Towards Multiscale and Multisource Remote Sensing Mineral Exploration Using RPAS: A Case Study in the Lofdal Carbonatite-Hosted REE Deposit, Namibia

Traditional exploration techniques usually rely on extensive field work supported by geophysical ground surveying. However, this approach can be limited by several factors such as field accessibility, financial cost, area size, climate, and public disapproval. We recommend the use of multiscale hyperspectral remote sensing to mitigate the disadvantages of traditional exploration techniques. The proposed workflow analyzes a possible target at different levels of spatial detail. This method is particularly beneficial in inaccessible and remote areas with little infrastructure, because it allows for a systematic, dense and generally noninvasive surveying. After a satellite regional reconnaissance, a target is characterized in more detail by plane-based hyperspectral mapping. Subsequently, Remotely Piloted Aircraft System (RPAS)-mounted hyperspectral sensors are deployed on selected regions of interest to provide a higher level of spatial detail. All hyperspectral data are corrected for radiometric and geometric distortions. End-member modeling and classification techniques are used for rapid and accurate lithological mapping. Validation is performed via field spectroscopy and portable XRF as well as laboratory geochemical and spectral analyses. The resulting spectral data products quickly provide relevant information on outcropping lithologies for the field teams. We show that the multiscale approach allows defining the promising areas that are further refined using RPAS-based hyperspectral imaging. We further argue that the addition of RPAS-based hyperspectral data can improve the detail of field mapping in mineral exploration, by bridging the resolution gap between airplaneand ground-based data. RPAS-based measurements can supplement and direct geological observation rapidly in the field and therefore allow better integration with in situ ground investigations. We demonstrate the efficiency of the proposed approach at the Lofdal Carbonatite Complex in Namibia, which has been previously subjected to rare earth elements exploration. The deposit is located in a remote environment and characterized by difficult terrain which limits ground surveys.


Introduction
The increasing demand for raw materials worldwide conflicts, with the growing difficulties to find new mineral deposits, outlines the need for innovative approaches in mineral exploration. Technological advancements, such as green technologies and smart devices, have led in particular to an

Geological Setting: Demonstration on the Lofdal Deposit
The Lofdal area was chosen as demonstrator, but is only one of the localities identified on remote sensing data. We focus on Lofdal because it is the area which has been studied most.
The Lofdal Alkaline Carbonatite Complex, aged~750 Ma [32], is located on the Bergville and Lofdal farms, 35 km NW from Khorixas in the Kunene region of Namibia. The Paleo-Proterozoic Huab Basement Complex hosts the alkaline complex and is made up of metasedimentary gneiss and schist. The geometry is composed of a plug-like syenite-carbonatite intrusion known as the Main intrusion with a smaller plug-like body, called the Emanya intrusion, 5 km to the SW. The complex contains a swarm of roughly E-W striking carbonatite dykes as well as smaller plugs of calcite carbonatite [32] ( Figure 1). These carbonatite plugs and dykes combined with the occurrence of phonolite, syenite, and mafic dykes, and plugs form together the alkaline intrusive complex [33]. Jung et al. [34] has proposed a two-stage model of emplacement based on isotopic compositions to explain the syenite-carbonatite suite intrusion. The first stage is partial melting of the upper-mantle followed by the second stage of differentiation of unexposed alkali-basaltic magma.
The Lofdal Alkaline Carbonatite Complex is unique in its unusually high enrichment of HREEs compared to the LREEs [35]. Xenotime-(Y) is the principal HREE host mineral and generally occurs in the iron-rich calcite carbonatite dykes. The xenotime is associated with iron oxides, thorite, apatite and synchysite-(Ce) [35]. Although the genesis of REE mineralization at Lofdal is still debated, it has been suggested lately that HREE enrichment is likely related to hydrothermal processes [36].
Several areas with high potential for REE explorations have been identified by the exploration team in the past. Such an area, known as Area 4, which lies~1 km SE from the main intrusion, has received much attention, and has been studied extensively over the years [29,30]. However, for this study we focus on Area 1, a relatively understudied area compared with Area 4, but very typical of the entire site in terms of mineralogy. Area 1 is located North-East of the main intrusion and consists of dozens of carbonatite dykes, some of which cross-cut each other (Figure 1). The dykes are tens of meters long and are 0.5 cm to 1 m wide. This characteristic geological setting is used to evaluate the benefits of RPAS for the geological mapping of complex and small-scaled structures.
Remote Sens. 2019, 11, x FOR PEER REVIEW 4 of 30 Several areas with high potential for REE explorations have been identified by the exploration team in the past. Such an area, known as Area 4, which lies ~1 km SE from the main intrusion, has received much attention, and has been studied extensively over the years [29,30]. However, for this study we focus on Area 1, a relatively understudied area compared with Area 4, but very typical of the entire site in terms of mineralogy. Area 1 is located North-East of the main intrusion and consists of dozens of carbonatite dykes, some of which cross-cut each other (Figure 1). The dykes are tens of meters long and are 0.5 cm to 1 m wide. This characteristic geological setting is used to evaluate the benefits of RPAS for the geological mapping of complex and small-scaled structures.

General Workflow
The proposed approach uses data captured from platforms with decreasing altitudes (Figure 2) and envisages a holistic workflow taking into account local characteristics, such as vegetation cover and geology, as well as data availability. This is especially beneficial as the resolution of the data being captured increases systematically ( Figure 3). Dedicated analyses were performed on the data at varying resolution levels and are summarized in Figure 4. Each analysis is explained in detail in section "Processing and field validation".

