A Spectral and Spatial Comparison of Satellite-Based Hyperspectral Data for Geological Mapping

Abstract

The new generation of satellite hyperspectral (HS) sensors provides remarkable potential for regional-scale mineralogical mapping. However, as with any satellite sensor, mapping results are dependent on a typically complex correction procedure needed to remove atmospheric, topographic and geometric distortions before accurate reflectance spectra can be retrieved. These are typically applied by the satellite operators but use different approaches that can yield different results. In this study, we conduct a comparative analysis of PRISMA, EnMAP, and EMIT hyperspectral satellite data, alongside airborne data acquired by the HyMap sensor, to investigate the consistency between these datasets and their suitability for geological mapping. Two sites in Namibia were selected for this comparison, the Marinkas-Quellen and Epembe carbonatite complexes, based on their geological significance, relatively good exposure, arid climate and data availability. We conducted qualitative and three different quantitative comparisons of the hyperspectral data from these sites. These included correlative comparisons of (1) the reflectance values across the visible-near infrared (VNIR) to shortwave infrared (SWIR) spectral ranges, (2) established spectral indices sensitive to minerals we expect in each of the scenes, and (3) spectral abundances estimated using linear unmixing. The results highlighted a notable shift in inter-sensor consistency between the VNIR and SWIR spectral ranges, with the VNIR range being more similar between the compared sensors than the SWIR. Our qualitative comparisons suggest that the SWIR spectra from the EnMAP and EMIT sensors are the most interpretable (show the most distinct absorption features) but that latent features (i.e., endmember abundances) from the HyMap and PRISMA sensors are consistent with geological variations. We conclude that our results reinforce the need for accurate radiometric and topographic corrections, especially for the SWIR range most commonly used for geological mapping.

1. Introduction

Hyperspectral remote sensing has been widely used for mineral exploration and geological mapping [1,2,3]. Recently launched satellite-based hyperspectral sensors have made hyperspectral (HS) data more accessible than ever before. Satellite-based HS data have a large spatial footprint, which can help in understanding local to regional geology and potentially identifying the large distal footprints of mineral systems [4,5].

HS data from different sources often overlap spatially and spectrally, but their integration (to leverage relative advantages or differences in coverage of the various available sensors) and comparison (to assess their relative consistency and resolving ability) remain a challenge, mainly due to various inherent sensor and pre-processing differences [1,2,6]. Furthermore, mixed pixel effects often result in scale-dependent spectral signatures, which also present challenges when integrating data from sensors with different ground sampling distances.

In this contribution, we investigate the applicability of these different datasets for geological mapping and evaluate their consistency by comparing the following key aspects:

• Ability to resolve diagnostic mineral absorptions, based on a qualitative assessment of endmember spectra.
• Qualitative consistency between mapping products, including spectral indices and spectral abundance maps, with respect to mapped regional geology.
• Quantitative correlation between the hyperspectral reflectance estimates, band ratios and abundance maps.

These results are integrated to highlight potential inconsistencies that result from sensor effects and/or radiometric, atmospheric and topographic correction artefacts.

Hyperspectral Remote Sensing

Many substances, including minerals, absorb light at specific wavelengths, imprinting a signature in reflected light that can be used to remotely detect and characterize them [7,8,9]. The visible-near infrared (VNIR) region (400–1100 nm) is commonly affected by absorption features due to electronic processes, which can help identify transition metal ions (e.g., Fe, Cr, REEs) [7,10,11]. At longer wavelengths, the SWIR (1100–2500 nm) range is influenced by vibrational absorptions from covalent bonds associated with compounds like H₂O, OH⁻, CO₃²⁻, etc. [1,8,12,13]. The SWIR region is thus useful for mapping many important alteration and rock-forming mineral phases, including Al-OH bearing minerals (e.g., muscovite, kaolinite, smectite), sulphates (e.g., jarosite and gypsum), Mg-OH bearing minerals (e.g., biotite, hornblende, phlogopite), and carbonate minerals (e.g., calcite, dolomite) [14,15,16].

The era of spaceborne HS data acquisition was inaugurated with the launch of Hyperion aboard NASA’s EO-1 mission in 2000 [17]. This was followed by two significant HS sensors from the Chinese space agency, including Tiangong-1 (2011–2016) [18] and GaoFen-5 (2018–present) [19]. Since 2019, five hyperspectral sensors have been launched: HISUI (Hyperspectral Imager Suite) by JAXA (Japan Aerospace Exploration Agency) [20], EMIT (Earth Surface Mineral Dust Source Investigation) by NASA (National Aeronautics and Space Administration) [21], and DESIS (DLR Earth Sensing Imaging Spectrometer) by DLR (German Aerospace Center) [22] are onboard the International Space Station (ISS) at a height of ±220 km, while PRISMA (PRecursore IperSpettrale della Missione Applicativa) by ASI (Agenzia Spaziale Italiana, launched in 2020) and EnMAP (Environmental Mapping and Analysis Program) by DLR (launched in 2023) are orbiting in a sun-synchronous orbit at a height between 600 and 700 km.

