*5.4. Oxidized Surface Abundance*

The oxidized surface endmembers (3, 6, and 12) have the highest abundance fraction at the vent as shown in Figure 8a. This agrees with a field observation shown in Figure 8b which highlights the matching dominant oxidized surface at the vent wall.

**Figure 8.** (**a**) Abundance map oxidized surface endmember; yellow areas indicate the highest fraction of oxidized surface meanwhile the black areas indicate the lowest fraction of oxidized surface; (**b**) Field photograph of an oxidized surface of the vent wall (red box and the line shows the approximate location of the field photograph)

#### *5.5. Sulfate Mineral Abundance*

The sulfate mineral endmembers (2, 7, 10, 11, and 15) have the highest abundance fraction around the lava pond and there are four most prominent areas for the sulfate (Figure 9a). This surface mineral looked as if it had been dusted by snow (white color) commonly identified as *thernadite* (*Na*2*SO*4) [41]. This can be directly seen from a true color image. This mineral formed as the flow cooled, a thin sublimate coating formed on the surface of the lava [41]. Figure 9b,c shows the *thernadite* formed in surface lava at Holuhraun.

**Figure 9.** *Cont.*

(**c**)

**Figure 9.** (**a**) The Abundance map for the sulfate mineral endmember, the yellow areas indicate the highest fraction of sulfate mineral meanwhile the black areas indicate the lowest fraction of the sulfate mineral; (**b**) Field photograph of sulfate mineral (white surface) formed on the surface of lava (the red boxes and lines show the approximate location of the field and aerial photo respectively); (**c**) aerial photograph of sulfate mineral (white surface) formed on the surface of lava (The numbers on the image indicate the approximate location of the sulfate for both the abundance and photograph).

#### *5.6. Water Abundance*

The water abundance (Figure 10a) has the highest abundance fraction at the location mainly recognized as a glacial river (Figure 10b). Endmember 14 represents water which is characterized by a relatively low reflectance and has the highest reflectance in the blue wavelength. Water has high absorption and virtually no reflectance in the NIR-SWIR wavelengths range (Figure 5f).

**Figure 10.** (**a**) The abundance map for water endmember, the highest abundance fraction indicated by a yellow color, and the lowest abundance fraction indicated by a black red box shows the approximate location of the aerial photograph); (**b**) aerial photograph of the glacial river.

#### *5.7. Noise Abundance*

Figure 11 shows the abundance map corresponding to endmember 9. We consider this endmember as representing noise due to an unrecognized spectral signature since this spectrum is characterized by saturated reflectance in channels ~2000 nm and ~2400 nm (Figure 5g). The saturated reflectance could be due to corrupted bands in some pixels.

**Figure 11.** The abundance map for the noise endmember, the highest abundance fraction is indicated by the yellow color, and the lowest abundance fraction is indicated by the black color.

#### *5.8. False Color Abundance*

The abundance results depicted as false color (R: Oxidized surface; G: Sulfate mineral; B: Basalt) images show that the majority of rocks or minerals in the study area are dominated by basalt as shown in the blue color in Figure 12a. The other colors such as magenta and yellow indicate a mixture. The mixture phenomenon is illustrated in Figure 12b, as the surface has 0.25 oxidized surface mix with 0.75 basalt resulting in the magenta color; and 0.25 oxidized surface mix with 0.75 sulfate mineral resulting in the yellow color pixel.

**Figure 12.** (**a**) False color of abundance highlighting for R: oxidized surface; G: sulfate mineral; and B: Basalt; (**b**) Illustration of the mixed pixels in the area, 0.25 oxidized surface mix with 0.75 basalt resulting in the magenta color; and 0.25 oxidized surface mix with 0.75 sulfate mineral resulting in the yellow color pixel.

#### *5.9. Validation*

The very high-resolution aerial photograph was used for ground truth. The aerial photograph was classified into oxidized surface, sulfate, basalt, and water using visual image interpretation and used for validation of the unmixing results. We only validate three endmembers for basalt—oxidized, sulfate, and water—since the noise and hot material cannot be detected based on visual interpretation. We classified the endmembers that have fractional abundance > 0.5. Validation was based on 150 randomly generated point samples within each class. Table 1 show the validation results, with a resulting mean overall accuracy 79% and mean Kappa index of 0.73. This result shows that the abundances have moderate agreement with the sample points.