General Workflow
The proposed approach uses data captured from platforms with decreasing altitudes (Figure 2) and envisages a holistic workflow taking into account local characteristics, such as vegetation cover and geology, as well as data availability. This is especially beneficial as the resolution of the data being captured increases systematically ( Figure 3). Dedicated analyses were performed on the data at varying resolution levels and are summarized in Figure 4. Each analysis is explained in detail in section "Processing and field validation".
In the first step, we applied commonly used spectral analyses, known for their versatility and robustness, on satellite-and aircraft-based imagery. We chose ASTER data for the first level of resolution and performed lithological mapping and area identification (top section of Figure 4). We mapped the main lithologies of the region by performing selected band ratios. The next level of resolution is provided by plane-based imaging. We performed band ratios and minimum wavelength mapping (MWM) on HyMap imagery and used the results to refine an area of interest (Top section of Figure 4). Once we chose and refined the area of interest, we moved to the next resolution level using RPAS-based sensors. RPAS-based sensors can capture the spectral and topographical data needed to study small scale features (middle and bottom sections of Figure 4).  In the first step, we applied commonly used spectral analyses, known for their versatility and robustness, on satellite-and aircraft-based imagery. We chose ASTER data for the first level of resolution and performed lithological mapping and area identification (top section of Figure 4). We mapped the main lithologies of the region by performing selected band ratios. The next level of resolution is provided by plane-based imaging. We performed band ratios and minimum wavelength mapping (MWM) on HyMap imagery and used the results to refine an area of interest (Top section of Figure 4). Once we chose and refined the area of interest, we moved to the next resolution level using RPAS-based sensors. RPAS-based sensors can capture the spectral and topographical data needed to study small scale features (middle and bottom sections of Figure 4).
Field scouting backed by literature show that the REE occurrences are hosted in carbonatite dykes in the study area [33]. Field observation showed that these structures are topographically elevated in a relatively flat surrounding terrain. Therefore, the best field approach was to capture RPAS-based data over the area and to use both spectral and topographic information in parallel.  In the first step, we applied commonly used spectral analyses, known for their versatility and robustness, on satellite-and aircraft-based imagery. We chose ASTER data for the first level of resolution and performed lithological mapping and area identification (top section of Figure 4). We mapped the main lithologies of the region by performing selected band ratios. The next level of resolution is provided by plane-based imaging. We performed band ratios and minimum wavelength mapping (MWM) on HyMap imagery and used the results to refine an area of interest (Top section of Figure 4). Once we chose and refined the area of interest, we moved to the next resolution level using RPAS-based sensors. RPAS-based sensors can capture the spectral and topographical data needed to study small scale features (middle and bottom sections of Figure 4).
Field scouting backed by literature show that the REE occurrences are hosted in carbonatite dykes in the study area [33]. Field observation showed that these structures are topographically elevated in a relatively flat surrounding terrain. Therefore, the best field approach was to capture RPAS-based data over the area and to use both spectral and topographic information in parallel. Field scouting backed by literature show that the REE occurrences are hosted in carbonatite dykes in the study area [33]. Field observation showed that these structures are topographically elevated in a relatively flat surrounding terrain. Therefore, the best field approach was to capture RPAS-based data over the area and to use both spectral and topographic information in parallel. structure-from-motion multi-vision-stereo (SfM-MVS) photogrammetry was performed on the orthophotos obtained by a fixed-winged RPAS and a Digital Elevation Model (DEM) was created. The photogrammetry procedure is explained further in Appendix A. Subsequently, the DEM was used to automatically map the dykes (Middle section of Figure 4). The structural and textural analyses were performed using TecGEMS, a Python-based toolbox, which is being developed by Andreani L., Pohl E. and Gloaguen R. [37].
The final level of resolution consists of hyperspectral RPAS-based imagery which was captured with a hexacopter (Bottom section of Figure 4). The RPAS-based hyperspectral data requires specific preprocessing steps and is performed with an in-house toolbox called MEPHySTo (See Appendix A) [28]. After the preprocessing of the RPAS-based hyperspectral imagery, we performed a combination of spectral analyses on the data to map potential REE zones. Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and MWM were used and compared with one another. MWM is usually preferred as it is fitted to a specific spectral signature. On the other hand, due to the low concentration of target minerals and limited spatial resolutions and SNR of existing sensors, absorption features might be difficult to identify. In that case, classification methods, such as SID and SAM, can provide an indirect tool to identify either the target minerals or proxies. Field work was then performed in the refined study area and included traditional field exploration techniques, such as field mapping and representative sampling.
The workflow developed here is versatile as it can easily be adapted to a wide range of commodities and makes use of commonly available datasets and procedures. The use of airborne data is not required but provides valuable information if the data are available. The parameters used in the workflow can easily be tuned by a geologist with some knowledge of remote sensing techniques to adapt the procedure to the geological conditions and the type of ore deposit targeted.
The photogrammetry procedure is explained further in Appendix A. Subsequently, the DEM was used to automatically map the dykes (Middle section of Figure 4). The structural and textural analyses were performed using TecGEMS, a Python-based toolbox, which is being developed by Andreani L., Pohl E. and Gloaguen R. [37].
The final level of resolution consists of hyperspectral RPAS-based imagery which was captured with a hexacopter (Bottom section of Figure 4). The RPAS-based hyperspectral data requires specific preprocessing steps and is performed with an in-house toolbox called MEPHySTo (See Appendix A) [28]. After the preprocessing of the RPAS-based hyperspectral imagery, we performed a combination of spectral analyses on the data to map potential REE zones. Spectral Angle Mapper (SAM), Spectral Information Divergence (SID), and MWM were used and compared with one another. MWM is usually preferred as it is fitted to a specific spectral signature. On the other hand, due to the low concentration of target minerals and limited spatial resolutions and SNR of existing sensors, absorption features might be difficult to identify. In that case, classification methods, such as SID and SAM, can provide an indirect tool to identify either the target minerals or proxies. Field work was then performed in the refined study area and included traditional field exploration techniques, such as field mapping and representative sampling.
The workflow developed here is versatile as it can easily be adapted to a wide range of commodities and makes use of commonly available datasets and procedures. The use of airborne data is not required but provides valuable information if the data are available. The parameters used in the workflow can easily be tuned by a geologist with some knowledge of remote sensing techniques to adapt the procedure to the geological conditions and the type of ore deposit targeted.

Data Acquisition and Preprocessing
Preprocessing of the data, which entails image corrections of multi-and hyperspectral data sets, is a paramount step for the accurate analyses of spectral information, and is performed before further analysis such as spectral classifications. Although they are based on the same principles, the preprocessing steps will differ for each platform according to specific requirements; satellite sensors Remote Sens. 2019, 11, 2500 7 of 28 will require atmospheric corrections unlike RPAS-borne sensors which will need corrections such as lens distortions and band co-registration. The satellite, plane-, and RPAS-based data sets we used have increasing spectral and spatial resolutions and their spectral ranges are visualized in Figure 5. A list of all the platforms and sensors used in the workflow is summarized in Table 1 and the spatial footprints of each data set can be found in Figure 6. Three areas (known as Area 1, Area 4, and Area 5) were surveyed with RPAS, and Area 1 was chosen as demonstration in this article due to its representative geology in the complex and its potential of high HREEs concentration.
performed on the RPAS-based data.

Data Acquisition and Preprocessing
Preprocessing of the data, which entails image corrections of multi-and hyperspectral data sets, is a paramount step for the accurate analyses of spectral information, and is performed before further analysis such as spectral classifications. Although they are based on the same principles, the preprocessing steps will differ for each platform according to specific requirements; satellite sensors will require atmospheric corrections unlike RPAS-borne sensors which will need corrections such as lens distortions and band co-registration. The satellite, plane-, and RPAS-based data sets we used have increasing spectral and spatial resolutions and their spectral ranges are visualized in Figure 5. A list of all the platforms and sensors used in the workflow is summarized in Table 1 and the spatial footprints of each data set can be found in Figure 6. Three areas (known as Area 1, Area 4, and Area 5) were surveyed with RPAS, and Area 1 was chosen as demonstration in this article due to its representative geology in the complex and its potential of high HREEs concentration.