For this study, we have obtained high-quality (cloud, haze and dust-free scenes) from three satellite-based HS sensors, EnMAP, EMIT, and PRISMA (Figure 1). These sensors (

Table 1. Relevant sensor specifications of the hyperspectral datasets used in the study.

		EnMAP	EMIT	PRISMA	HyMap
Spatial resolution (in m)		30	60	30	5
Spectral sampling (in nm)	VNIR	6.5	7.4	12	15
	SWIR	10	7.4	12	13–17
Spectral range (in nm)	VNIR	420–1000	381–2493	400–1010	450–1350
	SWIR	900–2450		920–2505	1400–2480
Spectral bands	VNIR	101	285	66	125
	SWIR	123		171

) were selected due to their data availability for our targeted sites, and because they are representative of the new generation of HS data.

Unfortunately, the data from these sensors were not acquired at the same time, nor were additional airborne hyperspectral (HyMap) data that we have also included for the two study sites. However, while the abundance and composition of soil or vegetation are susceptible to significant seasonal changes and weather fluctuations, geological features are not expected to change significantly between seasons and years, especially in arid regions with large outcropping areas. Soil moisture remains an uncertainty factor; however, for the main analyses, we have removed the water bands, which should reduce its effect on the results. Thus, with some caveats and uncertainties, we have obtained satellite datasets based on their quality (minimal cloud cover, haze and dust) rather than minimizing temporal separation. As a result, there are several year gaps between the different datasets.

All remotely sensed hyperspectral data, including satellite and airborne platforms, require calibration and correction to convert the measured digital numbers to at-sensor radiance, and then we must remove atmospheric and topographic effects to estimate surface reflectance [23]. These pre-processing and correction workflows can be very complex and are a significant source of uncertainty that can greatly influence geological mapping results (e.g., [21,24,25]). Conveniently, most satellite data vendors (e.g., the EMIT, PRISMA and EnMAP teams) now provide corrected datasets, which are (supposedly) ready for analysis. We have used these vendor-corrected data (cf. Section 3.1) for our comparisons, as we consider that (1) these are the most likely data to be used by general practitioners and (2) differences in the applied pre-processing and radiometric correction workflows are likely to result in substantially different reflectance spectra, for which we would like to understand the implications for geological mapping.

2. Geology of the Study Area

Namibia has a generally arid environment with limited vegetation and extensive bedrock exposure. Combined with relatively little seasonal variation and low population density, this makes it an ideal location for testing geological remote sensing techniques. Moreover, Namibia hosts significant critical raw materials needed for the global shift towards green energy, providing geologically interesting remote sensing targets including REE bearing carbonatites [26,27]. We have selected two carbonatite complexes for our comparison study: Marinkas-Quellen and Epembe [28]. A combination of HS data availability and extensive prior study [29,30] made these two carbonatite sites best suited for this study.

Carbonatites are geochemically anomalous igneous rocks containing >50% carbonate minerals [31,32], generally associated with intraplate rift settings [32]. As highly fractionated and/or very-low-degree partial melts, carbonatite magmas are known to host economic abundances of rare earth elements (REEs) [32].

Carbonatites are generally divided into the following categories based on their chemistry: calcio-carbonatite (calcite-rich), magnesio-carbonatite (dolomite-rich) and ferroan-carbonatite (siderite- or ankerite-rich) [31,32]. This primary mineralogy is often accompanied by fenitization, a type of alkali metasomatism [29,32,33]. Importantly for our study, these mineralogical variations are theoretically detectable using VNIR and SWIR hyperspectral data [11,34,35].

REEs are also known to exhibit unique spectral absorptions in the VNIR range [35], due to electronic field transitions [36,37], and some less prominent absorptions have been reported in the SWIR range as well [11]. The strength and type of REE absorption depend on the spatial coverage, the host mineral and its concentration ratio. However, even in carbonatites, REE absorptions are typically too subtle to be identified at airborne and satellite scales, pertaining to their larger ground sampling distances. Thus associated mineralogy needs to be used as a proxy to estimate potential for REE mineralization.

2.1. Marinkas-Quellen

Marinkas-Quellen is located in Southern Namibia, near the border with South Africa. The carbonatite complex here is Cambro-Ordovician in age and characterized by sequential intrusions of silica- and carbonate-rich magmas [30], which cooled to form nepheline syenite and carbonatite, respectively (Figure 2B). The carbonatites include calcio, magnesio, and ferroan varieties (Figure 2B) [26,30]. The late ferroan-carbonatite dykes are known to be rich in REEs, Mn and Th, while both the calcio- and the magnesio-carbonatite serve as hosts for the heavy REEs [30].

