Invisible Herders, Visible Traces: A Spatio-Chemical and Remote Sensing Approach to Deciphering Pastoral Landscapes in Dhofar, Oman

2.1. Study area: the Khor Rori area, SW Oman

The Khor Rori area, located approximately 40 kilometres east of Salalah in the Dhofar Governorate, was selected as the study area for this research (Figure 1). The area is situated at the mouth of Wadi Darbat, the largest drainage system feeding into the coastal plain, which empties into a bar-built discharge that historically served as a natural harbour (Steele et al. 2024). The exceptional preservation of this landscape is partly attributable to its early inclusion in the UNESCO World Heritage List as part of the ‘Land of Frankincense’ designation in 2000 (UNESCO 2000).

Figure 1

The Dhofar region presents a compelling setting for studying pastoralism due to its remarkable long-term continuity as a socio-ecological adaptation—a continuity fundamentally enabled by its distinctive monsoon-driven environment. The landscape is shaped by the Southwest Monsoon, locally known as the Khareef, a climatic system active for at least 10,000 years (Tigani ElMahi 2001). This seasonal phenomenon, occurring between mid-June and mid-September, brings moisture-laden winds from the southwest Indian Ocean, producing persistent low-level stratus clouds, fog, and intermittent drizzle rather than heavy rainfall. While the region remains generally arid, the Khareef transforms the mountains into a verdant refuge surrounded by desert, delivering approximately 300 mm of annual precipitation, with effective moisture availability substantially exceeding this due to fog interception (Arnold et al. 2020; Tigani ElMahi 2001). The terrain extends from coastal lowlands through foothills to mountain escarpments dissected by deep wadis (ephemeral watercourses), producing distinct ecological zones historically utilised by pastoral communities (Ramadan, Al-Awadhi & Charabi 2021; Tigani ElMahi 2001).

Archaeological evidence suggests that Neolithic communities were present in Dhofar by the late 7^th to early 6^th millennium BCE, within the broader emergence of pastoral societies in the Omani peninsula from around 6500 BCE or possibly earlier (Charpentier et al. 2023). Although the progressive aridification of the Rubʿ al-Khali has been invoked to explain a movement of herding groups towards the more stable environments of the Dhofar mountains (Newton and Zarins 2010; Tigani ElMahi 2001), Neolithic sites are now known across the coastal plain, the mountains, the Nejd and the desert interior. This distribution suggests a more complex system of mobility and interaction between ecological zones (Charpentier et al. 2023). More recent archaeological research in Dhofar has documented a wide range of settlement forms and occupation strategies. Within this broader framework, research conducted by the DHOMIAP project has revealed permanent settlements with monumental architecture alongside seasonal open-air sites, confirming occupation from the Early Iron Age (from the 9^th–8^th century BCE) through the Late Iron Age (c. 1^st century BCE–2^nd century CE) and into the Islamic period (c. 10^th–11^th century CE) (Lischi 2022; Lischi 2025; Rougeulle 2008). Elsewhere in Dhofar, excavations at Iron Age sites such as Shakil and Halqoot revealed material signatures—rooms with bedrock floors, deep accumulations of burnt cattle dung, and a striking scarcity of portable artefacts—more consistent with repeated, long-term pastoral use than with sedentary agricultural occupation (McCorriston et al. 2020).

Contemporary pastoralism in Dhofar employs three primary livestock species—cattle (Bos indicus), goats (Capra hircus), and camels (Camelus dromedarius)—each managed within specific ecological zones through dynamic herding strategies calibrated to seasonal cycles (McCorriston et al. 2020; Tigani ElMahi 2001). The distinct Khareef seasonality shapes a transhumant regime in which each species follows specific movement patterns. During winter and summer, cattle remain on the Khatum and Qatan plateaus, while during the Khareef, they disperse across the plateaus, foothills, and occasionally coastal plains. Camels occupy the foothills and Qatan during the dry seasons but are moved to the coastal plains and Nejd during the monsoon to avoid the slippery mountain slopes, where these heavy animals risk serious injuries. Goats are herded in the foothills and escarpment for most of the year but relocated to the foothills, coastal plains, and Nejd during the Khareef due to their intolerance of wet conditions (Tigani ElMahi 2001; Tigani ElMahi 2011b). Environmental pressures, particularly predation by hematophagous flies, further structure pastoral practice, with herders systematically collecting and burning manure to eliminate fly breeding areas (Tigani ElMahi 2011a). The material culture of these communities reflects adaptation to mobility constraints, characterised by minimal permanent construction: circular habitat areas (strayat), livestock enclosures (duguf), and small stone enclosures (al-masadid) represent the primary architectural features (Tigani ElMahi 2001).