Satellite and Plane-based Sensors
We favor open source satellite data to ensure availability and decrease costs. ASTER data was specifically chosen for this project because of its higher amount of SWIR bands compared to other multispectral satellite sensors such as Landsat. However, Sentinels, SPOT and Landsat data could also be considered. Geological spectral features are mostly situated in the SWIR region of the electromagnetic spectrum and thus other studies have proven the effectiveness of lithological mapping using ASTER imagery [39,40]. The ASTER scene over the study area in Namibia was acquired on 21/09/2006. The sensor parameters are listed in Table 1. The ASTER data was atmospherically corrected using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) tool in ENVI. The FLAASH atmospheric correction tool can be applied to VNIR and SWIR data and is based on the MODTRAN model [41]. Due to the varying resolutions of the spectral channels of the ASTER data, the VNIR bands were resampled from 15 m resolution to 30 m to match the SWIR band resolution before further analyses.
HyMap is an aircraft-mounted hyperspectral sensor operated by the HyVista corporation and developed by Integrated Spectronics [42]. The sensors parameters are summarized in table 1. Available aerial hyperspectral data over Lofdal were processed in order to map relatively small scale geological features. The HyMap data provided by HyVista were already fully corrected and readyto-use with a spatial resolution of 4.5 m. The preprocessing done by the HyVista team consists of radiance and atmospheric calibration as well as geometric corrections [42].

RPAS-based Sensors
Surveys were performed with two types of RPAS, a fixed-wing system and a multicopter (cf. Table 2). Selected areas are mapped using hyperspectral imaging (HSI) and SfM-MVS photogrammetry. The two sensors require specific flight characteristics and acquisition procedures [28]. The frame-based hyperspectral camera needs a stable platform at the time of image acquisition to allow a sufficient integration time and low spatial shift between single bands. The fixed-wing RPAS (Sensefly eBee) has a longer flight time than the multicopter, and can thus cover a larger area size and acquires nadir stereo-photos as input for SfM-MVS photogrammetry. The eBee has a maximum flight

Satellite and Plane-based Sensors
We favor open source satellite data to ensure availability and decrease costs. ASTER data was specifically chosen for this project because of its higher amount of SWIR bands compared to other multispectral satellite sensors such as Landsat. However, Sentinels, SPOT and Landsat data could also be considered. Geological spectral features are mostly situated in the SWIR region of the electromagnetic spectrum and thus other studies have proven the effectiveness of lithological mapping using ASTER imagery [39,40]. The ASTER scene over the study area in Namibia was acquired on 21/09/2006. The sensor parameters are listed in Table 1. The ASTER data was atmospherically corrected using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) tool in ENVI. The FLAASH atmospheric correction tool can be applied to VNIR and SWIR data and is based on the MODTRAN model [41]. Due to the varying resolutions of the spectral channels of the ASTER data, the VNIR bands were resampled from 15 m resolution to 30 m to match the SWIR band resolution before further analyses.
HyMap is an aircraft-mounted hyperspectral sensor operated by the HyVista corporation and developed by Integrated Spectronics [42]. The sensors parameters are summarized in Table 1. Available aerial hyperspectral data over Lofdal were processed in order to map relatively small scale geological features. The HyMap data provided by HyVista were already fully corrected and ready-to-use with a spatial resolution of 4.5 m. The preprocessing done by the HyVista team consists of radiance and atmospheric calibration as well as geometric corrections [42].

RPAS-based Sensors
Surveys were performed with two types of RPAS, a fixed-wing system and a multicopter (cf. Table 2). Selected areas are mapped using hyperspectral imaging (HSI) and SfM-MVS photogrammetry. The two sensors require specific flight characteristics and acquisition procedures [28]. The frame-based hyperspectral camera needs a stable platform at the time of image acquisition to allow a sufficient integration time and low spatial shift between single bands. The fixed-wing RPAS (Sensefly eBee) has a longer flight time than the multicopter, and can thus cover a larger area size and acquires nadir stereo-photos as input for SfM-MVS photogrammetry. The eBee has a maximum flight time of 50 min and is equipped with a Canon Powershot S110 digital camera with a resolution of 16MP (Table 2). A predefined flight-route with GPS waypoint is created and uploaded onto the RPAS before take-off. The flight line spacing is determined in such a way that there is a high overlap of the aerial photographs (85% forward lap and 70% side lap).

Satellite and Airborne Approach
Even though it is not possible to directly detect small carbonatite bodies with ASTER imagery due to its low spatial resolution, we used indicators in the regional geology to locate ideal areas where they might occur. Normally, carbonatites occur as small bodies within alkalic intrusive complexes, and the emplacement of carbonatite intrusions are commonly accompanied by hydrothermal alteration of the country rock [43,44]. The hydrothermal fluids can introduce REEs into the system. Thus, we have summarized two main indicators to locate possible occurrences of REE-bearing carbonatites. (1) We identify alkali rocks which can be associated with carbonatites. (2) We locate regions of deformation which could have facilitated hydrothermal activity ( Figure 7A).
With these two indicators in mind, a smaller area of interest was chosen and analyzed using HyMap data ( Figure 7B). However, we were not able to detect REE-rich regions with the HyMap data. To map REEs, their concentrations within a pixel must be such that absorption features will occur in the spectra. This is only possible if the pixel size matches the areas of high concentration. REEs are usually found in small concentrations zones within larger rock formations. A higher ground resolution than what HyMap (resolution of 4.5 m) can offer is thus required. Thus, we decided on a different indicator for possible REE-rich areas. A link between the presence of iron and REEs has been suggested by Salminen et al., (2005). With the HyMap data, we were able to select promising regions containing both these carbonatite and iron features. We selected three promising regions to demonstrate our field work approach involving RPAS ( Figure 7C,D). One area was specifically chosen because the exploration company who owns the exploration rights to Lofdal (Namibian Rare Earths) gave us access to the area (defined by Namibian Rare Earths as Area 1) and provided geochemical analyses of rock samples to validate our procedure.
RPAS-based data was then used to map the numerous dykes over a given area ( Figure 7C) and characterized with RPAS-based hyperspectral data ( Figure 7D). Field measurements with the handheld XRF and spectrometer indicated that there were primarily two sets of dykes, one set enriched with iron, and the other set contained very little iron. This allowed us to spectrally differentiate between the dykes based on chemical composition. Harmer and Nex [45] have noted that the highest REE concentrations are usually found in iron-rich carbonatites. Therefore, we decided to focus on the iron-rich dyke to test whether we can map REE-rich zones using RPAS-based sensors. Remote Sens. 2019, 11, x FOR PEER REVIEW 11 of 30 RPAS-based data was then used to map the numerous dykes over a given area ( Figure 7C) and characterized with RPAS-based hyperspectral data ( Figure 7D). Field measurements with the handheld XRF and spectrometer indicated that there were primarily two sets of dykes, one set enriched with iron, and the other set contained very little iron. This allowed us to spectrally differentiate between the dykes based on chemical composition. Harmer and Nex [46] have noted that the highest REE concentrations are usually found in iron-rich carbonatites. Therefore, we decided to focus on the iron-rich dyke to test whether we can map REE-rich zones using RPAS-based sensors.