2.2. Epembe

The carbonatite at Epembe covers an area of 7 × 0.4 km. Unlike Marinkas-Quellen, this carbonatite is mainly composed of coarse-grained calcio-carbonatite (søvite) (Figure 2A) with very narrow zones of magnesio-carbonatite [29,32]. It is Mesoproterozoic in age (1216 ± 2.4 Ma, [38]) and is emplaced within a sub-vertical NW-SE trending dextral regional shear zone [39]. The modal mineralogy of the Epembe carbonatite is characterized by coarse calcite (with minor dolomite and ankerite) and accessory minerals, including relicts of K-feldspar, aegirine, apatite, and pyrochlore [40]. The carbonatite complex is surrounded by an asymmetric fenite alteration halo [29].

3. Methods

3.1. Data Preparation and Correction

Satellite HS data providers generally provide data with different levels of preprocessing, including 1. geometric and sensor-corrected data (level L0); 2. radiometrically calibrated top-of-atmosphere radiance (level L1); and 3. atmospherically corrected surface reflectance estimates (level L2). As most geological mapping applications now use vendor-corrected reflectance estimates, we have used level L2 data (

Table 2. Details of the hyperspectral scenes used in the study.

		EnMAP	EMIT	PRISMA	HyMap
Processing level		L2A	L2A RFL	L2D	Reflectance data
Acquisition date	Marinkas Quellen	17 May 2023	7 January 2023	21 October 2020	26 February 2018
	Epembe	17 April 2023	3 May 2023	27 November 2022	6 August 2014
Acquisition time (local)	Marinkas Quellen	11:29	10:24	11:05	14:20
	Epembe	11:49	12:01	11:27	11:52

) for our comparison study. It is thus important to note that our comparison assesses both sensor performance and inter-sensor inconsistencies stemming from inaccuracies in their respective pre-processing workflows (conversion from recorded digital numbers to on-ground estimated reflectance spectra). These different pre-processing workflows are briefly summarized below.

3.1.1. PRISMA

The L1 radiance data generation runs a flat field correction, dark subtraction and a key quality check algorithm to generate and update the radiometric, spectral and geometrical parameters of the scene. Subsequently, the L2 processor utilizes the L1 top-of-atmosphere radiance data alongside the panchromatic band as inputs to generate the ultimate atmospherically corrected surface reflectance data [41].

3.1.2. EnMAP

The Python-based Atmospheric Correction module (PACO [42,43]) is developed by DLR and is employed to derive the EnMAP L2 data (which we used in this study). Although PACO is largely based on the ATCOR-IDL code [44], it is known for its unique and rigorous calibration and validation procedure [42]. The model utilizes sensor configuration parameters, such as the sun zenith angle and observation off-nadir angle, to automatically determine the biome density, ozone level, and the season corresponding to the scene. These parameters are then applied for ozone, cirrus, and haze correction. The process generates distinct outputs, including image masks, aerosol optical thickness, wavelength maps, and bottom-of-atmosphere reflectance images [42]. The ultimate L2A reflectance product is derived from the combination of these outputs. EnMAP is also unique in providing a range of acquisition parameters for the user to choose from, such as view angles, which can be important in areas with significant topographic variation.

3.1.3. EMIT

EMIT signifies a potentially important leap in HS imaging, using a pioneering detector that captures the entire VNIR-SWIR spectrum on a single sensor [21]. This technology is earmarked for upcoming NASA missions like the SBG (surface biology and geology) mission [45] and could help to mitigate discrepancies between the VNIR and SWIR ranges. EMIT uses the radiative transfer modelling tools for its final reflectance product as well as for its radiometric calibration to ensure consistency in assumptions and reduce model-induced errors [21]. It also uses the bright cloud features in a scene to monitor its calibration for the shorter wavelengths (VNIR). Further, it employs cross-calibration using other on-orbit instruments to assess the absolute accuracy of the preset calibration protocols [21,25]. This novel pre-processing workflow of EMIT ensures consistent product quality throughout its lifespan [21]. With a spatial resolution of 60 m, EMIT adequately fulfils its large-scale scientific mission objectives or mineral and dust flow mapping [21]. EMIT’s deployment in low-earth orbit aboard the ISS also offers opportunities for data acquisition at high temporal frequency.

3.1.4. HyMap

The airborne HS datasets were acquired using the HyMap sensor from a height of 2000 m, resulting in 5 m spatial sampling. Geometric and radiometric corrections were conducted by the HyVista Corporation, which acquired the data. Atmospheric correction was performed using a continental aerosol model and a mid-latitude summer atmospheric model, with an ozone count of 340 ppm and a 75 km visibility, to estimate ground reflectance.

