AnalySize: New software for analyzing and unmixing sediment grain size distribution spectra

Abstract:

Grain size distribution (GSD) data are a widely used tool in Earth sciences, particularly in understanding sediment transportation and sourcing. Although large data sets are regularly generated, detailed numerical analyses, such as grain size unmixing, are not routinely performed.

Unmixing of GSD data involves approximating a given data set by a small number of GSDs, known as end members. These end members, along with their relative abundances, can be used to fully characterize the variability of the data. End member analysis (EMA), which fits one set of end members to a single data set, is one the most robust ways to do this. This approach estimates the form of the end members from the data set itself; hence it is a non-parametric approach. Available algorithms, however, either produce sub-optimal solutions, or are time consuming.

To aid investigators in exploring the full potential of their data, we introduce AnalySize, which is a GUI based tool that allows for comprehensive processing and unmixing of grain size data obtained from laser diffraction particle grain size analyzers. AnalySize brings together methods from other disciplines in Earth sciences as well as introducing new techniques and improvements to provide a complete software package for unmixing GSD data. The software utilizes the rapid HALS-NMF algorithm from hyperspectral image analysis to perform non-parametric EMA, which is demonstrated to yield results that are an improvement over algorithms currently used in GSD analysis.

Non-parametric EMA, however, is often unable to clearly identify discrete unimodal grain size sub-populations, which can more detailed information about sediment sources. To alleviate this, we introduce a new algorithm to perform parametric EMA, whereby an entire GSD data set can be unmixed into unimodal parametric end members (e.g., lognormal or Weibull end members). This allows individual grain size sub-populations to be more readily identifiable in highly mixed data set. This new approach overcomes the shortcomings of techniques used to unmix single specimen data, which are widely used, but inherently flawed. These new tools, along with a suite of standard analysis statistics and plots, are packaged in a user friendly GUI, which is capable of processing and wide range of data formats.