# Principal component analysis (PCA) and normalization of radiometric data?

I did some experiments with airborne radiometric data in a region in southern Brazil and I would like to know your suggestion about the following topic:

I tried to classify high, medium and low values of radiometric variables in the region by empirical methods (visualization only). Then I summarize the results in a table (Fig.1) where high values were attributed a red color, while medium and low were attributed yellow and blue, respectively. I did not estimate a range of values to define of what is high or medium, for example. I only checked the alkaline suites (labeled on the image) responses in comparison with their surroundings to classify. A example of the grids used as the basis of this classification are presented in the Fig.2 while the Fig.1 represents the table.

I would like to know a alternative of how can I try to find a better classification system for this type of variables (which consist most of ratios maps).

I thought in first normalize the data using the min-max normalization, where all the gridded images will have a range between 0 and 1. All my grids have different range of values and units. My advisor told me to use a technique of “mean + 2 or 3 standard deviation” to identify anomalies. But since the radiometric data are not presented in a normal distribution, I could not use this method. mean + 3 sd are usually identified as outliers in normal distribution and do not represent anomalies. So I think this idea could be discarded.

1) Is the min-max normalization a good idea? Or I should think in another type of normalization to analyze all the variables? Is there a more automated-driven approach that I could follow to classify the data?

The min-max normalization is given by the equation in the link: https://en.wikipedia.org/wiki/Normalization_(statistics) I have seen some works using z-score (standard score) for radiometric data. I could be wrong, but this type of normalization is suitable for when your population is normally distribute (Gaussian distribution), which is not the case for radiometric data.

The second step was to apply a Principal Component Analysis (PCA) or Self-Organizing Maps (SOM) on it. I would like to create a map of PC1 and PC2, for example. Something similar with the Lima and Marfurt 2018 (p.2279) maps (Fig.3). However, the authors did not cite the routine used in MATLAB.

2) Is there a option to create a similar map of Lima and Marfurt (2018) using Oasis? If someone can suggest some references about the topic and that was performed using Oasis Montaj or ArcGIS routines I would be grateful. References: Lima and Marfurt 2018. https://www.researchgate.net/publication/327611793_Principal_component_analysis_and_K-means_analysis_of_airborne_gamma-ray_spectrometry_surveys

• Have you not answered your first question (is there a better method) by the comments and reference leading to your second question? ... Use PCA. With regard to how to obtain the routines, have you considered contacting the authors directly? – Jeffrey J Weimer Jun 26 at 12:47