Machine-learning Based Max-Tree Image Segmentation for Automated Identification of Coherent and Quasi-Coherent Waves in Plasma Fluctuation Spectrograms*

A Machine-learning based analysis—max-tree image segmentation [P. Salembier, et al. IEEE Trans. on Image Proc. (1998)]—is being developed to automate identification of coherent or quasi-coherent waves in measured plasma density fluctuation spectrograms and determination of their properties. Results from Monte Carlo simulations with synthetic signals are presented that show how to distinguish coherent or quasi-coherent waves from background instrumental noise and turbulence. Efficiently determining plasma waves that can grow and cause cross-field transport of particles and energy advances our understanding. The simulations show the probability of noise producing high-fluctuation power wave-like features with various properties, including the area of the feature as well as the ratio of signal to background level. For instance, the probability of noise features with different areas in the spectrogram, P(area), scales as area^-1.4 for large areas. These results provide guidance for distinguishing features produced by noise from those produced by coherent or quasi-coherent waves present in the plasma. The resulting analysis tool is intended to replace subjective visual identification of waves in plasma fluctuation spectrograms and validation of simulations.