Given this rich presence of active and recently abandoned pastoral camps embedded within a landscape of documented long-term pastoral continuity, this region provides essential ground truth for establishing the methodological approach. Within this landscape, 15 seasonal pastoral campsites were selected and documented following the methodological protocol described in Biagetti et al. (2026). These settlements typically comprise distinct functional zones: a penning area for livestock enclosure, a domestic living space, and associated features such as fireplaces. Campsite selection was based on physical visual inspection, prioritising locations that exhibited diverse activity areas. The documentation process involved GPS mapping, photographic recording, and detailed sketching. Occupation history was assessed through multi-temporal satellite imagery analysis through Google Earth’s historical imagery (2003–2023), combined with in situ sediment sampling (January 2025). Campsites were finally revisited in January 2026 to check whether they had been used during the last monsoon season. Two locations initially recorded as campsites (S11, S12) were reclassified during the second field visit as vegetation patches rather than pastoral encampments. Their samples were retained for two reasons: vegetation cover was sparse enough for the sediment surface to remain visible to the sensor, and their compositions extend the range of low-enrichment values against which the anthropogenic–lithogenic balance is calculated. Of the 13 campsites, three (S4, S7, S10) showed evidence of continued seasonal use at the time of the second field visit in January 2026, while the remaining sites were last visible as active campsites in the historical satellite record between 2016 and 2023, indicating abandonment within roughly the last decade.

A total of 1,441 soil samples were collected following a multi-resolution spatial sampling strategy: a dense 2-metre grid for domestic zones and 4-metre grids in penning areas where geochemical signatures are more homogeneous. Additional targeted sampling was conducted at discrete features such as fireplaces. All samples were collected at a 5 cm depth to capture the anthropogenic horizon while removing superficial aeolian deposits. Spatial coordinates were recorded using a pole-mounted GNSS receiver (UTM Zone 40N, EPSG:32640). Additionally, 150 control samples were collected across the wider pastoral landscape to characterise the regional geochemistry, allowing for a statistically robust estimation of the baseline signal. This follows the standard CAMP protocol of c. 150 off-site control samples per study region (Biagetti et al. 2026), a fixed sampling effort that ensures baselines remain comparable across the project’s study area. Control locations avoided visible pastoral activity and disturbed contexts such as tracks, active grazing areas, and recently abandoned structures.

2.2. Case study: campsite S4

To test the feasibility of upscaling geochemical detection through remote sensing, the analysis begins with contemporary settlements where anthropogenic signals are expected to be strongest. From the 13 documented campsites, campsite S4 (Figure 2) was selected based on two criteria: it remained in active seasonal use, ensuring strong geochemical signatures; and it was the only site exhibiting the complete range of functional zones characteristic of pastoral settlements in the region.

Figure 2

Campsite S4 is located in the northeastern portion of the Khor Rori area and was identified as a seasonally used camel corral situated on limestone bedrock. The campsite comprises three distinct activity zones: a roughly circular fenced penning area of approximately 170 m² (c. 15 m diameter), a small domestic space of approximately 12 m² positioned to the southeast of the corral, and a fireplace (c. 0.5 m²) located to the southwest (Figure 3). The penning enclosure consisted of a semi-complete metal mesh fence, with some sections collapsed on the ground, while visible dung accumulation within the corral interior confirmed recent animal presence. A total of 61 soil samples were collected across these activity zones, following the multi-resolution sampling strategy outlined in Section 2.1 (Figure 3).

