Wideband Model Order Estimation Using Machine Learning

We apply two classification machine learning methods, one versus all logistic regression and a simple neural network, to estimate the model order or number of sources for parametric direction of arrival estimation using a wideband dataset with low snapshots and a range of signal to noise ratios. This particular scenario is of interest because it is common in the radar ice sounding datasets collected by the Center for Remote Sensing of Ice Sheets. We compare results with six standard model order estimation methods that use various functional forms for the cost function. We also compare against a numerically tuned method which is related to the machine learning methods, but the simple optimization method was not as capable as the machine learning methods and the cost function formulation is different. We find that the machine learning based methods outperform all the compared methods for our test cases, suggesting that this is a promising solution to the wideband model order estimation problem.