It is important to note that the HyMap scene from Epembe was received with vegetation patches masked out. However, in the other scenes, no such masking was employed, as at 30m and 60m resolutions, most pixels contain some vegetation.

3.2. Data Co-Registration

Co-registration is a crucial step for data integration and comparison. The hyperspectral data from each sensor at both sites were spatially subset to the respective regions of overlap (Figure 3). These subset scenes were then co-registered to the HyMap dataset using the python package Automated and Robust Open-Source Image Co-Registration Software (AROSICS) 1.11.0 [46]. This is a Python-based toolbox that automatically identifies comparable bands between two datasets (using a Fourier transform) and then automatically calculates tie points that can be used to co-register. This co-registration was applied using a global affine transform [46], which was applied (to resample the data onto the same grid) using Rasterio [47], and scikit-image [48]. Note that this regridding was performed as late as possible in our workflow, such that hyperspectral data were kept in their original form until a final resampling of the analysis results, as described later, was carried out to minimize the influence of this regridding on our results.

3.3. Spectral Analyses

Geological mapping via HS remote sensing has a rich history. Recent strides in machine learning have led to the use of more sophisticated models for mapping geological patterns and lithological variations [49,50]. However, as this study focuses on comparative analysis, we apply relatively simpler but well-established techniques that facilitate comparisons between sensors. Specifically, we focused on spectral indices, known for their robustness in mineral-specific mapping, and linear spectral unmixing (using manually selected “pure pixel” endmember spectra). All data analyses and processing were carried out using hylite 1.23 [51] and scikit-image 0.22.0 [48].

3.3.1. Spectral Index Analysis

Spectral indices, arithmetic combinations of carefully selected bands, are a robust method for mineral identification and mapping and have been widely used for multispectral and hyperspectral data analysis [9,52]. The fundamental understanding is that dividing one spectral band by another produces an image with relative insensitivity to many common detrimental effects (e.g., residual illumination and atmospheric signatures), as these can largely cancel out during the division. Bands are normally selected such that the denominator represents the reflectance at the minima of a target absorption feature, while the numerator is a function of the reflectance at one or both adjacent shoulders [9,53]. This helps in discriminating mineral types, by suppressing errors that are relatively constant between adjacent bands while enhancing local variations associated with mineral absorptions [54,55].

For our comparisons, we have applied several established spectral indices to map the spectrally relevant phases in both study sites. At Epembe, the spectral indices delineate the calcite-rich carbonatite dyke from the surrounding altered country rock (spectrally characterized by kaolinite and hematite-dominated spectra). For Marinkas-Quellen, spectral indices were employed to map the extent of calcite-bearing and dolomite-bearing carbonatite lithologies. The formulae used, and their sources, are described in Table 3.

Many of these formulae were originally defined for multispectral sensors (e.g., ASTER) and have since been re-defined and updated by various authors [54,55,56] to suit hyperspectral satellite data (e.g., PRISMA). These formulae were kept consistent for all the sensors, to enable a fair comparison between them. Also note that most of the formulae use an average of many bands, which intentionally reduces the influence of sensor-specific spectral shifts or sampling differences, allowing for fairer comparisons than with more highly band-specific indices (potentially at the expense of discriminating power—however, that is not our aim here).

Table 3. Spectral indices formulae were used in this study. Note that the ∗ operator represents multiplication, and the operator represents the averaging of all bands between the specified wavelengths. All values represent wavelengths in nanometers.

Mineral Proxy	Study Site	Target Absorption Position	Spectral Index	Reference
Calcite	Marinkas-Quellen and Epembe	2337 nm	$\frac{2190:2224}{2293:2345}\ast \frac{2375:2430}{2293:2345}$	[54,55]
Hematite	Epembe	570 nm	$\frac{600:630}{550:580}$	[57]
Kaolinite	Epembe	2195 nm, 2225 nm	$\frac{1600:1700}{2145:2185}\ast \frac{2295:2365}{2185:2225}$	[54,55]
Dolomite	Marinkas-Quellen	2320 nm	$\frac{2298 + 2342}{2320}$	[54]

Table 3. Spectral indices formulae were used in this study. Note that the ∗ operator represents multiplication, and the operator represents the averaging of all bands between the specified wavelengths. All values represent wavelengths in nanometers.

