A Machine-Learning Method to Integrate Arctic Supersite Observations and Diagnose Weather Element Occurrence

Abstract

The accurate detection and quantification of light precipitation is problematic, particularly in the Arctic region. Satellite and ground-based observations of light precipitation are frequently underestimated at high latitudes. Remote sensing and in-situ observations from the Iqaluit, NU supersite (64oN, 69oW) were integrated to train, develop, and validate a random forest (RF) model that can diagnose precipitation type and other weather element occurrences. Observations from multiple lidars, optical disdrometers, traditional precipitation gauges and meteorological aerodrome (METAR) reports from 2015–2020 were integrated and used in the RF model development. The model was trained at Iqaluit, validated over different time periods, and applied to another region (Whitehorse, YT; 61oN, 135oW). Results indicate the importance of accurate visibility observations to train the model. Overall, the RF model was capable of distinguishing precipitation types and demonstrated the potential to be used at all sites/networks where similar automated and cost-effective instruments are already deployed (e.g. radar sites, airports with ceilometers, etc.). This would reduce the dependency on METARs while improving weather element occurrence accuracy.