Progress in Machine Learning Based Detection and Management of Sea Clutter for the Scanning Radars at the ARM ENA Site

Description

The quality of radar measurements performed by ARM is subject to significant impacts from non-meteorological echoes, with the source and severity depending on location and radar operating parameters. At coastal locations, sea clutter, significant signal (above the detection level) scattered by the ocean surface is a primary concern for both X-band and Ka/W-band radars scanning at low elevation angles. Within the past few years, a surge in progress has considerably advanced machine learning skill in the area of image analysis (Lecun 2015). Casting our application, the identification of features in a two dimensional array of radar data (polar or cartesian), as a deep learning image analysis problem provides a direct opportunity to capitalize on this progress. Using the convolutional neural network (CNN) architecture, we treat radar data as images of pixel depth N, where each pixel is an N-dimensional vector consisting of all relevant moments and polarimetric variables. While CNNs typically operate on images of pixel depth one or three (e.g. monochrome or RGB images), this is by no means a requirement. We address the clutter detection problem in two phases. Our first (supervised learning) phase uses use a human “expert” to label datasets of radar images into the classes of “no echo”, “clutter-only echo”, “hydrometeor-only echo”, or “clutter plus hydrometeor echo”, plus an “unclassified” class. From these, a CNN-based deep neural network classifier is trained and tested. The second (unsupervised learning) phase capitalizes on differences in the spatiotemporal properties of clutter and meteorological echoes to extend skill beyond phase one. For example, clouds and precipitation form clusters that coherently advect through clutter fields in a way that can be recognized in a sequence of scans. Here we present our progress in this effort.