2 Location and Geological setting
3.4 Classification Mapping of Alteration Patterns
Chapter 3 42
Electron Microprobe (EMP) measurements were performed using a JEOL JXA 8900 RL instrument and included low-resolution BSE images for overview, accumulated BSE images as high-resolution imaging method and quantitative wavelength dispersive analyses of alteration minerals. The instrument was run at 15 kV potential and at 10 µm beam diameter.
For the thin sections signals for Si, Al, Na, K, Ca, Ba, Mg, Sr and Fe were detected on the five spectrometers, while additionally also the BSE signal was recorded in the COMPO-mode.
Grain mounts were only analyzed qualitatively, as quantitative analysis was not possible due to small grain size, low crystalinity, and mineral instability under the beam.
3.3.3 X-ray Diffraction (XRD)
We measured 35 powdered samples (clay to silt fraction) with a Philips powder diffractometer at the Dept. for Cristallography in Göttingen. Measuring time was 5 and 15 seconds (day and nighttime measurements respectively) at 0.2 ° 2-theta. Scree samples were dried before measurement at 110 °C.
3.4 Classification Mapping of Alteration Patterns
The aim of spectral classification was to map ignimbrite outcrops and alteration patterns based on the results from ground-truth samples, correlated spectral measurements and geochemical analysis. For this purpose we tested various methods in order to find the best one for our particular regional, spatial and spectral setting.
The classification method used was the Support Vector Machine Classification algorithm available in ENVI. This algorithm separates classes using a decision surface that maximizes the margin between different classes (optimal hyperplane). Detailed information about this classification algorithm can be found in Boser et al. (1992), Chih-Wei Hsu and Chih-Jen Lin (2003) and Cortes and Vapnik (1995.).
The classification was conducted in two steps using twelve different training sites (ROIs) (see table 1). ROIs were chosen on the basis of detailed fieldwork and represent different alteration settings and lithologies. The nine ASTER reflectance bands (after applying the vegetation suppression algorithm provided in ENVI) were combined with 7 bands of Minimum Noise Fraction (MNF) transformed (Green et al. 1988) reflectance data. The two last eigenchannels were removed for classification as a method of noise suppression.
Additionally, the surface emissivity product was used for classification, as thermal infrared provides additional information about silicate minerals present (Hecker et al., 2012). This
Manuscript II ‐ Mapping patterns of mineral alteration in volcanic terrains using ASTER data and field spectrometry in Southern Peru
Chapter 3 43
data is derived from level 1B radiance at sensor data after atmospheric correction and separation of the emissivity component from kinetic temperature component (Gillespie et al., 1998).
Examples of the spectral characteristics of the respective training areas are given in Fig.
4A and B. Although differences in the nine reflectance bands are not obvious, MNF transformation allows enhancement spectral differences. The addition of these bands increases dimensionality in the dataset although some information might also get lost using MNF bands only. Emissivity data (zoom) shows subtle differences for the respective classes with the “lava” class showing highest values. The composite mask was applied and the SVM classification was started. The penalty parameter was set to 50 and the probability threshold to 0.5. The gamma parameter automatically set by ENVI was not changed. Classes and reference samples for the definition of ROIs are given in Table 1.
As the automatic classification output proved insufficiently accurate, the rule classifier was used to optimize class thresholds. The “Soil/Regolith” class was set to background value. The accuracy of the classification was assessed by calculating a confusion matrix using a second set of ground-truth ROIs that was selected together with the ROIs used for classification but kept apart for quality assessment. The resulting class image was exported to ArcGIS. For the “ignimbrite” class a buffer image was calculated in order to better highlight isolated pixels in the image. The classes defined by the ROIs and separated by this procedure are summarized in Table 1.
Fig. 4: A: Plot showing examples for the different classes of SVM classification. The first nine bands are ASTER reflectance data (in % multiplied by 10), bands ten to 14 are emissivity data (scale factor 100), bands 15 to 21 are MNF transformed bands that were used to increase dimensionality and to reduce noise. B: Zoom to the emissivity bands (square in Fig. 5 A) in order to highlight spectral differences.
Table 1: Description of ROIs used for classification and reference samples.
Regions of interest for SVM classification
Class Description Reference sample*
Zeolithe Fault zone in ignimbrite at Lomada Atansa with zeolithes ATAN‐11‐08, ATAN‐11‐09, ATAN‐11‐05 ORCO scree
Alteration at Cerro Orconccocha with a high degree of silicification, clay minerals, iron oxides,
rutile ORCO‐11‐ 10
CHAV scree higher‐T
Alteration at Cerro Palla Palla with clay minerals, a high degree of amorphous silica and some
oxides CHAV‐11‐09, CHAV‐11‐12
CHAV high‐T silicification Silicified rock at Cerro Palla Palla, up to 99 wt‐% SiO2
CHAV‐11‐08, CHAV‐11‐07, CHAV‐11‐06, CHAV‐11‐17
CHAV scree Alteration scree at Cerro Palla Palla, clay minerals, amorphous silica
CHAV‐11‐01, CHAV‐11‐02, CHAV‐11‐03, CHAV‐11‐04
Iron oxides Alteration with dominating iron oxides and clay minerals at Cerro Palla Palla CHAV11‐16, CHAV‐11‐14
Weathered Ignimbrite Weathered ignimbrite, some clay minerals ATAN‐11‐01
CCAR scree Alteration scree at Cerro Carhuarazo, amorphous silica, iron oxides CCAR‐11‐08, CCAR‐11‐09 CCAR scree with clay
minerals Alteration at Cerro Carhuarazo, clay minerals, amorphous silica, iron oxides only spectral indices Soil/Regolith (unclassified) Covered to partially covered ground only validation in the field
Lava Lava flows and volcanic rocks only validation in the field
Ignimbrite Ignimbrite outcrops only validation in the field
* training areas were determined using reference samples, spectral indices and ground‐truthing experience.
Manuscript II ‐ Mapping patterns of mineral alteration in volcanic terrains using ASTER data and field spectrometry in Southern Peru
Chapter 3 45
4 Results and Discussion