Figure 3

2.3. pXRF spectrometry for multi-element determination

The geochemical analysis of soil samples collected from the Khor Rori area was conducted using portable X-ray fluorescence (pXRF) spectroscopy, following the laboratory protocol established by the CAMP project (Biagetti et al. 2021; Biagetti et al. 2026). The machine used to analyse the samples was a Bruker CTX 800 portable benchtop XRF analyser equipped with a rhodium target X-ray tube, and a large-area silicon drift detector was employed for all measurements. Prior to analysis, air-dried bulk sediment samples were sieved through a 1 mm mesh directly into plastic sample cups prepared with 4.0 µm polypropylene thin film, a preparation method designed to increase reliability and reproducibility while providing sample homogenization (Middleton 2004; Williams, Taylor & Orr 2020). The instrument was configured using the Bruker ‘GeoExploration’ application with the ‘Oxide3phase’ method, combining three measurement phases at different voltages and filters to detect elements ranging from magnesium to uranium. Integration times were set at 30 seconds each for intermediate and heavy element detection phases, and 100 seconds for light elements such as phosphorus and magnesium, which lie at the edge of pXRF detection limits.

The employment of the pXRF method for identifying anthropogenic signatures in pastoral contexts builds upon foundational multi-element approaches that were primarily introduced by Barba and Ortiz (1992) and Middleton and Price (1996), who demonstrated that repetitive human activities leave characteristic elemental patterns in occupation surfaces. Although pXRF raises technical and methodological concerns that require standardised protocols (Hunt and Speakman 2015; Williams, Taylor & Orr 2020), its rapid, non-destructive, and cost-effective nature was essential for analysing the 1,441 samples.

2.4. Compositional data analysis (CoDA) and geostatistical modelling

Geochemical data derived from pXRF analysis are inherently compositional, meaning that elemental concentrations represent proportions of a whole rather than independent variables. When one element increases the others must decrease proportionally, creating dependencies that violate the assumptions of classical statistics and can generate spurious correlations (Aitchison 1982). The analysis therefore follows a Compositional Data Analysis (CoDA) framework, in which data reside in the simplex, a space that inherently accounts for the constant-sum and positivity constraints (Filzmoser, Hron & Templ 2018; Greenacre and Wood 2024; Martín-Fernández, Buxeda i Garrigós & Pawlowsky-Glahn 2015). Within this framework the meaningful information lies in the ratios between elements rather than their absolute values.

Before the analysis, data preparation addressed zeros and outliers, two fundamental challenges in compositional datasets. Values below detection limits were imputed using methods that respect compositional principles such as scale invariance and subcompositional coherence (Palarea-Albaladejo and Martín-Fernández 2015), while outlier detection employed robust Mahalanobis distances computed on Orthonormal Log-Ratio (OLR) coordinates derived from a balanced basis, using the Minimum Covariance Determinant (MCD) estimator (Filzmoser and Hron 2008); the outlier threshold was set at the chi-squared quantile with α = 0.975. Outliers representing meaningful subpopulations rather than measurement errors were retained. Furthermore, log-ratio transformations were then applied to convert the data from the simplex into Euclidean space: Centred Log-Ratio (CLR) coordinates for exploratory analyses such as Principal Component Analysis, Orthonormal Log-Ratio (OLR) coordinates, also termed isometric log-ratio (Egozcue et al. 2003), for geostatistical modelling where proper Euclidean covariance structures are required, and symmetric log-ratio balances for correlation analysis, as standard correlations are invalid for compositional data (Kynčlová, Hron & Filzmoser 2017).

Geostatistical techniques were used to model the spatial distribution of elemental signatures. Variogram analysis characterised the spatial structure of the data (Hilal et al. 2024; Matheron 1963), the Linear Model of Coregionalization (LMC), jointly with the variogram analysis, described the spatial autocorrelation of OLR-transformed variables and their cross-correlations (Goulard and Voltz 1992; Tolosana-Delgado and Mueller 2021). Cokriging predicted elemental distributions at unsampled locations, incorporating spatial structure and cross-correlations while providing uncertainty estimates (Tolosana-Delgado, Mueller & Van Den Boogaart 2019). Finally, Maximum Autocorrelation Factors (MAF) analysis enhanced spatially structured signals by separating coherent patterns from random variation (Tolosana-Delgado and Mueller 2021). A detailed presentation of this workflow is provided in Biagetti et al. (2026), using the code available in the following repository: https://github.com/project-camp/camp-public. Model performance was assessed through cross-validation diagnostics.

2.5. Multispectral and hyperspectral remote sensing

Materials interact with electromagnetic radiation in characteristic ways, producing spectral signatures that encode mineralogy, organic matter content, moisture and texture (Clark 1999; Hunt 1977). Hyperspectral sensors resolve these through numerous narrow, contiguous bands, whereas multispectral sensors use broad, discrete bands insufficient for subtle absorption features (Goetz 2009) but often extending into the shortwave infrared, critical for soil mineralogy and moisture.

