Turbulence Profiling Neural Networks Using Imaging Shack-Hartmann Data for Wide-Field Image Correction

Wide-field image correction of turbulence-induced phase requires tomographic reconstruction of each layer of turbulence. Before reconstruction can occur, the layers must be counted and ranged. A new signal-to-noise ratio metric for detecting a single layer of turbulence in a multi-layer atmosphere from SLOpe Detection And Ranging (SLODAR) measurements of Shack-Hartmann wave-front sensor (SHWFS) data is presented. 12,000 1-4 layer atmosphere profiles are procedurally defined by Fried length, layer altitude, and a minimum layer SNR requirement. Each profile is measured in simulation by a SHWFS in a 1.5 meter telescope with a 2.5 arcminute field of view over a 200 millisecond window. The simulation outputs are used as a 5-fold cross validation training data set for convolutional neural networks (CNNs) that count and range layers. The counting network achieved 92.6% accuracy and all ranging networks scored above 97.8% validation accuracy. We find that layers with SNR below 1 accounted for a majority of the misclassified points for all networks. We conclude that CNNs are a good candidate for wide-field image correction systems imaging through turbulence due to their ability to accurately profile the atmosphere from short time windows of collected data.