Mineral Proxy	Study Site	Target Absorption Position	Spectral Index	Reference
Calcite	Marinkas-Quellen and Epembe	2337 nm	$\frac{2190:2224}{2293:2345}\ast \frac{2375:2430}{2293:2345}$	[54,55]
Hematite	Epembe	570 nm	$\frac{600:630}{550:580}$	[57]
Kaolinite	Epembe	2195 nm, 2225 nm	$\frac{1600:1700}{2145:2185}\ast \frac{2295:2365}{2185:2225}$	[54,55]
Dolomite	Marinkas-Quellen	2320 nm	$\frac{2298 + 2342}{2320}$	[54]

3.3.2. Spectral Abundance Mapping

Spectral unmixing is based on the premise that the signal from a hyperspectral pixel is an abundance-weighted (linear) combination of pure material endmember spectra [58,59]. Endmembers can be identified if (near) pure pixels exist anywhere in a scene and used to estimate fractional abundances using, e.g., least-squares non-negative unmixing [60,61]. These fractional abundances can then be represented as a spectral abundance map representing (in a low-dimensionality space) the latent variables that determine hyperspectral variations within the scene (i.e., the spatial distribution of materials in the scene) [54,61]. A latent variable is an unobservable quantity that represents hidden aspects of the data, inferred indirectly through mathematical modelling based on observable variables.

This approach has some limitations: pure pixels can be uncommon in satellite data due to their coarse spatial resolution, while some materials also mix in scale-dependent (non-linear) ways [59]. However, the sparse representation of an HS image as weighted combinations of selected endmember pixels does present an interesting way to compare latent information between sensors. Hence, we have manually defined fixed endmember pixel locations that we consider (based on geological mapping and spectra from the EnMAP sensor) to be representative “pure” spectra for each of the geological classes of interest and extracted corresponding endmember spectra libraries by sampling these locations for each sensor. We used EnMAP as a reference for selecting endmembers at both scenes because it matched closely with the in-situ measurement spectra compared to the other sensors. Non-negative linear unmixing was then employed to derive spectral abundance maps using pysptools 0.15.0 [62].

Given the different hyperspectral responses of the target endmembers at Marikas and Epembe, we applied this unmixing to different spectral subsets. At Epembe, bands between 450 and 1150 nm, and 2000 and 2480 nm, were considered (due to the presence of hematite absorptions in the VNIR and carbonate absorptions in the SWIR), while at Marinkas, only bands between 2000 and 2480 nm were considered (as carbonates are the main minerals of interest here). Note also that water absorption bands were intentionally excluded from the range used for the unmixing, due to typically high noise within these spectral ranges in remotely sensed data.

3.4. Comparing Sensors by Quantifying Hyperspectral Consistency

Meaningfully comparing hyperspectral data from different sensors is non-trivial as different sensors, acquisition methods and correction techniques can result in systematic biases in, e.g., estimated reflectance. Furthermore, many applications use analyses that are tuned to very specific spectral properties, rather than entire hyperspectral data cubes. It is thus necessary to perform both “broad” comparisons that compare wide spectral ranges and “narrow” comparisons that focus on application-specific spectral ranges and analysis approaches.

Given these challenges, we have formulated a set of four consistency metrics that range from spectrally broad comparisons to specific ones related to, e.g., a single spectral index. To ensure interpretability and maintain consistency across these metrics, we use the widely applied score for these comparisons. Note that these R² scores reflect only the consistency between two observed spectral datasets, not their accuracy, because no ground truth is known or assumed.

We first compute $R^2$ scores using the mean-square deviations between the pairs of reflectance values (one from each sensor, across all bands). This is consistently expressed relative to the sensor with the highest standard deviation (Equation (1)) and quantifies the deviation of the measured reflectance pairs from a 1:1 line. Values thus range from 1 (perfectly consistent data with a 1:1 relationship) to 0 (inter-sensor differences are equal to the intra-sensor variance, i.e., no correlation) and negative infinity (inter-sensor differences are arbitrarily greater than the intra-sensor variance).

$${R_{abs}}^2=1-\frac{(senso{r}_1-senso{r}_2{)}^2}{max\left(var\right(senso{r}_1), var(senso{r}_2\left)\right)}$$

(1)

For many applications, this absolute consistency metric is overly conservative because processing methods are often sensitive only to relative spectral variations (e.g., hull corrections will remove variations in absolute reflectance while amplifying local minima that result from absorption features). Hence, our second-order consistency metric ${R_{prop}}^2$ is computed as above but using residuals to the best-fit linear regression between two sensors rather than the 1:1 line. This metric is insensitive to some types of systematic bias (i.e., those that preserve linear relationships), and results in high scores so long as spectral variations are proportionally consistent between the sensors. This metric is equal to the $R^2$ score (coefficient of determination) of a linear regression between each pair of reflectance measurements, so it ranges from 0 (no correlation) to 1 (perfect correlation). It can be calculated using the sum of squared regression residuals (SSR) and total sum of the squared variation of the chosen dependent variable (SST). Note that the choice of dependent and independent variables in the linear regression does not affect the score.

$${R_{prop}}^2=\frac{SSR}{SST}$$

(2)

To compute the absolute (R_abs²) and relative (R_prop²) consistency metrics between the sensors, we resampled all of the reflectance data to match the coarsest sensor resolution (EMIT, with 60 m pixels), using a spatially averaged downsampling consistent with an assumption of approximate linear mixing. Spectrally, the data were downsampled to match the spectral resolution of the HyMap sensor (~125 bands).

