CISL Seminar Series: Deep Learning for Downscaling Gridded Daily Temperature and Precipitation in Complex Terrain

Title: Deep Learning for Downscaling Gridded Daily Temperature and Precipitation in Complex Terrain

Speaker: Mr. Yingkai (Kyle) Sha

Department of Earth Ocean and Atmospheric Sciences, University of British Columbia Global Climate Models (GCM) cannot resolve surface meteorological variables over complex terrain properly because of their coarse grid spacing and overly smoothed terrain input. Statistical downscaling is proposed for bridging the gap between GCM outputs and their various downstream applications (e.g., hydrology and climate risk management), which requires localized and fine grained meteorology fields. Traditional gridded statistical downscaling methods are challenged by their limited generalization abilities across different regions, seasons and numerical model inputs. This study provides a novel Convolutional Neural Network (CNN) gridded downscaling approach for 2-m daily minimum/maximum temperature and precipitation. The CNN-based downscaling takes low resolution GCM temperature/precipitation, elevation and ETOPO1 high resolution elevation as inputs, performs resolution enhancement from 0.25-degree to 4-km and is trained on Parameter-Elevation Regressions on Independent Slopes Model (PRISM) data. Through the evaluation of metrics of bias, texture and distribution, this CNN-based downscaling approach has demonstrated improved performance over a single-site regression approach. It can be generalized to perform well across different spatial areas, including those not included in training area. The CNN also can downscale inputs from different GCMs. The CNN-based downscaling proposed by this study is especially useful for regions where in-situ observations and high resolution gridded analysis are not available.

Biography Yingkai (Kyle) Sha is a second year PhD candidate from Department of Earth Ocean and Atmospheric Sciences, University of British Columbia whose research includes the adaptation of deep learning methods on the bias-correction, downscaling and evaluation of weather and climate model outputs.

Refreshments will be served at 9:45 a.m.!

Date: Monday, December 2, 2019

Time: 10:00am

Location: Mesa Lab, Main Seminar Room

 

Building:

Room Number: 
Main Seminar Room

Type of event:

Will this event be webcast to the public by NCAR|UCAR?: 
No
Announcement Timing: 
Wednesday, November 20, 2019 to Monday, December 2, 2019
Calendar Timing: 
Monday, December 2, 2019 - 10:00am to 11:00am

Posted by Taysia Peterson at ext. 1222, taysia@ucar.edu

Quick Facts

NCAR is managed by the nonprofit University Corporation for Atmospheric Research on behalf of NSF and the UCAR university community.

  • NCAR is not a federal agency and its employees are not part of the federal personnel system.
  • Our activities complement those of the federal agencies and we work closely with them.

More about us