No abstract

We apply a multilayer perceptron machine learning (ML) regression approach to infer electromagnetic (EM) duct heights within the marine atmospheric boundary layer (MABL) using sparsely sampled EM propagation data obtained within a bistatic context. This paper explains the rationale behind the selection of the ML network architecture, along with other model hyperparameters, in an effort to demystify the process of arriving at a useful ML model. The resulting speed of our ML predictions of EM duct heights, using sparse data measurements within MABL, indicates the suitability of the proposed method for real-time applications. Key Points:• Artificial neural networks can quickly characterize electromagnetic wave propagation within the marine atmospheric boundary layer • Bistatic sampling of propagation factors is practical for estimating EM duct height with neural network models • The rationale underpinning the design of our machine learning models is carefully explained: to demystify the process

We show that Gaussian process regression (GPR) can be used to infer the electromagnetic (EM) duct height within the marine atmospheric boundary layer (MABL) from sparsely sampled propagation factors within the context of bistatic radars. We use GPR to calculate the posterior predictive distribution on the labels (i.e., duct height) from both noise-free and noise-contaminated array of propagation factors. For duct height inference from noise-contaminated propagation factors, we compare a naïve approach, utilizing one random sample from the input distribution (i.e., disregarding the input noise), with an inverse-variance weighted approach, utilizing a few random samples to estimate the true predictive distribution. The resulting posterior predictive distributions from these two approaches are compared to a "ground truth" distribution, which is approximated using a large number of Monte-Carlo samples. The ability of GPR to yield accurate and fast duct height predictions using a few training examples indicates the suitability of the proposed method for real-time applications.

No abstract