**Table 1.** Validation of the endmembers that have abundance > 0.5.

#### **6. Discussion**

#### *6.1. Comparison with the Existing Spectral Index Technique*

The correlation between the spectral index images and the abundance image was analyzed. We only correlated the three endmembers since there are no reference spectral indices for sulfate mineral, hot material, and noise. Here we compared the basalt, oxidized, and water abundance images with the mafic, oxidized, and water index images proposed by Inzana et al., Podwysocki et al. and Xu respectively [42–44] (Appendix B). We applied these indices to the hyperspectral image and compared them with the result from each abundance. Figure 13a–c shows the scatter plots results. The *R<sup>2</sup>* values were 0.46, 0.91, and 0.77 for the basalt, oxidized surface, and water, respectively. The oxidized surface and water indicate a good correlation with the indices (Figure 13b,c). This suggests that both oxidation and water generated from a spectral index are properly validated [2,44]. Meanwhile, basalt shows a low correlation with the mafic index (Figure 13a) suggesting that the estimates of the basalt surface from the unmixing technique is an overestimation, since the basalt abundance shows the older lava flows as mafic with a relatively high fraction compared to the mafic index that only showed for fresh lava flow. This being due to a full spectrum of hyperspectral can easily differentiate between basalt surface and non-basalt.

**Figure 13.** *Cont.*

**Figure 13.** Linear regression analysis between the spectral index images and the (**a**) basalt abundance; (**b**) oxidized surface abundance; and (**c**) water index.

#### *6.2. Number of Endmembers*

The determination of the number of endmembers is critical, since underestimation may result in a poor representation of the mixed pixels, whereas overestimation may result in an overly segregated area [16]. Table 2 shows the relationship between the number of endmember and the number of pixels that have fractional abundance > 0.5 and the mean correlation with mafic, oxidized, and water index. We considered abundance >0.5 as high abundance. As the number of endmembers increase, the number of pixels also increases for an oxidized surface, sulfate mineral, water, and noise abundances, respectively. This is due to an increase of endmembers that is detected for each group. Meanwhile, the basalt abundance shows the opposite, as the endmembers increase the number of pixels with abundance >0.5 decreases. These results show that as more endmembers are considered the mixing of basalt with other endmembers increases resulting in a decrease of the fractional abundance of basalt. According to the results, we considered the 15 endmembers as an optimum number for this study since they have the highest mean correlation with mafic, oxidized, and water index. Clearly, the selection of appropriate endmembers in such a diverse volcanic environment, considering the particularities of the FENIX dataset, is of great importance in order to obtain accurate unmixing results. In addition, since only a small number of the available materials spectra are expected to be present in a single pixel, the abundance vectors are often sparse [45].


**Table 2.** Comparison number of pixels that have abundance >0.5, *R*<sup>2</sup> and number of endmembers.

#### *6.3. Size of Lava Field Area*

As the methods were only tested on a subset area of the lava field vent, to apply the methods for the entire lava flow is challenging for several reasons. (1) The high spatial heterogeneity typically gives rise to mixed pixels containing multiple materials and it will increase the number of endmembers detected by SMACC [40]. (2) Different illumination occurs within the different flight lines for the entire lava flow (Figure 3b) since the data acquisition time is acquired between 16.56 and 17.58 local times which results from the very low sun angle during the acquisition. This problem can be approached by collecting ground truth spectra, extensive calibration, and atmospheric correction using simultaneous and constrained calibration of multiple hyperspectral images through a new generalized empirical line model purposed by Kizel et al. [37]. (3) The computation time to perform unmixing also must be considered for the entire lava field since the area is relatively large (84 km2) and the hyperspectral data contains 622 channels with a 3.5-meter spatial resolution. In order to process the full set of data we need to consider using high performance computing (HPC) [46].

#### *6.4. Using Full Optical Region for Mapping Recent Lava Flow (VIS-SWIR-TIR)*

Hyperspectral VIS-SWIR image data is effective for discrimination mafic, oxidation, sulfate etc. However, not all the minerals and surface type are always mapped uniquely with VIS-SWIR hyperspectral data. A typical surface such as rock forming minerals associated with unaltered rocks and alteration minerals associated with altered rocks can be identified with TIR (Thermal Infrared) data [47–49]. Image processing methods that have become standard for hyperspectral VNIR/SWIR data analysis also work for hyperspectral TIR data [47]. Vaughan et al [47] showed that pixel classification techniques based on spectral variability within the scene and mineral libraries for matching spectral emissivity features can be used for TIR-derived mineral maps using SEBASS hyperspectral TIR image data. Hyperspectral TIR instruments operational for airborne surveys are also available in the NERC Airborne Research Facility with a Specim AisaOWL sensor [48]. A synergistic use of airborne data from both FENIX (VIS-SWIR) and OWL (TIR) allows great potential for lava discrimination in future study due to the complementary nature of the reflective (VIS-SWIR) and emissive (TIR) spectral regions. This might significantly improve our understanding of physical lava surface properties. Specifically, VIS-SWIR imaging spectrometers can discriminate surface materials and TIR data acquisitions can help to identify the thermal characteristics of different materials [47–49]. For instance, combining emissivity spectra with reflectance spectra in a mixing model would improve discriminating lava from surfaces [50–52].

