Machine learning: A viable option to improve precipitation retrievals in cold regions

Reza Ehsani and Ali Behrangi
Department of Hydrology and Atmospheric Sciences
The University of Arizona

Machine Learning (ML) is a term in computer science, but recently, it has received tremendous attention from the entire scientific community. Due to the popularity of ML, hydrologist and atmospheric scientists have also started incorporating ML techniques to address challenging issues. Precipitation retrieval has been an exigent topic, especially in high latitudes (i.e. poleward of 50°) and over frozen surfaces. Given the importance of quantifying high‐latitude precipitation, and ample challenges that the current precipitation products face over these regions, the present research investigates precipitation retrieval in higher latitudes using several ML algorithms. CloudSat provides direct observations of snow   and light rainfall at high latitudes with unprecedented signal sensitivity which can be considered as the baseline for precipitation rate. However, due to its nadir-only observation, it does not provide sufficient temporal sampling. ML techniques can help us to retrieve precipitation rate using brightness temperature from Microwave Humidity Sounder (MHS) and Advanced Very High-Resolution Radiometer (AVHRR), providing a valuable alternative for precipitation estimation in high latitudes. For this purpose, we have matched up Cloudsat, MHS and AVHRR data for the period 2007-2010 to create a database for ML training and testing. Then precipitation rate is retrieved from brightness temperature at different frequencies (~11 to ~190 microns) together with climatic variables such as near-surface air temperature. The results indicate that ML algorithms are capable of both identifying precipitation events and estimating precipitation rates with relatively high accuracy compared to the current physically-based retrieval methods in high latitudes.

Video Presentation

Back to Agenda | Back to All Abstracts