Deep Learning for High-Resolution Precipitation Downscaling: Case Study in New Jersey
Presentation Type
Poster
Faculty Advisor
Clement Alo
Access Type
Event
Start Date
26-4-2024 9:45 AM
End Date
26-4-2024 10:44 AM
Description
Understanding the local-scale impacts of precipitation, particularly during extreme events, requires fine-scale data at regional levels. Several statistical downscaling methods have been developed to refine Global Climate Model (GCM) output from coarse resolutions (100–600 km) to finer resolutions. However, deep learning models have recently emerged as a promising approach for this purpose. This study evaluates the effectiveness of a deep learning technique, specifically convolutional neural networks (CNNs), in providing high-resolution climate data for analyzing extreme precipitation events in New Jersey. A three-layer CNN structure was employed and trained using input variables from ECMWF Reanalysis v5 (ERA5), including wind, geopotential height, temperature, and humidity at different pressure levels (850mb, 700mb, and 500mb). The CNN model was then utilized to downscale daily Coupled Model Intercomparison Project Phase 6 (CMIP6) GCM data from 200km to 4km resolution. This downscaling was achieved through a transfer learning function that established a relationship between the GCM climate variables at coarse resolution and finer-scale observed precipitation. To evaluate the model’s extrapolation ability, a hold-out cross-validation approach was used,. with training data from 1980–2005 and validation data from 2006–2008. During the validation period, bias and root mean square error metrics showed that the deep learning model performed well at capturing the spatial distribution of rain in New Jersey. This suggests that the CNN model can effectively downscale the GCM data and generate accurate, high-resolution precipitation projections for the region.
Deep Learning for High-Resolution Precipitation Downscaling: Case Study in New Jersey
Understanding the local-scale impacts of precipitation, particularly during extreme events, requires fine-scale data at regional levels. Several statistical downscaling methods have been developed to refine Global Climate Model (GCM) output from coarse resolutions (100–600 km) to finer resolutions. However, deep learning models have recently emerged as a promising approach for this purpose. This study evaluates the effectiveness of a deep learning technique, specifically convolutional neural networks (CNNs), in providing high-resolution climate data for analyzing extreme precipitation events in New Jersey. A three-layer CNN structure was employed and trained using input variables from ECMWF Reanalysis v5 (ERA5), including wind, geopotential height, temperature, and humidity at different pressure levels (850mb, 700mb, and 500mb). The CNN model was then utilized to downscale daily Coupled Model Intercomparison Project Phase 6 (CMIP6) GCM data from 200km to 4km resolution. This downscaling was achieved through a transfer learning function that established a relationship between the GCM climate variables at coarse resolution and finer-scale observed precipitation. To evaluate the model’s extrapolation ability, a hold-out cross-validation approach was used,. with training data from 1980–2005 and validation data from 2006–2008. During the validation period, bias and root mean square error metrics showed that the deep learning model performed well at capturing the spatial distribution of rain in New Jersey. This suggests that the CNN model can effectively downscale the GCM data and generate accurate, high-resolution precipitation projections for the region.