Spectral Analyses
Band ratios were specifically chosen for their proven robustness and applied to the ASTER imagery [47]. A combination of SWIR and VNIR bands enhances geological units, making lithological discrimination possible [45]. The band ratios chosen for this purpose were taken from Abrahams and Hook [45] and are as follows; B4/B7, B4/B3, and B2/B1 in the R-G-B channels, respectively. Band 7 is used to distinguish magnesium hydroxide and amphiboles while iron can be detected with the VNIR bands [48].
We performed band ratios and minimum wavelength mapping (MWM) in parallel on the HyMap data of a given area to investigate the geological features. Numerous studies have established the potential of band ratios to map iron occurrences, a general feature to distinguish lithologies [49,50]. Thus, an iron band ratio (769 nm/458 nm) was performed on the HyMap data. MWM is a procedure whereby the wavelength position of the deepest absorption features is mapped, and has been successfully used to distinguish minerals with strong spectral features [51]. This works best with hyperspectral data because of the high spectral resolution and contiguity of the bands allowing it to

Spectral Analyses
Band ratios were specifically chosen for their proven robustness and applied to the ASTER imagery [46]. A combination of SWIR and VNIR bands enhances geological units, making lithological discrimination possible [47]. The band ratios chosen for this purpose were taken from Abrahams and Hook [47] and are as follows; B4/B7, B4/B3, and B2/B1 in the R-G-B channels, respectively. Band 7 is used to distinguish magnesium hydroxide and amphiboles while iron can be detected with the VNIR bands [48].
We performed band ratios and minimum wavelength mapping (MWM) in parallel on the HyMap data of a given area to investigate the geological features. Numerous studies have established the potential of band ratios to map iron occurrences, a general feature to distinguish lithologies [49,50]. Thus, an iron band ratio (769 nm/458 nm) was performed on the HyMap data. MWM is a procedure whereby the wavelength position of the deepest absorption features is mapped, and has been successfully used to distinguish minerals with strong spectral features [51]. This works best with hyperspectral data because of the high spectral resolution and contiguity of the bands allowing it to display narrow absorption features. This powerful method was used to map the distinct carbonate absorption feature at 2340 nm [52].
RPAS-based hyperspectral data was used to characterize dykes in a chosen area. We again used an iron band ratio to map and distinguish dykes in the image; the 760 nm wavelength was divided by the 880 nm wavelength to enhance the iron feature. We then mapped zones of REE occurrences within individual dykes using the same hyperspectral data. Pixels outside the analyzed dykes were masked. Both (1) direct and (2) indirect methods were used to identify zones possibly containing REEs within a dyke.
(1) Based on the results of laboratory tests, MWM was used to map the 800 nm absorption feature of Nd. Spectra measured by the Rikola imager in the laboratory has an approximate 7 nm shift to higher wavelengths (See Appendix B). Considering this spectral shift, a range was set between wavelengths 795 nm and 820 nm to map the deepest absorption feature within that region. Although it is the most direct way to map REE absorption features, noise in the spectra made it challenging to confidently detect minute absorption features.
(2) The spectral matching methods SAM and SID are supervised classification methods. SAM is a widely deterministic classification method in hyperspectral remote sensing and many studies have shown its effectiveness in mineral mapping [53,54]. Although SID is not as widely used, it also determines the similarity between a pair of spectra. SID, however, is a stochastic approach that can characterize spectral similarities more effectively than SAM [55]. We used the spectrometer to measure spectra along the dykes and created a library from spectra containing REE absorption features at 580 nm, 740 nm, and 800 nm. Of the in situ spectra collected, two measurement points showed positive REE absorption features. These spectra were averaged and used as the reference spectra for both SAM and SID methods. The idea is to find sections which are spectrally similar to host rocks containing REEs.

Extraction of structural features
The high-resolution DEM ( Figure 8) obtained from SfM-MVS photogrammetry allows us to map linear features such as carbonatite dykes, which form narrow and shallow topographic ridges as a result of differential erosion. For this purpose, we propose to use a ridge detection algorithm. The algorithm locates the points of maximum curvature in the DEM which correspond to the surface trace of dykes. One common approach to analyze the curvature of elevation data is to use an eigenvalue analysis of the Hessian matrix [56,57]. The Hessian matrix is a 2 × 2 matrix composed of second-order partial derivatives of the input image, whereas the second-order partial derivatives are defined as a convolution with derivatives of Gaussian filter at scale σ. The idea behind eigenvalue analysis of the Hessian matrix is to extract the principal directions and magnitude in which the local second order structure of the image can be decomposed. Each pixel is associated to a set of eigenvectors |λ1| or |λ2|. Linear features are characterized by a very small magnitude of λ1 (ideally close to zero) and a large magnitude of λ2, whereas point features will be characterized by similar magnitudes of λ1 and λ2 and features without preferential directions will have low magnitudes for both λ1 and λ2.
To apply this technique, we applied a low pass-filter to remove small artifacts such as boulders as well as accelerate the processing. The eigenvalue analysis was then performed on the resampled DEM using an initial Gaussian filter. Once the absolute magnitude of eigenvectors is extracted, a threshold is used in order to remove weak ridges (i.e., associated to a low curvature). This threshold was set to 75% of the maximum magnitude based on previous studies and experimentation. Ridges were then identified using a local maxima function (i.e., pixels higher than their neighbors along the x or y axis were tagged as "ridges"). Finally, identified ridges were vectorized by connecting neighboring tagged pixels. The final step is a simplification of extracted linear features. Segments smaller than 20 pixels (ca. 4 m) were discarded and the trace of remaining segments is simplified using the approach of Visvalingam and Whyatt [58]. One drawback of this approach is that other curvilinear features such as the edges of gullies may be falsely detected as well. For this reason, a manual cleaning is crucial and was performed in order to discard irrelevant features in a last step. Remote Sens. 2019, 11, x FOR PEER REVIEW 13 of 30

Validation and field procedure
Validation points were measured with a handheld XRF and a portable spectroradiometer, while their positions were recorded by a handheld GPS. Strike and dip of the outcropping carbonatite dykes in the study area were also measured. Both methods rapidly provide information about rock chemistry during field work and allow a better field planning.
A Spectral Evolution PSR-3500 portable spectroradiometer was used to measure the spectra of rock surfaces and hand samples. Spectra are recorded in the VNIR/SWIR region of the electromagnetic spectrum (400-2500 nm) using a contact probe (8 mm spot size) with artificial illumination. The spectral resolution in the VNIR range is 3.5 nm while the spectral resolution is 7 nm in the SWIR range. A precalibrated PTFE panel (Zenith polymer; >99% reflectance in the VNIR range and >95% in the SWIR) was used to convert the radiance values to reflectance. Each spectral record consists of 10 individual measurements taken consecutively and averaged.
A Bruker S1 Titan 800 handheld X-ray fluorescence spectrometer was used both in the field and laboratory, operated in Brukers GeoChem calibration mode. The Bruker features a Rh target X-ray tube with a maximum acceleration voltage of 50kV at a maximum current of 0.2 mA and a Silicon Drift Detector (SDD). The X-ray emission in GeoChem mode is recorded between 0 to 45 kV in two separate phases: the 1st phase includes heavy elements (≥Fe) and is recorded using a TiAl filter and 45 kV acceleration voltage, and the 2nd phase includes the light elements (≤Fe) without a filter and 15 kV. The handheld XRF measured in situ geochemical information of the rock outcrops, giving us an estimate of iron and, via indicators (i.e., Y or Zr), the REE contents. The rapidly captured geochemical data roughly indicate REE mineralized areas, which aids us in determining focus areas. Consecutive XRF and XRD analysis of rock samples were measured in the laboratory to verify the