Hyperspectral imagery was acquired from the Wyvern Dragonette-001 satellite (6U cubesat, ~550 km altitude) on June 10, 2025, capturing 23 spectral bands spanning 503–799 nm at a 5.3 m ground sampling distance. Commercial data were provided through ESA’s Network of Resources Initiative and delivered via UP42. To complement this, multispectral imagery was obtained from Sentinel-2 on June 11, 2025 (one day after the Dragonette acquisition). Although soil sampling was conducted in January 2025, both sampling and imaging fell within the extended dry window prior to the Khareef monsoon, so surface reflectance properties remained unaffected by seasonal moisture or vegetation cover. Imagery was atmospherically corrected using the 6S radiative transfer model (Py6S interface; Vermote et al. 1997; Wilson 2013) with a maritime aerosol profile, given the study area’s proximity to the Arabian Sea coastline. To exploit complementary sensor strengths, Sentinel-2 NIR plateau bands (B8, 842 nm; B8A, 865 nm) and SWIR bands (B11, 1614 nm; B12, 2202 nm) were reprojected onto the Dragonette reference grid by bilinear interpolation and concatenated through layer-stacking, creating a fused dataset of 27 spectral bands. Initial testing indicated that prioritising spectral richness over spatial sharpening yielded superior model accuracy.

The feature space was further enlarged through spectral indices known to correlate with specific biophysical properties (Poggio and Gimona 2017). These include NDVI for vegetation abundance (Rouse et al. 1974) and SAVI for soil brightness adjustment (Huete 1988), and MSAVI2, which removes the need for an externally specified soil-adjustment factor (Qi et al. 1994):

Moreover, specific mineralogical indices were generated, including the Clay Index (SWIR2/SWIR1) responding to clay mineral absorption features (Drury 1987) and the Calcite Index, targeting the mineral alterations often associated with anthropogenic soil compaction and chemical weathering. The Bare Soil Index (BSI), which contrasts the SWIR and red bands against the NIR and blue bands, was also included as a measure of exposed soil relative to vegetation cover (Rikimaru, Roy & Miyatake 2002):

Moisture-sensitive indices such as NDMI (Gao 1996) and NDWI (McFeeters 1996) were included to capture the differential moisture retention capacities of organic-rich corral soils versus the surrounding arid lithology.

This multi-source expansion follows Poggio and Gimona (2017), who report that multi-sensor integration improves soil property model accuracy over single-sensor approaches. Ground sample coordinates were transformed to match the coordinate reference system of the fused raster (WGS 84 / UTM Zone 40N, EPSG:32640). Where multiple samples fell within a single pixel, compositional aggregation was performed using the geometric mean in simplex space. This procedure reduced the initial dataset of 1,441 discrete samples to 708 unique pixel-level observations (a 51% reduction due to high spatial clustering of sampling points). Because this study implements the CAMP protocol for the specific target of pastoral campsites, sampling was designed around the spatial scale of the geochemical features themselves rather than the pixel grid of the satellite data. A sampling design matched to sensor resolution would retain a higher proportion of independent observations and improve the correspondence between ground and image scales in future applications. In cases of ‘mixed pixels’ containing both pastoral and background samples, we prioritised samples representing pastoral activities. This preserves the integrity of the anthropogenic signal: given that over half the dataset required aggregation, averaging with background samples would have diluted the chemical signature and risked false negatives in training.

A rigorous background removal protocol isolated the anthropogenic signal. A custom script identified control points within a 500-metre buffer zone of each site (see Figure 4). This distance threshold was large enough to extend beyond the ‘anthropogenic halo’ of the campsites (typically <100 m for ephemeral sites) while remaining within the same lithological unit. The geometric mean of these control points—the appropriate measure of central tendency under CoDA—was used to perturb sample values, perturbation being the compositionally valid equivalent of subtraction. This filters out the geological background, leaving residuals representing anthropogenic enrichment.

