[Paper Review] DeepSD: Generating High Resolution Climate Change Projections through Single Image Super-Resolution

Thomas Vandal, Evan Kodra+4 more|arXiv (Cornell University)|Mar 8, 2017

Meteorological Phenomena and Simulations52 citations

TL;DR

DeepSD uses stacked super-resolution CNNs with multi-scale inputs to statistically downscale climate variables, improving high-resolution precipitation projections vs BCSD and other methods.

ABSTRACT

The impacts of climate change are felt by most critical systems, such as infrastructure, ecological systems, and power-plants. However, contemporary Earth System Models (ESM) are run at spatial resolutions too coarse for assessing effects this localized. Local scale projections can be obtained using statistical downscaling, a technique which uses historical climate observations to learn a low-resolution to high-resolution mapping. Depending on statistical modeling choices, downscaled projections have been shown to vary significantly terms of accuracy and reliability. The spatio-temporal nature of the climate system motivates the adaptation of super-resolution image processing techniques to statistical downscaling. In our work, we present DeepSD, a generalized stacked super resolution convolutional neural network (SRCNN) framework for statistical downscaling of climate variables. DeepSD augments SRCNN with multi-scale input channels to maximize predictability in statistical downscaling. We provide a comparison with Bias Correction Spatial Disaggregation as well as three Automated-Statistical Downscaling approaches in downscaling daily precipitation from 1 degree (~100km) to 1/8 degrees (~12.5km) over the Continental United States. Furthermore, a framework using the NASA Earth Exchange (NEX) platform is discussed for downscaling more than 20 ESM models with multiple emission scenarios.

Motivation & Objective

• Motivate the need for higher-resolution local climate projections beyond coarse Earth System Models (ESMs).
• Propose a spatially aware statistical downscaling method by adapting image super-resolution techniques to climate data.
• Develop DeepSD: an augmented stacked SRCNN framework using multi-scale inputs and topographic features.
• Evaluate DeepSD against established downscaling methods on CONUS daily precipitation.
• Discuss scalability to large ensembles of climate model simulations on NASA's NEX platform.

Proposed method

• Represent climate variables as image-like multi-channel data (precipitation, elevation, etc.).
• Construct three-layer SRCNNs to map low- to higher-resolution grids using patch-based learning.
• Stack SRCNNs with independent training at each upscaling step to form DeepSD.
• Incorporate static high-resolution topography (elevation) as input channels at each scaling level.
• Train with Adam optimization; use 51x51 patches; replicate padding at test time to maintain sizes.
• Evaluate against BCSD and ASD approaches using metrics like Bias, Corr, RMSE, and Skill.

Experimental results

Research questions

• RQ1 Can a stacked super-resolution CNN framework exploit spatial dependencies in climate data to improve downscaling accuracy over traditional methods?
• RQ2 Does incorporating high-resolution topographic features improve downscaled precipitation performance across scales?
• RQ3 Are DeepSD downscaled outputs robust across seasons and capable of handling extreme precipitation better than baseline methods?
• RQ4 Can DeepSD scale to downscale ensembles of climate model simulations on platforms like NASA’s NEX?

Key findings

• DeepSD outperformed Lasso, ANN, and SVM in bias, correlation, and RMSE on 1000 CONUS locations.
• DeepSD achieved higher correlation (0.914) and lower RMSE (2.529 mm/day) than BCSD (Corr 0.849, RMSE 4.414 mm/day).
• DeepSD provided competitive skill (0.947) to BCSD (0.955) while offering faster runtimes (DeepSD 71s vs BCSD 13s per year).
• DeepSD maintained strong performance across seasons, with higher correlation and lower RMSE than BCSD overall.
• Extremes: DeepSD remained more stable than BCSD when downscaling increasingly extreme precipitation events.