Validation and field procedure
Validation points were measured with a handheld XRF and a portable spectroradiometer, while their positions were recorded by a handheld GPS. Strike and dip of the outcropping carbonatite dykes in the study area were also measured. Both methods rapidly provide information about rock chemistry during field work and allow a better field planning.
A Spectral Evolution PSR-3500 portable spectroradiometer was used to measure the spectra of rock surfaces and hand samples. Spectra are recorded in the VNIR/SWIR region of the electromagnetic spectrum (400-2500 nm) using a contact probe (8 mm spot size) with artificial illumination. The spectral resolution in the VNIR range is 3.5 nm while the spectral resolution is 7 nm in the SWIR range. A precalibrated PTFE panel (Zenith polymer; >99% reflectance in the VNIR range and >95% in the SWIR) was used to convert the radiance values to reflectance. Each spectral record consists of 10 individual measurements taken consecutively and averaged.
A Bruker S1 Titan 800 handheld X-ray fluorescence spectrometer was used both in the field and laboratory, operated in Brukers GeoChem calibration mode. The Bruker features a Rh target X-ray tube with a maximum acceleration voltage of 50kV at a maximum current of 0.2 mA and a Silicon Drift Detector (SDD). The X-ray emission in GeoChem mode is recorded between 0 to 45 kV in two separate phases: the 1st phase includes heavy elements (≥Fe) and is recorded using a TiAl filter and 45 kV acceleration voltage, and the 2nd phase includes the light elements (≤Fe) without a filter and 15 kV. The handheld XRF measured in situ geochemical information of the rock outcrops, giving us an estimate of iron and, via indicators (i.e., Y or Zr), the REE contents. The rapidly captured geochemical data roughly indicate REE mineralized areas, which aids us in determining focus areas. Consecutive XRF and XRD analysis of rock samples were measured in the laboratory to verify the field measurements as well as to validate spectral field information. Selected samples were sent for whole rock geochemical assay using ICP-MS fusion for major and selected trace elements.

Results from Lofdal
We performed dedicated band ratios on the ASTER data as mentioned in the section "Spectral analyses", to identify areas which may contain carbonatite-hosted REEs and define a region of interest. The result enabled us to discriminate between the main lithologies of amphiboles, syenites, and gneisses, and possible alteration zones were highlighted by identifying iron ( Figure 9A). The structure of the region was also discerned from the band ratio result, which were not entirely evident in the plain RGB image. We were able to map major faults by locating areas with offsets in the lithological layering ( Figure 9A.1). We also detected highly curved structures indicating intense folding in the region ( Figure 9A.2). The faulting and folding show that the area has undergone regional ductile deformation. field measurements as well as to validate spectral field information. Selected samples were sent for whole rock geochemical assay using ICP-MS fusion for major and selected trace elements.