Both of these metrics are sensitive to the entire spectral range captured by a sensor (or selected subset thereof, e.g., we have removed bands affected by atmospheric water absorbtions, as mentioned previously). However, many spectral analyses including, e.g., spectral indices, use only a few (2–6) bands, meaning most of the spectral variability within and between sensors will have no influence on the results. Our third consistency metric is thus application-specific and relies on the selection of one or more relevant spectral indices (sensitive to, e.g., carbonate mineral abundance). The results of these spectral indices can also be compared using the R² score (relative to the 1:1 line), such that our third consistency metric, R_feat², can be derived by substituting band ratio results into Equation (1) instead of reflectance estimates.

Lastly, some analyses (e.g., unmixing) express spectra not in absolute terms but as linear mixtures of specific pixels (i.e., endmembers) in the dataset(s). These latent features can be consistent between two sensors even when very significant spectral distortions or biases are present (as can be introduced by, e.g., over- or under-compensating atmospheric corrections), provided they equally affect each pixel in an image. This latent feature comparison is also application-specific and relies on the selection of a set of endmember pixels with consistent locations between the two datasets. As described previously, the spectra from these are then used to linearly unmix (using non-negative least squares; [51]) the respective images, and deviations of the resulting spectral abundances (A) from the best-fit linear regression are used to derive an ${R_{lat}}^2$ score that ranges from 0 to 1 by substituting spectral abundances into Equation (2) instead of reflectance estimates.

To summarize, we have computed a set of four $R^2$-score-based comparison metrics for each pair of sensors investigated in this study. ${R_{abs}}^2$ is interpreted as a measure for the consistency of the coregistered reflectance measurements (across all bands), while ${R_{prop}}^2$ measures this spectral consistency while allowing for some (linear) biases and offsets. ${R_{feat}}^2$ and ${R_{lat}}^2$ are used as context-specific comparisons that test, respectively, the consistency of an application-specific band ratio and the consistency of latent spectral abundances (that likely reflect the distribution of specific materials of interest).

Note that we carried out spectral analyses (spectral index analysis and linear unmixing) at the inherent spatial resolutions of the HS scene, and then resampled the results to 60 m for computing the R²_feat and R²_lat metrics.

4. Results and Interpretation

4.1. Qualitative Comparison

Selected rock samples from the Marinkas-Quellen and Epembe carbonatites were spectrally analyzed using a Specim AisaFENIX mounted in a SiSuROCK, developed by Specim, Spectral Imaging Ltd., Oulu, Finland, core scanner, to characterize their spectral response under laboratory conditions. The resulting ~1 mm spatial sampling spectra (from the weathered face of hand samples) show subtle but distinct neodymium (Nd) absorptions at 580 nm, 750 nm and 800 nm, as well as pronounced carbonate absorption at 2337 nm (Figure 4B).

However, in this case, the hyperspectral data corresponding to these ground location pixels at airborne (5 m spatial sampling) and satellite (30 m spatial sampling) only show the carbonate (here calcite) absorption, with no REE signatures apparent (Figure 4). At these larger ground sampling distances, the Nd absorptions are likely too diluted (by non-REE-bearing phases) to be detectable, unlike the relatively more pronounced and spatially more abundant carbonate feature.

Endmember spectra for both study sites were selected using a combination of N-finder [63] on the EnMAP dataset (to identify relatively pure pixels) and geological knowledge of the area (as described in Section 3), and then they were kept consistent for comparison between the datasets (i.e., the same pixel coordinates and corresponding on-ground area were used to derive sensor-specific endmember spectra from each dataset).

The carbonatites at Marinkas-Quellen exhibit distinct carbonate absorptions in all of the datasets (except the PRISMA data; Figure 5), with spatial variations in absorption minima wavelength corresponding to dolomite (2321 nm) and calcite (2337 nm) [53] matching known geological variations (Figure 5A,B) [14]. The signature absorptions of dolomite and calcite are discernible in the endmembers obtained from the HyMap, EnMAP, and EMIT scenes. In the endmembers obtained from the PRISMA scene, however, these absorptions are obscured by noise or artefacts, likely from the atmospheric correction (Figure 5).

A third endmember was also selected to represent the granitic country rock (Figure 2). While the identifiable signature absorptions of the primary minerals like quartz and feldspar exist beyond the SWIR range, this endmember is characterized by white mica and clay absorptions due to surficial weathering [26]. These three endmembers, loosely referred to as dolomite, calcite, and country rock, were used for the spectral unmixing analysis.

