Mitigating Checkerboard Artifacts for Enhanced Precipitation Nowcasting: A Comparison of Upsampling Techniques

In the field of precipitation nowcasting, the application and advancement of deep learning techniques have enabled resource-efficient predictions. In particular, U-Net variants and attention-based architectures achieve computational reduction by extracting features with wide receptive fields through downsampling and upsampling processes. However, upsampling methods can induce checkerboard artifacts when spatially adjacent pixels in high-resolution feature maps are computed from different low-resolution pixels, resulting in overlooked dependencies compared to those derived from identical pixels. This leads to discrepancies with the ground truth patterns, ultimately degrading the performance of prediction models. This paper introduces upsampling techniques known to prevent checkerboard artifacts in the super-resolution domain into precipitation prediction models, aiming to improve performance while minimizing increases in model complexity. At the upsampling stage, we incorporate sub-pixel convolution or decouple the upsampling and channel reduction processes, comparing performance against models using transposed convolution, the standard upsampling approach in U-Net. Additionally, the Checkerboard Artifacts Score (CAS) is proposed to quantify the degree of checkerboard artifacts in images, which is applied to each model for analysis. CAS is defined as the ratio of errors between pixels forming artifact boundaries to errors between all adjacent pixels. In experiments, sub-pixel convolution and the combination of nearest neighbor or bilinear interpolation with subsequent convolution record lower CAS values than transposed convolution, while also demonstrating improved performance across metrics including NSE, CSI, and RMSE. Notably, sub-pixel convolution exhibits pronounced performance with balanced POD and FAR, while the bilinear approach generates spatially natural patterns with competitive performance. Analysis of the experimental results suggests that the reduction of checkerboard artifacts contributes to performance improvement. Furthermore, this work highlights the importance of upsampling method selection in video prediction tasks and provides practical guidance for model design.