[Paper Review] Climate Downscaling: A Deep-Learning Based Super-resolution Model of Precipitation Data with Attention Block and Skip Connections

Chia-Hao Chiang, Zheng-Han Huang+7 more|arXiv (Cornell University)|Mar 26, 2024

Cryospheric studies and observations6 citations

TL;DR

The paper presents a deep CNN architecture with residual attention blocks, skip connections, and auxiliary topography data to downscale heterogeneous precipitation data (ERA5 to TCCIP) in Taiwan, outperforming alternative methods in multiple metrics.

ABSTRACT

Human activities accelerate consumption of fossil fuels and produce greenhouse gases, resulting in urgent issues today: global warming and the climate change. These indirectly cause severe natural disasters, plenty of lives suffering and huge losses of agricultural properties. To mitigate impacts on our lands, scientists are developing renewable, reusable, and clean energies and climatologists are trying to predict the extremes. Meanwhile, governments are publicizing resource-saving policies for a more eco-friendly society and arousing environment awareness. One of the most influencing factors is the precipitation, bringing condensed water vapor onto lands. Water resources are the most significant but basic needs in society, not only supporting our livings, but also economics. In Taiwan, although the average annual precipitation is up to 2,500 millimeter (mm), the water allocation for each person is lower than the global average due to drastically geographical elevation changes and uneven distribution through the year. Thus, it is crucial to track and predict the rainfall to make the most use of it and to prevent the floods. However, climate models have limited resolution and require intensive computational power for local-scale use. Therefore, we proposed a deep convolutional neural network with skip connections, attention blocks, and auxiliary data concatenation, in order to downscale the low-resolution precipitation data into high-resolution one. Eventually, we compare with other climate downscaling methods and show better performance in metrics of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Pearson Correlation, structural similarity index (SSIM), and forecast indicators.

Motivation & Objective

• Motivate the need for high-resolution precipitation downscaling in regions with strong climatic forcings and data heterogeneity.
• Develop a deep learning architecture that can correct bias and produce high-resolution precipitation from low-resolution inputs.
• Incorporate auxiliary topography data to improve downscaling accuracy.
• Compare the proposed method against statistical and machine learning baselines to demonstrate performance gains.

Proposed method

• Use a cascading convolutional bias correction backbone to transform low-resolution input into high-resolution feature maps.
• Incorporate skip connections to facilitate residual learning and fuse multi-level feature maps.
• Integrate residual attention blocks (CBAM-based) to emphasize channel and spatial features during learning.
• Apply a one-step image upscaling layer using pixel shuffle, followed by a corrective convolution to reduce artifacts.
• Fuse auxiliary elevation/topography data with processed features to enhance precipitation downscaling accuracy.
• Train end-to-end under supervised learning with MSE loss using ERA5 as input and TCCIP observations as ground truth.

Experimental results

Research questions

• RQ1 Can a deep CNN with residual attention and skip connections effectively downscale heterogeneous precipitation data from ERA5 to high-resolution observations?
• RQ2 Does incorporating topographic auxiliary data improve downscaling performance in a region with strong climatic forcings like Taiwan?
• RQ3 How does the proposed model compare to statistical downscaling and other ML/DL climate downscaling approaches across multiple evaluation metrics?

Key findings

• The proposed model delivers superior downscaling performance compared with baselines across multiple metrics including MAE, RMSE, Pearson correlation, and SSIM.
• The architecture is specifically designed to handle heterogeneous data pairs and strong regional forcings, improving bias correction.
• Topography data fusion contributes to higher-quality downscaled precipitation outputs.
• The model demonstrates favorable qualitative and quantitative results against methods such as YNet, FSRCNN-ESM, DeepSD, QM, BCSD, and others in the study.
• Training on ERA5 inputs with daily precipitation targets from TCCIP over Taiwan shows effective downscaling with a small input-to-output scaling factor.