Unlike Marinkas-Quellen, the Epembe carbonatite is mainly calcitic and is spatially associated with syenite and fenite intrusions (Figure 2). These can be spectrally identified by the presence of the Mg-OH absorption at 2300–2339 nm [29], but the partial overlap with CO₃²⁻ features makes it challenging to identify in mixtures with calcite (due to the carbonate absorption at 2337 nm). Extensive alteration and near-surface weathering have resulted in high clay and iron contents in the granitic host rock, adding further spectral complexity relative to Marinkas-Quellen [29]. Thus, to separate the Epembe carbonatite from the surrounding altered host rock in the study areas, a kaolinite phase (minimum at 2205 nm) and a hematite phase (characterized by an inflection between 570 and 600 nm) were identified as the characteristic mineral phases used in the following analyses.

The signature inflection point around 570–600 nm, which is attributed to the presence of hematite, remains consistent across all sensors (Figure 6A). EnMAP and EMIT showed larger absorption depths for both kaolinite and calcite thanHyMap. As for PRISMA, the kaolinite signature appears weak, and the calcite absorption feature at 2337 nm appears entirely absent (Figure 6B,C). Figure 6 further illustrates a considerable variation in spectral responses from different sensors, especially in the relative depths of the target absorption features.

Lastly, we also note that the Epembe HyMap scene exhibits a ~5 nm shift in band position relative to the Marinkas-Quellen HyMap scene, likely due to the recalibration of the instrument between surveys (highlighting challenges comparing even data from the same sensor). This results in non-trivial differences in the sampling of the carbonate absorption feature, apparently degrading the ability of this sensor to precisely identify carbonate absorptions at Epembe.

4.2. Absolute and Relative Consistency

Absolute reflectance estimates from EnMap, EMIT and HyMap are relatively comparable (${R_{abs}}^2$ = 0.2–0.7), while the PRISMA results show no correlation (negative ${R_{abs}}^2$ scores) with the other sensors. The cross-plots (Figure 7) suggest that this lack of correlation results from significant deviations in the SWIR range. Additionally, the sensors exhibiting the strongest correlation vary between the two sites. At Marinkas-Quellen, HyMap and EnMAP displayed both highest relative (${R_{prop}}^2$ = 0.87), and absolute (${R_{abs}}^2$ = 0.76) consistency. At Epembe, EnMAP and EMIT show the greatest absolute and relative consistencies (${{{R_{abs}}^2=-2.88, R}_{prop}}^2$ = 0.71). Note that the relative consistency will, by definition, be higher than the absolute consistency.

4.3. Feature Consistency

A feature consistency analysis, as described in Section 3.4, was performed to compare spectral indices known to be sensitive to the depth of the carbonate absorption position, ${\mathrm{F}\mathrm{e}}^{3+}$ absorption (associated with iron oxides in the regolith), and the kaolinite doublet absorption (Table 3). Visual inspection of the spectral index maps from Marinkas-Quellen suggests that absorption features associated with calcite and dolomite-bearing carbonatite are aligned with the known geology (Figure in Supplementary, Figure S1). For visual comparison, the results of spectral indices from the Epembe site have been plotted as false-color ternary images (Figure 8). These show that the targeted carbonatite dyke can be delineated using data from any of the sensors. However, the level of clarity (sharpness and distinctiveness of the dyke) varies depending on the sensor used and its inherent spatial and spectral resolutions.

The ${R_{feat}}^2$ metrics calculated for all of the spectral indices at both study sites are presented in Figure 9. At Marinkas-Quellen, the comparison among HyMap, EnMAP, and PRISMA yielded an ${R_{feat}}^2$ of 0.90 for calcite, but dolomite yielded lower values (−3.8 to 0.62). At Epembe, the hematite index produced the most consistent results, followed by calcite. EMIT and EnMAP, which exhibit the most prominent kaolinite absorption (Figure 6D), produced an ${R_{feat}}^2$ of 0.32, higher than all other sensor pairs for this band ratio.

4.4. Latent Consistency

Unlike the locally sensitive indices presented above, which are tailored to specific absorption features, spectral abundances are generated by unmixing entire spectra. However, unlike the comparisons of absolute reflectance, these abundances should reflect the latent features controlling spatial variation within the dataset, which ideally correlate with the geology. Even severe spectral distortions (such as those from, e.g., errors in atmospheric correction) should not influence these latent features, so long as they are uniformly present. They will, however, be sensitive to potentially subtle spatial variations induced by variations in geology, vegetation or illumination (e.g., topographic effects).

The spectral abundance maps (Figure 10) generated for Epembe reveal intriguing disparities compared to the spectral index outputs (Figure 9). Significantly, the distinct spectral features of the targeted minerals observed in the EnMAP and EMIT scenes appear to result in a precise delineation of the carbonatite dyke that aligns well with the regional geological map [28] (see Figure S2). In contrast, the HyMap and PRISMA data show a more scattered and less well-defined geometry (refer to Supplementary Figure S2).

