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

TL;DR

This work tackles high-resolution precipitation downscaling in regions with strong climatic forcings by introducing a deep learning model that integrates cascading bias correction, residual skip connections, and residual attention blocks. The architecture uses a one-step pixel-shuffle upscaling layer and fuses high-resolution topography as auxiliary input to enhance local-scale accuracy. Evaluated on ERA5–TCCIP data for Taiwan, it outperforms statistical baselines (QM, BCSD) and several DL-based approaches (DeepSD, FSRCNN-ESM, YNet) across $MAE$, $RMSE$, $Corr$, and $SSIM$, as well as forecast indicators like $POD$, $FAR$, and $TS$. The results demonstrate the practical potential of ML-based climate downscaling for coastal/island regions, while highlighting opportunities in model interpretability and cross-domain integration with GCMs/RCMs.

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.

Figures (16)

• Figure 1: Homogeneity and heterogeneity of the datasets. (a) Homogeneous data pair. One can still tell that the low-resolution one (downsampled one, on the left-hand side) is the similar to the high-resolution one (original data, on the right-hand side). The downsampled data is structural similar to the origin one, that there are same patterns at corresponding regions, but only with some lacks of high-frequency information in low-resolution one. (b) Heterogeneous data pair. It is hard to tell the structural similarity between them through patterns, since one is from ERA5 reanalysis data (left-hand side), and the other is from observations (right-hand side). Neither patterns nor the value ranges are similar.
• Figure 2: A simple building block with a skip connections. The skip connection (or "shortcut connection") is to feed forward an identity $x$ to the output of a series of layers. It is commonly used in a deep neural network for training optimization.
• Figure 3: Pixel shuffle from ESPCN. This one-step upscsaling layer rearranges the pixels along channel axis into a larger image. The number of channels needs to be adjusted to the square of the scaling factor. One can change the number of filters of the convolutional layer right before pixel shuffle layer for preferred scaling factor.
• Figure 4: Diagram of attention blocks. As illustrated in (a), a residual attention block (RAB) consists of a channel attention block (CAB) and a spatial attention block (SAB), and a skip connection from input feature map forwarding to refined feature map. (b) shows the structure of CAB, including a global max pooling, a global average pooling, and a shared multi-layer perceptron unit. (c) illustrates the structure of SAB, including a pair of channel-wise max and average pooling, a concatenation layer, and a convolutional layer. Both activation functions in (b) and (c) are using sigmoid activation.
• Figure 5: Overview of proposed model architecture. Main stream of the model is a series of residual attention block, attached with a convolutional layer and a ReLU activation. Refined intermediate feature maps are fed into a convolutional layer with 1$\times$1 kernel size to limit the number of parameters and pick up the most significant one at different levels for forwarding. The first fusion layer combines all refined feature maps, while the second one additionally concatenate elevation data and interpolation of original input. The image upscaling layer is adopting the pixel shuffle technique from ESPCN, after a convolutional layer with the number of filters equal to square of scaling factor.