–ii) show that pixels with elevated reflectance values appear along the megadune stoss slopes on the eastern margin of the image. Subsequently, the resultant SAM classification pixels were selected based on samples containing quartz content in excesses of 55%. A visual comparison shows that the SAM classified pixels are in good agreement with the brightest pixels identified in the ferric-coated quartz index (Figure 12a ii–iii). The ASTER assessment utilised the high absorption feature of quartz, centred about the thermal region of 9000–9400 nm, defined by TIR band T12, and the quartz index T14/T12 was applied (Figure 7b). The ASTER dataset has a considerably reduced spatial resolution of 90 m per pixel; however, the resulting image compares favourably with the WV-3 classification, and elevated reflectance values dominate the megadune stoss slopes along the eastern margin of the image (Figure 12a).

Mafic

Based on thin section analysis and the literature encompassing the Semail ophiolite complex (MacLeod, Johan Lissenberg, & Bibby, 2013; Prigent et al., 2018), olivine forsterite and pyroxene enstatite minerals of mantle harzburgites, and pyroxene augite minerals of crustal gabbro are considered a good representation of mafic spectra (Figure 7a). The molecular structure of these minerals contains characteristic absorption features near 950–1100 nm, corresponding to WV-3 band V8 (Figure 7a). The output image from band index S4/V8 reveals pixels with elevated reflectance values on Wadi Al Batha deposits in the NE, and on the active superimposed dunes along the megadunes of the western margin (Figure 12b i–ii). The SAM classification map was produced from the mean spectra of sampled zones containing more than 9% mafic material, and a spatial agreement exists with the mafic index, particularly on the western margin of the image and in wadi Batha (Figure 12b ii–iii). The higher representation of SAM pixels, however, may be because of spectral interference with more abundant mineral species (Figure 8a).

Carbonate

Carbonate mineral identification utilised the band ratio (S6/S8)*(S3/S8), necessitated by their absorption features in band S8, coupled with relatively high reflectance in bands S6 and S3 (Figure 7a). Results reveal that pixels with elevated reflectance values are concentrated on the megadunes of the western half of the WV-3 scene and decrease eastwards (Figure 12c i-ii). Results are particularly distinct along the N–S trending stoss slopes, in corridors that lack active superimposed dunes, on exposed Hawasina outcrops, and on Wadi Al Batha (Figures 3, 12c i-ii). The SAM classification utilised the mean spectra of zones where samples contained more than 55% carbonate material, and a visual comparison indicates a good accordance with the carbonate index map (Figure 12c ii–iii). The unclassified wadi sediment by SAM is unsurprising, as the proximal Wadi Al Batha sample contains less than 24% carbonate material (Figure 8b). The high index values present over the wadi sediment in the carbonate index map is probably because of interference with a high clay content (42%; Figure 8b), which contains the same absorption feature in band S8 (e.g. chlorite in Figure 7a).

4.2.2 Sentinel-2 classification

With only two bands in the SWIR region of the electromagnetic spectrum, the Sentinel-2 dataset is inherently limited in spectral resolution, reducing the data dimensionality and degrees of freedom available for mineral differentiation. This limitation affects the ability to accurately perform linear spectral unmixing, contributing to challenges in classification accuracy (e.g. Lillesand, Kiefer, & Chipman, 2015; Tompolidi et al., 2020). The quality assessment of Sentinel-2 classification was determined by statistical (Table 3) and visual (Figures 13 and 14) comparisons to the calibrated WV-3 scene. To enable a direct comparison between the two datasets, the Sentinel-2 scene was masked to include the same area as the WV-3 scene (Figure 13), and WV-3 pixels were resampled to 10 m per pixel to match raster resolutions. The scene was further masked to include only the megadunes (excluding interdunes) before running a confusion (error) matrix analysis and calculating a Kappa coefficient (Table 3). Statistically, when comparing all 13 sample spectra run through the SAM algorithm (Figure 13a), a poor overall accuracy of 26.6% with a corresponding Kappa coefficient of 0.18 was observed between the two datasets. However, isolating classes that contain distinct differences in quartz, carbonate and mafic mineralogy (i.e. coloured diamonds in the ternary plot in Figure 13a) yielded a more favourable overall accuracy and Kappa coefficient of 78.8% and 0.69, respectively (Table 3). High producer's and user's accuracy were achieved for both quartz-rich and carbonate-rich sample classes, while much lower accuracy was observed for the mafic-rich sample class. Mafic material is commonly less than 10% of the bulk sample content, and the low accuracies indicate that spectral mixing and confusion between mafic content and other classes is to be expected in Sentinel-2 classifications.

TABLE 3. Confusion matrix of classified Sentinel-2 and Worldview-3 datasets (see Figure 13).

Sentinel-2 classification	Ground truth pixels (WV-3)
	High mafic	High carbonate	Hiqh quartz	High quartz 2	User's accuracy
High mafic	52	24	14	10	52%
High carbonate	32	129	2	0	79%
Hiqh quartz	5	2	26	9	62%
High quartz2	0	1	8	198	96%
Total	89	156	50	219
Producer's accuracy	58%	83%	52%	90%
Overall accuracy = 78.79%			Kappa coefficient = 0.69

Visually, the SAM pixel classification of VW-3 and Sentinel-2 shows that the general trends across the scene are comparative (Figure 13b, c). Further inspection indicates that misclassified pixels are typically concentrated on smaller superimposed dunes and slip faces (Figure 14). These misclassifications are thought be attributed to a combination of the high angle lee slope affecting spectral reflectance, the lower spatial resolution of Sentinel-2, and differences in scene acquisition dates and active sediment movement. Overall, these observations highlight the sensitivity of the Sentinel-2 dataset to spectral mixing compared with the WV-3 datasets. Subsequent Sentinel-2 dataset classification was therefore conducted by isolating geomorphic provinces and utilising the mean spectra of sampled zones that contain pronounced mineralogical heterogeneity (Figures 15-18).