Results from Lofdal
We performed dedicated band ratios on the ASTER data as mentioned in the section "Spectral analyses", to identify areas which may contain carbonatite-hosted REEs and define a region of interest. The result enabled us to discriminate between the main lithologies of amphiboles, syenites, and gneisses, and possible alteration zones were highlighted by identifying iron ( Figure 9A). The structure of the region was also discerned from the band ratio result, which were not entirely evident in the plain RGB image. We were able to map major faults by locating areas with offsets in the lithological layering ( Figure 9A.1). We also detected highly curved structures indicating intense folding in the region ( Figure 9A.2). The faulting and folding show that the area has undergone regional ductile deformation. We selected an area of interest close to alkali rock (the syenite intrusion) showing regional deformation, as a promising site for the occurrence of carbonatites. This area was investigated in more detail with HyMap data. The results of the iron band ratio and the regions rich in carbonates identified using MWM were overlain and the product can be seen in figure 10. One of the selected areas containing both iron and carbonatites was used to demonstrate our RPAS-based field surveying. We selected an area of interest close to alkali rock (the syenite intrusion) showing regional deformation, as a promising site for the occurrence of carbonatites. This area was investigated in more detail with HyMap data. The results of the iron band ratio and the regions rich in carbonates identified using MWM were overlain and the product can be seen in Figure 10. One of the selected areas containing both iron and carbonatites was used to demonstrate our RPAS-based field surveying.
The selected area contains numerous thin dykes. To map the dykes in this area, a better resolution than what satellite or plane-based data can offer is needed. The photogrammetric DEM derived from the eBee over a specific area of interest has a ground resolution of 4.9 cm and covers of approximately 1 km 2 . Field validation supported by the DEMs to locate structures of interest confirmed that the dykes are mostly composed of carbonatites. Geological maps and available information provided by local geologists indicated that the dykes have not been thoroughly mapped in this area.
The dykes are not easily discernible from the DEM alone, and performing solely a manual extraction would be arduous and subjective. We, therefore, automatically mapped the dykes using morphological feature extraction ( Figure 11A-D). The specific techniques and parameters are outlined in Table 3. A ridge detection algorithm was performed on elevation data ( Figure 11A,B) and lineaments were automatically extracted ( Figure 11C). Automatically extracted features were refined manually to map dykes in Area 1. We were able to map the dykes rapidly and remotely using this procedure. Approximately 45 dykes were mapped in the area. Overall, the dykes show a North-easterly trend.
A small area containing a complex pattern of dykes was selected for precise hyperspectral imaging with the multicopter mounted Rikola ( Figure 11D). The selected area contains numerous thin dykes. To map the dykes in this area, a better resolution than what satellite or plane-based data can offer is needed. The photogrammetric DEM derived from the eBee over a specific area of interest has a ground resolution of 4.9 cm and covers of approximately 1 km 2 . Field validation supported by the DEMs to locate structures of interest confirmed that the dykes are mostly composed of carbonatites. Geological maps and available information provided by local geologists indicated that the dykes have not been thoroughly mapped in this area.
The dykes are not easily discernible from the DEM alone, and performing solely a manual extraction would be arduous and subjective. We, therefore, automatically mapped the dykes using morphological feature extraction ( Figure 11A-D). The specific techniques and parameters are outlined in table 3. A ridge detection algorithm was performed on elevation data ( Figure 11A,B) and lineaments were automatically extracted ( Figure 11C). Automatically extracted features were refined manually to map dykes in Area 1. We were able to map the dykes rapidly and remotely using this procedure. Approximately 45 dykes were mapped in the area. Overall, the dykes show a Northeasterly trend. A small area containing a complex pattern of dykes was selected for precise hyperspectral imaging with the multicopter mounted Rikola ( Figure 11D).    We identified two types of dykes (D1 and D2) in the Rikola scene ( Figure 12A), based on geometrical and compositional characteristics. The high spatial resolution data indicates that the dykes have two main orientations. One set of dykes strikes at~35 • NE (D1), whereas the other trends around 60 • NE (D2). D2 cross-cut D1 creating a sinistral offset in D1 by~2 m. This allows us to establish the age relationship between the two dykes (i.e., D1 intruded before D2) and thus infer that the area has at least two generations of dyke intrusions. A band ratio to enhance the iron feature on the hyperspectral data confirmed the field observation that D1 is relatively enriched in iron compared to D2 ( Figure 12B). An accuracy assessment was performed to quantitatively assess the classification of the iron-rich dyke on the hyperspectral RPAS-based imagery. The image was grouped into three classes; iron-rich zones, iron-poor zones, and soil. A confusion matrix shows that the overall accuracy of the dyke classification with the iron band ratio was 84%. Moreover, the band ratio classified 81.78% of the pixels correctly and only 3.52% pixels were incorrectly classified as the iron-poor dyke and 11.54% of the pixels were mistakenly classified as soil. Most of the pixels incorrectly classified using the iron band ratio can be attributed to the numerous rock pieces and weathered surfaces surrounding the dikes. Additionally, extended limestone dissolutions and downslope precipitations were identified during our field observations. However, with these accuracies we are confident in using an iron band ratio to differentiate the dykes. Moreover, we can use their specific geometry in addition to the iron content and extend the mapping of these carbonatite dykes over the entire region. using an iron band ratio to differentiate the dykes. Moreover, we can use their specific geometry in addition to the iron content and extend the mapping of these carbonatite dykes over the entire region. The band ratio result was confirmed in situ both with field spectroscopy and a portable XRF. Determining the precise geometry and relationships of the two sets of dykes is only made possible with the high spatial resolution hyperspectral RPAS-based images. XRF analyses were taken of rock samples from both types of dykes to validate the classification. The results of samples taken at the same point locations where spectral measurements were taken indicate that a representative dyke from D1 contained 21% iron (Fe) and 0.9% yttrium (Y), whereas a dyke belonging to D2 contained 2.5% Fe and 0.5% Y. From the in situ portable XRF analyses, we determined that the D1 dyke not only have higher Fe contents than D2, but also higher amounts of yttrium, a proxy for HREE in rocks [59]. We thus decided that the D1 dykes were promising targets and tested the in situ measurement of REE from RPAS in operational mode for the first time. A selected area was surveyed and the resulting data was adequately processed. In order to minimize false positives, the REE mapping was limited to the dykes and surrounding areas were masked out. The mask was based on the iron band ratio, in which the "non-iron" pixels are discarded. In Figure 13, a segment of the iron-rich D1 dyke was chosen to demonstrate the methods we use to map REEs with hyperspectral RPAS-based imagery ( Figure 13A). Overall, the distribution of pixels positively identified as REEs with MWM are disseminated throughout the dyke segment. MWM indicated a higher concentration of REEs in the middle (at point A) and upper parts of the dyke segment ( Figure 13A). SAM and SID show pixels in red which are spectrally similar to REE hosting rocks. The reference spectra for both these methods were taken at points A and B on Dyke 1 and is shown in Figure 13A. The SID result shows positive pixels in the middle of the dyke section ( Figure  13B), which are in the same region as the positive pixels of the MWM result. On the other hand, the SAM result showed visibly less positive pixels, but still showed a very little amount in the middle and lower parts of the dyke section ( Figure 13C). The combination of all three methods provides us with possible indications of REE-rich zones within the dyke. The band ratio result was confirmed in situ both with field spectroscopy and a portable XRF. Determining the precise geometry and relationships of the two sets of dykes is only made possible with the high spatial resolution hyperspectral RPAS-based images. XRF analyses were taken of rock samples from both types of dykes to validate the classification. The results of samples taken at the same point locations where spectral measurements were taken indicate that a representative dyke from D1 contained 21% iron (Fe) and 0.9% yttrium (Y), whereas a dyke belonging to D2 contained 2.5% Fe and 0.5% Y. From the in situ portable XRF analyses, we determined that the D1 dyke not only have higher Fe contents than D2, but also higher amounts of yttrium, a proxy for HREE in rocks [59]. We thus decided that the D1 dykes were promising targets and tested the in situ measurement of REE from RPAS in operational mode for the first time. A selected area was surveyed and the resulting data was adequately processed. In order to minimize false positives, the REE mapping was limited to the dykes and surrounding areas were masked out. The mask was based on the iron band ratio, in which the "non-iron" pixels are discarded.
In Figure 13, a segment of the iron-rich D1 dyke was chosen to demonstrate the methods we use to map REEs with hyperspectral RPAS-based imagery ( Figure 13A). Overall, the distribution of pixels positively identified as REEs with MWM are disseminated throughout the dyke segment. MWM indicated a higher concentration of REEs in the middle (at point A) and upper parts of the dyke segment ( Figure 13A). SAM and SID show pixels in red which are spectrally similar to REE hosting rocks. The reference spectra for both these methods were taken at points A and B on Dyke 1 and is shown in Figure 13A. The SID result shows positive pixels in the middle of the dyke section ( Figure 13B), which are in the same region as the positive pixels of the MWM result. On the other hand, the SAM result showed visibly less positive pixels, but still showed a very little amount in the middle and lower parts of the dyke section ( Figure 13C). The combination of all three methods provides us with possible indications of REE-rich zones within the dyke. To validate our results, spectra from the hyperspectral images were compared with spectra taken by the handheld spectroradiometer of each dyke (Figure 14). Both handheld spectra show characteristic carbonate absorption features at ~2335 nm, whereas only Dyke 1 shows a broad iron To validate our results, spectra from the hyperspectral images were compared with spectra taken by the handheld spectroradiometer of each dyke (Figure 14). Both handheld spectra show characteristic carbonate absorption features at~2335 nm, whereas only Dyke 1 shows a broad iron charge-transfer feature in the VNIR (Figure 14A1,B1). This feature characteristically peaks at 780 nm and steeply dips from 900 nm [60]. The spectra from the Rikola range between 450 nm and 900 nm also show the iron feature in Dyke 1 (Figure 14A2,B2). XRD revealed that iron is mainly present in the form of goethite (FeOOH) and to lower contents as hematite (Fe2O3-below 3 wt-% in total). In contrast, no typical Fe-mineral was detected in the sample of Dyke 2, with the exception of a low content of chlorite. The XRD results also indicated that the samples from Dyke 1 and Dyke 2 contained 64% and 85% calcite, respectively (Tables 4 and 5), confirming that both dykes are indeed carbonatites. Both dykes contained REE-bearing minerals such as apatite and xenotime, whereas the sample from Dyke 1 contained more xenotime, one of the major REE-bearing minerals. The XRD results also confirms the in situ measurements from the portable XRF. Table 4. XRD results of field sample taken from D1 at validation point 1 shown in Figure 12A. XRD revealed that iron is mainly present in the form of goethite (FeOOH) and to lower contents as hematite (Fe 2 O 3 -below 3 wt-% in total). In contrast, no typical Fe-mineral was detected in the sample of Dyke 2, with the exception of a low content of chlorite. The XRD results also indicated that the samples from Dyke 1 and Dyke 2 contained 64% and 85% calcite, respectively (Tables 4 and 5), confirming that both dykes are indeed carbonatites. Both dykes contained REE-bearing minerals such as apatite and xenotime, whereas the sample from Dyke 1 contained more xenotime, one of the major REE-bearing minerals. The XRD results also confirms the in situ measurements from the portable XRF. Table 4. XRD results of field sample taken from D1 at validation point 1 shown in Figure 12A.  Table 5. XRD results of field sample taken from D2 at validation point 2 shown in Figure 12A. The whole rock geochemical analysis of the two samples shows similar results: Dyke 1 is more enriched in iron and REEs, especially HREEs, than Dyke 2 (Figures 15 and 16). This verifies the XRD results (with higher contents of xenotime in Dyke 1) as well as the in situ handheld XRF measurements, which also indicated that Dyke 1 has a higher amount of Y. The data from geochemical and mineralogical (XRD) investigation are consistent and support the field work and remote sensing findings.  The whole rock geochemical analysis of the two samples shows similar results: Dyke 1 is more enriched in iron and REEs, especially HREEs, than Dyke 2 (Figures 15 and 16). This verifies the XRD results (with higher contents of xenotime in Dyke 1) as well as the in situ handheld XRF measurements, which also indicated that Dyke 1 has a higher amount of Y. The data from geochemical and mineralogical (XRD) investigation are consistent and support the field work and remote sensing findings.    The whole rock geochemical analysis of the two samples shows similar results: Dyke 1 is more enriched in iron and REEs, especially HREEs, than Dyke 2 (Figures 15 and 16). This verifies the XRD results (with higher contents of xenotime in Dyke 1) as well as the in situ handheld XRF measurements, which also indicated that Dyke 1 has a higher amount of Y. The data from geochemical and mineralogical (XRD) investigation are consistent and support the field work and remote sensing findings.

