The goal is to increase the resolution of precipitation data into high quality image based results.
Our model comprises a series of convolutional layers with residual attention blocks as
our model spine, skip connections of feature maps at different levels, and a one-step
image upscaling layer. We utilize a cascade of convolutional layers to fulfill the tasks of bias correction and feature extractions. Attention blocks help model to learn “which to emphasize” and “where to focus”, locating the prominent values. Skip connections are used for avoiding gradient diminishing and decreasing training parameters. The one-step climate downscaling layers combine feature maps and apply a pixel shuffling layer for image upscaling. We add an influencing factor, topography data, to fusion layers which composed of another convolutional layers.
Above all, our main contributions are concluded below:
• We propose a deep learning model for heterogeneous precipitation simulation
data in climate downscaling problems with bias correction.
• The model is specially designed for an area with strong regional scale forcings
and can receive a very small number of precipitation data points from the simulation
and generates a corresponding high-resolution output.
• We conducted a comprehensive study and compared it with different types of
climate downscaling approaches, including statistical methods and other machine/
deep learning approaches to show that ours outperforms alternatives.
Prediction results and error (MAE)
Metrics: MAE, RMSE, Pearson Correlation, SSIM, and Forecast Indicators.