Figure 4

Finally, the target variable followed the CoDA framework (Section 2.4). Rather than predicting raw concentrations, a group balance was constructed through amalgamation: a log-ratio contrast between anthropogenic markers (P, K, Ca) and lithogenic elements (Si, Al, Fe, Ti, Mg) (Kynčlová, Hron & Filzmoser 2017). This gives a target that is mathematically coherent and geochemically meaningful. Chlorine was excluded despite its prominence at S4: the balance requires all constituent elements in both site and control samples, and Cl was absent from the control points, being detected only at the most recently active campsites. Code for the remote sensing analysis is available at https://github.com/project-camp/camp-remote-sensing.

3.1. Geochemical signatures and geostatistics at campsite S4

Following pXRF analysis, Spearman rank correlations were computed using symmetric pivot coordinates appropriate for compositional data (Kynčlová, Hron & Filzmoser 2017). The results reveal a coherent lithogenic cluster: Si, Al, Fe, Ti, and Mg intercorrelate strongly (rs ≥ 0.76), reflecting the geochemical fingerprint of the local limestone bedrock. Against this geological background, the most striking result is a set of strong negative correlations between these lithogenic elements and chlorine (Si–Cl r_s = –0.79; Al–Cl r_s = –0.79; Ti–Cl r_s = –0.78; Fe–Cl r_s = –0.77) (Figure 5).

Figure 5

This pattern suggests that chlorine from an external source is masking the relative proportion of rock-derived elements in the soil. Animal waste introduces mobile elements that can obscure the geochemical signature of the underlying parent material, complicating the distinction between anthropogenic and natural soil formation processes (Middleton and Price 1996). While dung decomposition enriches phosphorus—derived from its organic component—and calcium, concentrated in the form of spherulites (Marshall et al. 2018; Shahack-Gross et al. 2004), urine is the main pathway through which chlorine leaves the animal’s body. This fits ethnographic observations in Dhofar, where pastoralists supplement livestock diets with commercial concentrates containing salt (Tigani ElMahi 2011b). Such supplementation is common in arid pastoral systems, since natural pastures rarely provide sufficient sodium and chloride and heat-tolerant livestock must tolerate high salt intake (Belhadj Slimen et al. 2016; Habeeb, Marai & Kamal 1992). Most of this salt is passed in urine, leaving a chemical trace where animals are repeatedly penned.

Principal component analysis reinforces these findings: the first two components account for 92.2% of total variance, with PC1 alone capturing 85.3%, driven predominantly by chlorine. This confirms that chlorine variation represents the dominant source of geochemical differentiation within the site, which can be interpreted as a urine marker. PC2 (6.9% of variance) isolates a secondary gradient driven by P, Zn and Sr loadings, which might indicate a signal from organic-waste accumulation associated with dung decomposition. A biplot of the first two components resolves samples into two tight clusters corresponding to corral-interior versus exterior locations, while a single fireplace sample plots as a distinct outlier (Figure 6). Notably, the overlap between DOMESTIC and OUTSIDE samples was anticipated, given the low-intensity use of domestic areas compared to the corral enclosure.

Figure 6

These compositional patterns translate into spatially coherent structures suitable for geostatistical modelling. To address the constant-sum constraint of compositional data, cokriging was applied to transformed balances. The analysis compared a direct OLR-based model against a Maximum Autocorrelation Factors (MAF) rotation. Leave-one-out cross-validation indicated that the MAF-transformed model achieved superior performance: the omnidirectional model yielded 95% multivariate accuracy with 47% precision, while directional variogram analysis at four azimuths (0°, 45°, 90°, 135°) produced comparable accuracy (95%) with improved precision (73%). The gain in precision suggests that although the overall spatial structure is largely isotropic, subtle directional dependencies in the variance structure are captured more effectively by directional modelling, giving more reliable predictions with reduced uncertainty even in a roughly circular corral. In addition, cokriging interpolation was employed to visualise the spatial geochemical distribution across the site (Figure 7). As expected from the prior Spearman and PCA analysis, chlorine concentrations peak in the central corral area and decline sharply beyond the enclosure boundary, mirroring the pattern expected from concentrated urine deposition. Within the same framework, phosphorus and calcium exhibit a more diffuse enrichment centred on the corral interior, reflecting the broader spatial footprint of dung accumulation. Apart from that, potassium closely tracks chlorine, reinforcing the interpretation that both elements derive from the same urinary source. The lithogenic elements (Al, Fe, Ti, Si, Mg) display the inverse pattern, with the highest values in exterior samples and progressive decrease towards the corral’s centre.