Discussion
The sequence of exploration phases typically progresses from a large area to a small focused target. The integration of remote sensing data sets from different scales mirrors this fundamental process. REE deposits associated with alkaline carbonatite complexes are typically associated with a particular magmatic or magmatic-hydrothermal phase, which affects a significantly smaller area than the whole of the complex. The first step is to identify the carbonatite-alkaline complex, where the final result should reveal a focused target for drilling and detailed sampling prior to the evaluation of the resource. The aim of an exploration geologist is to have as much information as possible about the geology of the area before doing the field work, usually in the form of a detailed geological map. Satellite and aircraft-based remote sensing is usually the first step in any field exploration endeavor, as it provides basemaps that can be used to improve or generate geological maps. Traditional geological maps in former years are usually produced through field mapping (e.g., mapping on foot) and incorporates assumptions to fill in the blanks that were based on the accepted paradigms at the times of their production. Extensive mapping is usually required in mineral exploration, especially in remote locations, where existing geological maps are imprecise and in need of upgrading.
The discovery of mineral resources in remote regions and/or without conspicuous surface indicators presents a challenge for exploration. Rare earth deposits are often associated with small-scale, lithologic (e.g., carbonatites), and tectonic structures (e.g., shear zones), and are not easily detected by traditional remote sensing. While remote sensing has been increasingly used in recent years in the field of geology thanks to the development of better sensors (multi-and hyperspectral, and airborne and satellite-based) and increased computing power [61], it is not commonly and operatively used in mineral exploration. In this manuscript we argue that multiscale remote sensing approaches provide a cost effective way of exploring critical metals hosted in poorly accessible structures. We also demonstrate that field mapping and sampling with prior remote sensing is often efficient and more lucrative.
Previously, the use of remote sensing in the field of earth observation has focused more on the agricultural and environmental sectors. Topics such as precision farming [62,63] and environmental monitoring [64,65] have long dominated the remote sensing field. Lesser focus has been placed on the geological application of remote sensing. That being said, a few studies have successfully mapped large geological features via multispectral and/or hyperspectral satellite sensors, providing large scale estimates of the area in question [47,61,66]. The challenging aspect for researchers at the moment is the resolution gap between satellite-or aircraft-based imaging, and ground surveying. To address this challenge, an innovative approach is taken whereby RPAS are used to bridge the gap in the scale of data collection. This study not only uses one type of sensor and data scale, but introduces a work scheme involving various data sources, from satellite-and aircraft-based, to RPAS-based sensors and ground-based surveys. The work chain is then completed with field validation through geochemical and spectral analyses.
We used ASTER and HyMap data in combination with existing geological maps to improve field exploration and rock sampling and identify areas of interest. We were able to map the regional geology and major structures with band ratios using satellite data. Additionally, ASTER-TIR data can be used to map felsic rocks, such as the syenites we investigated in this study. Ratios such as a quartz index have been shown to be valuable tools in determining the locations of felsic rocks covering large area extents [67]. The major structures we mapped shows that large scale deformation took place and indicate that the area might have undergone hydrothermal activity. With this information we were able to locate a promising region of interest for the exploration of REE-hosted carbonatites. The spectrally and spatially higher resolution HyMap data allowed us to further analyze the area by identifying regions of overlapping carbonate and iron features and, thus, select a particular region of interest. Selecting the relevant samples was made possible with an improved geological map and by identifying the relevant areas anteriorly, decreased the amount of required samples. Ultimately, field work was improved, accelerated and simplified.
Currently, RPAS are mainly used for SfM-MVS photogrammetry and aerial photography. Studies such as mine and tailings monitoring [68,69], glaciology, and land slide monitoring [70][71][72], as well as topographic mapping [73], have largely used RPAS for SfM-MVS purposes and aerial photography. Hyperspectral RPAS-based surveying is still a new and developing field, and has almost primarily been used in precision agriculture and forestry [17,21,74]. There are however a few studies which uses hyperspectral RPAS-based surveying for environmental monitoring, such as the impact mining has on the environment [50]. Nonetheless, the use of RPASs for mineral exploration is an underutilized source for data acquisition. For this study we attempt to use the full potential of RPAS-based imaging. We not only use topographic information captured with RPAS, but we integrate it with spectral information from a hyperspectral RPAS-based sensor. Topographic information from RPAS-based data was used to automatically map the dykes in the area. This data was obtained through SfM-MVS photogrammetry using the fixed-winged eBee. As the eBee is able to cover 4 sq.km/hour, the majority of the dykes in Area 1 could be mapped in 20 min, and the processing time for the data takes a few hours. Traditional field mapping of the same area would have taken a geologist days. Altogether, this technique is much faster, safer and the procedure is cheaper than conventional field mapping.
During the field campaign, we selected promising areas and tested the detection of REEs from RPAS directly in the field for the first time. The RPAS-based Rikola data showed that it is possible to distinguish different types of dykes based on their iron content. Additionally, the high spatial resolution of the data made it possible to determine the age relationship between the dykes based on their positions to each other. Identifying REEs within an individual dyke has proven to be challenging. Ideally, MWM would have been the most direct method to map REEs, but noise introduced by RPAS-based imaging created uncertainties. Thus, to tackle this challenge, we used spectral classification methods as well, and we were able to adequately pinpoint possible REE enriched zones within a dyke by comparing all three results. Field validation is essential for the study of mineral exploration with remote sensing data. Each carbonatite body studied by means of satellite-, aircraft-, and RPAS-based surveys required proper geochemical validation. To validate the remote sensing data, rock samples were collected in the field for laboratory geochemical and spectral analyses. By determining the iron content of rock samples from each dyke with whole rock geochemical analyses and XRF, we validated the procedure to spectrally classify dykes based on iron content. Geochemical analyses also indicated that there were in fact more HREE in the iron-rich dykes, which focused our attentions on one set of dykes, increasing the efficiency of our investigations. Being able to successfully validate the remote sensing results demonstrates that our multisource approach to mineral exploration can easily be extended to other structurally bound deposits.
This study has integrated and synthesized information from multiple platforms with diverse sensors to provide improvements to the detection of REE mineralization within structurally bound carbonatite bodies. Although mapping of wide areas can be done with satellite acquisitions, it lacks the spatial resolution required to identify small structures. Thus, aircraft-and RPAS-borne sensors are used additionally to obtain data with an increased spatial resolution over a smaller, more defined area. Although we focus on the remote sensing aspects, it is not intended to replace traditional exploration procedures but rather to improve them. When taking costs into consideration, based on our previous experience, RPAS-based surveys usually become considerably cheaper beyond approximately nine days of field mapping by a qualified geologist. This calculation is based on logistical (e.g., import/export and transport) and operational (e.g., processing) factors and might slightly vary depending on the local conditions (e.g., remoteness and accessibility).
On the other hand, the use of RPASs for mineral exploration purposes does have a few challenging aspects. Environmental factors such as strong winds and bad weather conditions may prohibit the RPAS from flying. All these challenges however can be overcome with diligent flight planning. Hyperspectral RPAS-based surveying is still a developing field. This means that there are no standardized preprocessing methods universally used for the acquired data. It is, therefore, very important to practice verified preprocessing procedures to ensure well-corrected, quality data [75]. Furthermore, our data processing workflow can be adapted to other deposits and is not limited to carbonatite-hosted REEs. It can be adjusted to detect deposits associated with alteration or deposits characterized by specific structures and geometries. This approach can greatly improve exploration targeting for various commodities. By no means does this technique replace geologists in the field. With this approach we intended to improve traditional exploration methods by (1) reducing the exploration footprint, (2) increasing personal safety, as well as (3) accelerating the overall process.

