Versatile Machine Learning-based detection technology for supercooled liquid clouds in polar regions

Clouds are a key regulator of Earth’s surface energy balance. The cloud cover along with cloud phase are the primary factors modulating the amount of radiation incident on the surface. While ice clouds tend to have small optical depths, liquid clouds that can be sustained down to -40C have large optical depths with a strong influence on the long-wave surface net flux [1]. Recent observational studies highlighted the ubiquity of supercooled liquid clouds (SLC) in both polar regions and how their presence can significantly enhance surface melt [2,3]. Global climate models generally fail to reproduce observed polar SLC occurrences, and thus do not represent the surface energy balance correctly. Sustained high quality SLC observations are paramount for challenging models in the highly climate change susceptible polar regions.

Cloud phase identification from active remote sensors in mixed-phase clouds remains challenging. The identification of SLCs is well established if lidar observations are available, but only for the lowest SLW cloud layer. Without lidar observations or in presence of multi-layered clouds (in which case the lidar signal is completely attenuated) millimeter-wavelength cloud Doppler radars can offer a viable alternative; however, in mixed‐phase conditions, ice crystals dominate the radar signal, rendering the detection of liquid droplets from radar observables more difficult. A machine learning (ML) approach that accounts for vast volumes of data accumulated and verified can represent a major breakthrough in this research area.

The project aims at developing a ML-based cloud mask&phase identification technology by exploiting the morphological features in the multi-frequency radar Doppler spectra and the spatial coherency of SLW clouds. The algorithm will be trained on extended datasets of collocated radar-lidar profiles and ancillary observations (microwave radiometer, and radiosonde measurements) collected at the North Slope of Alaska Atmospheric Radiation Measurement facility (https://www.arm.gov/capabilities/observatories/nsa) and will be tested on data gathered in other polar field campaigns including data from the AWARE campaign in Antarctica (https://www.arm.gov/research/campaigns/amf2015aware). Major focus will be on adapting ML solutions to data from various locations using stochastic separation theorems recently discovered at the University of Leicester [4,5]. In contrast to conventional approaches that require re-training of an AI system for different applications, the project will present an ideal test-bed to assess how well ML can adapt to conditions that are different from those encountered at the training site in Alaska (e.g. enhanced turbulence/more pristine aerosol environment are common in Antarctica) and how the algorithm will be able to correct on-the-fly erroneous detections.

Because of the ability of cloud radars to penetrate multiple liquid layers, this radar‐based technique will not suffer from the extinction limitations of lidars and will be able to provide unprecedented cloud phase identification, a critical step towards a thorough validation of global climate models. In addition to offering a solution to a critical scientific problem, the proposal creates a framework for sustained use of AIs at different locations over long time intervals, thus enabling the reconstruction of long climate record using the entire dataset collected at polar site facilities by cloud Doppler radars.