1.2.2. Hyperspectral Imaging (HSI)

1.2.2. Hyperspectral Imaging (HSI) Hyperspectral images were generated using a Surface Optics 710VP camera, based on a dispersive system and a silicon sensor, providing a 4.5 nm spectral resolution in the 375 nm to 1047 nm range. Image dimensions are 696 × 520 pixels. Two 3200 K halogen lamps illuminate the object, and its intensity is determined right in front of the artwork using a solar power meter H110 series from Anaheim Scientific. Verification of the wave-Hyperspectral images were generated using a Surface Optics 710VP camera, based on a dispersive system and a silicon sensor, providing a 4.5 nm spectral resolution in the 375 nm to 1047 nm range. Image dimensions are 696 × 520 pixels. Two 3200 K halogen lamps illuminate the object, and its intensity is determined right in front of the artwork using a solar power meter H110 series from Anaheim Scientific. Verification of the wavelength calibration is achieved by using yellow, blue, red, and green reflectance standards

length calibration is achieved by using yellow, blue, red, and green reflectance standards CSS-04-020 AS-01178-060. Deconvolution of the light source and detector contributions

CSS-04-020 AS-01178-060. Deconvolution of the light source and detector contributions were calculated with Spectral Radiance Analysis Toolkit v3.5 using a Spectralon certified reflectance standard (model SRT-99-050 AA-00821-000) with a 5 cm × 5 cm reflective area. Classification, image analysis, and generation of hypercubes was performed with Harris Geospatial Solutions ENVI 5.5 software. HSI analysis methods used in this work allowed us to generate different images: pseudo color IR; spectral angle mapper (SAM); UV fluorescence; PCA-RGB composite image; and single PC images, which are described in detail in the work of Perez et al. [7], and references therein. Such images permit a detailed description of the techniques involved in the manufacture of the *Purísima Concepción.*

In brief, the classification workflow was: (i) images were background-subtracted and normalized to rule out contributions from light source and detector efficiency; (ii) endmembers were determined using the *pixel purity index* (PPI) criteria, and resulting pixels were processed following the process described in the work of Veganzones et al. [14] and Kale et al. [15]; (iii) endmembers were mapped by using the ENVI-SAM algorithm, thus resulting in an artificially colored image, where each color corresponds with a material with a characteristic spectrum; (iv) optimization of angle thresholds was achieved by using the ENVI *rule classifier* tool to determine the correct threshold angle for each class, and the results were tuned by examination of the corresponding angle distribution histograms, as in the work of Foglini et al. [16].

Principal component analysis (PCA) was performed to provide a direct classification of regions with spectral information variations, which could not be easily distinguishable in each of the *single-band* images. Resulting images provide a *reduced* set, in which material, spectral, or even manufacturing features are condensed. Most of the data variance is included in the first PC, thus we propose that an RGB-PCA image conformed with the first three PCs will significantly enhance such features. Mathematically, each PC is determined by a linear combination of the wavelengths, thus providing insights on chemical or physical properties linked with the associated spectra. We have shown in a previous work [7] that it is possible to go even further by considering a multidisciplinary approach for PCA image interpretation, allowing the description of manufacturing details on the studied object.