#### **7. Conclusions**

In this study, an application of potential spectral unmixing methods on 2014–2015 Holuhraun lava flow field was presented. In total, we acquired fifteen spectral endmembers and their abundances. The first endmember was chosen as the brightest pixel which represented saturated incandescent lava. We grouped these 15 endmembers into six groups (basalt, oxidized surface, sulfate mineral, hot material, water, and noise) based on the shape of the endmembers since the amplitude varies due to illumination conditions, spectral variability, and topography. The endmembers represent pure surface materials in a hyperspectral image. We concluded that the selection of appropriate endmembers in such a diverse volcanic environment, considering the particularities of the FENIX dataset, is of great importance in order to obtain accurate unmixing results. Combination of SMACC and LSMA methods offers an optimum and a fast selection for volcanic products segregation However, ground-truthing spectra are recommended for further analysis. A synergistic use of airborne data from both FENIX (VIS-SWIR) and OWL (TIR) gives a great potential for lava discrimination in future study due to the complementary nature of the reflective (VIS-SWIR) and emissive (TIR) spectral regions. This might significantly improve our understanding of physical lava surface properties.

**Author Contributions:** Conceptualization, M.A.; Supervision, A.H., M.O.U., I.J. and T.T.

**Funding:** The first author was supported by the Indonesia Endowment Fund for Education (LPDP) Grant No. 20160222025516, European Network of Observatories and Research Infrastructures for Volcanology (EUROVOLC), The European Facility for Airborne Research (EUFAR) and Vinir Vatnajökuls during his Ph.D. project.

**Acknowledgments:** Authors would like to thank Robert Askew and Catherine Gallagher from the Institute of Earth Sciences, University of Iceland for the fieldwork photos around the lava field. Authors also would also like to thank anonymous reviewers for their constructive comments for the manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A**

The bad channels in this data are located at 968 nm and 1014 nm. Figure A1 show the spectral reflectance before masking (Figure A1A) and after channel masking (Figure A1B).

**Figure A1.** Spectral reflectance of material (**A**) before masking; (**B**) after channel masking.

Figure A2. This shows the theoretical pixel size at the nadir for Fenix. The pixel size will be larger at the edges of the swath, in this study, the AGL is ~ 2400 m so according to the graph the optimal pixel size resample for the FENIX is ~3.5 m.

**Figure A2.** The theoretical pixel size at the nadir for FENIX, EAGLE, and HAWK. In this study we used FENIX airborne for data acquisition [39].

#### **Appendix B**

The mafic indices originated, developed by Inzana et al. [42] to distinguish mafic from non-mafic rocks are from Landsat TM image, expressed as follows:

$$\text{Mafic index} = \frac{\rho\_{1600\text{nm}}}{\rho\_{860\text{nm}}} \ast \frac{\rho\_{640\text{nm}}}{\rho\_{860\text{nm}}} \tag{A1}$$

where *ρ*1600nm is the measured reflectance at wavelength 1600 nm, *ρ*640nm is the measured reflectance at wavelength 640 nm, and *ρ*860nm is the measured reflectance at wavelength 860 nm.

The oxidized index originated designed any multispectral sensor with bands that fall within the red channel and blue channel [43], expressed as follows:

$$\text{Oxidized index} = \frac{\rho\_{640\text{nm}}}{\rho\_{500\text{nm}}} \tag{A2}$$

where *ρ*500nm is the measured reflectance at wavelength 500 nm.

We calculated the water index using the Modified Normalized Difference Water Index (MNDWI) [44]. This index enhances open water features while suppressing noise from built-up land, vegetation, and soil. This is expressed as follows:

> *Water index* <sup>=</sup> *<sup>ρ</sup>*600nm <sup>−</sup> *<sup>ρ</sup>*1600nm *ρ*600nm + *ρ*1600nm

where *ρ*600nm is the measured reflectance at wavelength 600 nm.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Technical Note*