Figure 7

3.2. Remote sensing: spectral differentiation and technical challenges

The Dragonette sensor (503–799 nm) lacks Short-Wave Infrared (SWIR) coverage, a limitation for detecting pastoral geochemical signatures. Model performance using Dragonette bands alone was unsatisfactory, indicating that this range was insufficient to capture the relationship between surface reflectance and geochemistry. The multi-sensor fusion described in Section 2.5 was therefore implemented to expand the feature space. Before model training, a background removal protocol was applied to isolate the anthropogenic signal from local geological variability.

The model was trained and evaluated on the geochemical data from all 13 campsites and the two locations recharacterised as vegetation patches. Following aggregation to pixel resolution, the dataset was split into training (80%) and testing (20%) subsets by stratified random sampling. Stratification was necessary because functional zones (e.g., fireplaces, middens) are not present at every site, and all activity classes had to be proportionally represented in both subsets. This introduces a risk of spatial autocorrelation between adjacent pixels, but we prioritised sampling density so that the full range of geochemical variance within each class was captured.

To establish a predictive baseline using the fused imagery, a Random Forest regressor was tuned by randomised search, yielding a test-set R² of 0.51. The modelling strategy then advanced to an Extreme Gradient Boosting (XGBoost) framework, leveraging sequential residual learning and built-in regularization (R² = 0.62). The XGBoost implementation was tuned using Optuna Bayesian optimisation, with the final configuration employing a learning rate of 0.066, maximum tree depth of 4, and 195 estimators, with L2 regularisation (λ = 0.0008) and subsampling parameters calibrated to mitigate overfitting. The final performance metrics for the optimised regression model are summarised in Table 1.

The regression results reveal a gap between training (R² = 0.95) and test set performance (R² = 0.62). This gap suggests that while the model captures internal variance and spectral-geochemical relationships within the training data, it fits those data closely, and generalisation to unseen pixels is further constrained by the landscape’s lithological heterogeneity. The moderate test R² indicates that while the model captures general regional geochemical trends and identifies anomalous enrichment zones, it exhibits higher uncertainty when predicting fine-scale concentration values. Since prospection relies on detection rather than precise quantification, we assessed the model’s ability to simply distinguish ‘enriched’ pixels from ‘background’ pixels. Crucially, this did not involve training a separate classifier; instead, ground-truth pixels were labelled as enriched or background at a balance of zero, the point at which anthropogenic markers exceed the lithogenic reference. The continuous values predicted by the regression model (Table 1) were then binarised against this labelling, with the optimal prediction threshold of 0.10 determined by maximising Youden’s J statistic (Sensitivity + Specificity – 1) (Youden 1950). The diagnostic performance of this approach is presented in Table 2.

The resulting ROC-AUC score of 0.86 indicates substantial discriminative capacity. Bootstrap confidence intervals (1000 iterations) give a 95% CI of 0.78–0.92, confirming performance statistically distinct from chance. At the Youden-optimised threshold, the model successfully identified 80 out of 96 enrichment pixels (Recall: 83.33%) alongside a specificity of 76%. The high recall ensures that few genuinely enriched areas are missed, while the moderate specificity means a proportion of background pixels are also flagged, an acceptable trade-off in a prospection context, where candidate locations are subsequently verified on the ground. The classification output (Figure 8) identifies high-enrichment pixels in areas where no ground-truth samples were collected, suggesting the learned relationships generalise beyond the training locations rather than memorising site-specific patterns.

Figure 8

Model performance was additionally assessed at the level of individual sampled locations (Table 3). Within-site Spearman correlations between measured and predicted enrichment balance were computed for 14 of the 15 locations (S11 and S12, reclassified as vegetation patches, are included as low-enrichment references), ranging from 0.54 (S15) to 0.89 (S4) with a median of 0.69. S12 is excluded owing to insufficient samples (n = 5). Site-level bias remains within ±0.06 throughout, indicating no systematic over- or under-prediction at particular locations.

The landscape-scale prediction (Figure 9) reveals spatially coherent patterns, with elevated values clustering around several sites, particularly near S8–S9 in the north-west and S10–S11 in the north-east. Isolated hotspots in the northern and eastern sectors indicate possible activity beyond the sampled locations; whether these are true pastoral signatures or artefacts of lithological variation requires ground-truth verification.

Figure 9