Conclusions and Outlook
Multisource and multiscale field exploration integrates the full range of available remote sensing data sets, from space-borne imaging over airborne data, to high-resolution RPAS-based hyperspectral data acquisition as well as ground validation. These diverse data sets provide indispensable information on the relief and mineral contents of the geological formations in areas of interest at different scales. However, spectral data alone can lack the sufficient discrimination potential either due to spatial and spectral sensor resolutions or noise produced during acquisition. The integration of geomorphic data can improve mapping. It is also important to have an accurate preprocessing scheme for the topographic and spectral data obtained from RPAS platforms. We demonstrated that our correction methods produced geometrically and spectrally accurate results. The use of RPAS-based exploration in the chain of multisource and multiscale data acquisition resulted in data with both high spectral and spatial resolution. Despite the fact that the use of RPAS in the field is not always possible due to environmental factors, RPAS-and/or ground-based surveying lowers the costs and improves the accuracy of field data, which shows its high potential in the use of future mineral exploration.
In the case of carbonatite hosted REEs in southern Africa we could demonstrate that: (1) Multispectral satellite data can be used for regional scale targeting and structural interpretation.
(2) The high spatial resolution from hyperspectral airplane-based data can allow us to locate regions of interest for further high resolution mapping. (3) High spatial resolution imagery captured with RPAS allowed us to study and map relatively small sized structures, and (4) by using high spatial resolution RPAS-based hyperspectral data we could identify promising areas with high REE concentrations. Based on the above findings, we suggest that a multisource and multiscale approach is beneficial for the exploration of structurally bound mineral deposits.
One current limitation is that the spectral range of low-cost, lightweight RPAS-based cameras only cover the visible and NIR part of the spectrum. This makes it impossible to detect important absorption features situated in the SWIR portion of the electromagnetic spectrum. However, with the rapid pace of technological advancements, it is only a matter of time when such a sensor will be developed. Additionally, RPAS-based magnetometers may offer subsurface solutions which can be integrated with surficial mineral mapping from hyperspectral RPAS-based data. These technologies will probably revolutionize the field of remote sensing geology. size additionally creates strong geometric patterns and thus impedes the detection of individual invariants. Therefore, manual georeferencing was applied in QGIS open-source software before further processing. 4.
Topographic correction: The relief of the area, such as slopes in various orientations with respect with sun incidence, influences the illumination within an image. Different sunlight incidence angles on the same material cause the radiance of that material to vary. The MEPHySTo toolbox may be used to perform respective topographic corrections. In this case, topographic correction was performed but did not provide any additional improvement therefore both corrected and uncorrected data can be used further. For this study case, uncorrected data was used in preference as topographic correction can introduce artifacts that affect the spectral content if the noise in the DEM is large compared to the relief in the scene. 5.
Radiometric correction: The hyperspectral radiance images are converted to reflectance using an empirical line correction. This is done by using known spectra from black, gray, and white PVC panels placed in the field while the images are being captured. 6.
Spectral smoothing: A Savitzky-Golay filter was used to smooth the data to decrease spectral noise. A successive series of low degree polynoms is fitted to each hyperspectral spectrum using linear least squares within a defined moving window (Savitzky & Golay 1964). A polynomial order of 3 was chosen with a window size of 5. 7.
Mosaicking: As a last step, the corrected and georeferenced images can be stitched together to create a mosaic of the entire area.

Appendix B. Suitability of the Rikola Imager
The first question that comes to mind concerning the potential of a lightweight hyperspectral imager for operational use is if the spectral resolution and SNR of the Rikola imager is sufficient for the direct detection of REEs. The Rikola imager was thus initially tested in the laboratory to capture hyperspectral images of cut samples to validate the use of the device in operational conditions. Cut samples allow the capture of spectra on a clean, unweathered surface to produce accurate spectral results. We stacked multiple acquisitions and averaged them to eliminate noise without erasing minute spectral features and thus produce a robust and reliable spectrum. A selected spectrum was compared with a spectrum taken by a portable spectrometer on the same point on the sample ( Figure A1). Nd absorption features can be seen in the spectrum from the Rikola and from the spectrometer. The slight shift in the observed Rikola spectrum (~7 nm shift to higher wavelength) can be attributed to the varying spectral resolutions of the two sensors in the device. This sampling effect generates a bias in the discrimination on the local minima that can reach up to half the spectral resolution. It should be noted that the shift might also be attributed to the insufficient spectral calibration of the sensor. It is also worth noting the section of the Rikola spectrum labeled as sensor noise in Figure A1 is caused by the internal setup of the camera. The camera consist of 2 sensors, one captures spectra from 506-636 nm while the other sensor captures the VNIR range from 650 to 900 nm [81]. This causes a spectral jump in the range of~640 nm (between the spectral ranges of the two sensors), but does not affect the rest of the spectrum in any way. The laboratory test confirms that the Rikola camera can detect REEs under laboratory conditions. We, therefore, concluded that it could be possible to identify REE signatures in field conditions.