**3. The SFE2D architecture**

To separate the spectral features, we created a 2D feature extraction scheme that combines 2D CWT with variance, kurtosis, and negentropy maximization algorithms (PCA, kICA, and nICA) to extract spectral features from wavelet spectra. The algorithm consists of three main stages (**Figure 5**):


An effective spectral feature extraction depends on computer hardware specifications, specifically for larger numbers of spectral features produced by increasing the number of scales and mother wavelet's directions. In practice, users can create a larger number of features by changing the CWT parameters in the SFE2D program.

**Figure 5.** *Schematic view of the SFE2D procedure.*

For large numbers of CWT features (100–1000 CWT features), PCA/ICA methods provide a way to decorrelate and separate spectral features from CWT images.

If needed, users can also reduce the dimensionality to summarize the features. By running the proposed workflow in the SFE2D program, the redundant frequency volumes could be reduced to a more manageable number of components. Taking advantage of the ICA statistical properties, we can keep the most geologically pertinent information within the spectral decomposed data.