At Marinkas-Quellen, the spectral abundance maps (Figure 10) align with the geological map, with some notable differences (Figure 3B). The dolomite-rich carbonatite matches well with the geological maps, while the spatial complexity of the calcite-rich carbonatite is captured in varying levels of clarity depending on the sensors. The most consistent results for the dolomite-rich and calcite-rich carbonatites were derived from HyMap and EnMAP (Figure 6A,B). Interestingly, PRISMA also produces comparable results, despite difficulties interpreting the endmember spectra (Figure 6D), that closely match the geological maps and result in high ${R_{lat}}^2$ scores when compared with the other sensors (Figure 11).

The results from EMIT also match the geological map [28] (Figure 10C), but the low spatial resolution is insufficient to capture some of the intricate geometries within each carbonatite body. As at Marinkas-Quellen, the abundance map derived from the PRISMA scene exhibits a notable spatial similarity with the EnMAP and HyMAP results, despite the apparently degraded endmember spectra (Figure 10D).

5. Discussion

The objective of the study was to compare the consistency of new-generation satellite HS sensors and a high-resolution airborne HS dataset in the context of geological mapping and mineral exploration applications. We show that there are significant disparities among the HyMap, EnMap, PRISMA and EMIT datasets we have investigated (Figure 7), most notably in the SWIR spectral range. Importantly, the relative consistency we observe between these sensors in the VNIR range suggests that these differences cannot be attributed to temporal effects (as these would be expected to largely manifest in the VNIR range, due to its sensitivity to vegetation). Furthermore, our consistency results do not show any relationship with acquisition time, with, e.g., HyMap data (the oldest used dataset) showing generally high consistency scores, or with, e.g., EnMAP (the newest datasets).

PRISMA stands out among the satellite sensors used in this study due to its comparatively extended operational history (since 2019) and coverage (EnMAP has only acquired data since late 2022 and EMIT since early 2023). However, we find that PRISMA spectra tend to be inconsistent with those derived from the other similar platforms (EMIT, EnMAP and HyMap), most notably in the SWIR range (Figure 7). These differences were sufficient to derive inconsistent mineral mapping results using spectral-indices methods (Figure 9), though significantly, they did not seriously affect the comparison of latent features via spectral unmixing (Figure 10 and Figure 11). Based on this, we suggest that the inconsistent results produced by PRISMA result from the vendor’s radiometric and atmospheric correction routines, which appear to severely distort spectra in the SWIR range (Figure 6 and Figure 10). This interpretation is supported by multiple other studies that have highlighted the superior data quality of PRISMA VNIR relative to SWIR spectra, especially in urban classification and vegetation-focused research [64,65]. That said, studies such as [54,56] successfully used the spectral index to map mineral phases like iron-oxyhydroxide, kaolinite, and calcite in a PRISMA L2 scene after employing additional corrections, including cross-track illumination corrections for VNIR and SWIR separately and heavy spectral smoothing [54,56]. The comparatively high R²_feat scores using spectral indices reinforce these results; spectral indices, especially those covering relatively small wavelength ranges, appear to be quite robust.

Our results (Figure 5 and Figure 6) clearly show that the EnMap and EMIT datasets best capture the known diagnostic absorption features in the SWIR range that are typically used for geological mapping. Comparatively, these features are obscured or distorted in the PRISMA (and to some extent in HyMap due to its coarser spectral-sampling) scenes (Figure 5 and Figure 6), resulting in, e.g., spectral indices that are less consistent with the geological maps and, as shown by our quantitative comparisons, with results from EnMap and EMIT (Figure 8, Figure 9 and Figure 10). As a result, we suggest that EnMAP consistently yielded the best results for geological mapping purposes, at least at our two study sites. Results from the EMIT sensor also have a high spectral quality but at a lower spatial resolution.

6. Conclusions

Using two test sites at arid locations in Namibia (Epembe and Marinkas-Quellen), we have tested and compared three new-generation satellite-based HS data sources—PRISMA, EnMAP, and EMIT—as well as airborne HyMap data. In the process, we develop and present a novel approach for qualitatively and quantitatively assessing the absolute, relative and application-specific consistency of spectra acquired using different platforms, with an emphasis on potential applications to geological and mineralogical mapping. Our results suggest that inconsistent results can be produced by applying the same analyses to data from different sensors. This is attributed to differences in radiometric and atmospheric correction approaches by the data providers and varying ground sampling distances. Most of the observed inconsistencies are present in the SWIR spectral range, which is particularly relevant to geological mapping applications. We conclude that all of the tested sensors could be used for geological mapping, but that PRISMA data likely require additional radiometric corrections to derive interpretable spectra, while EMIT’s lower resolution limits its ability to resolve smaller (10–100 m scale) geological variations.