Sharpening Precipitation Nowcasts with Deep Learning: Architecture-Dependent Effects of Loss Functions

Precipitation nowcasting, or the prediction of rainfall intensity over short lead times from sequences of previous observations, is important for applications including early warning and weather-dependent decision-making. Deep learning models have become widely used for precipitation nowcasting; however, conventional pixel-wise losses such as mean squared error (MSE) can produce spatially smooth forecasts and may inadequately represent fine-scale precipitation structures. This study evaluates the use of Fourier Amplitude and Correlation Loss (FACL), a frequency-domain loss designed to account for differences in spectral amplitude and spatial structure, for Indian monsoon precipitation nowcasting. Two deep learning architectures, ConvLSTM and Spatiotemporal UNet, are trained using both MSE and FACL with the IMERG-Final Run precipitation product. Their forecasts are evaluated using pixel-wise, structural, perceptual, and spatial metrics, including MSE, SSIM, LPIPS, FSS, and RHD. The results show that the impact of FACL is strongly architecture-dependent. For ConvLSTM, changing from MSE to FACL produces only small changes in the evaluated metrics. In contrast, for Spatiotemporal UNet, FACL produces substantial improvements across all evaluated metrics, with SSIM increasing from approximately 0.1327 to 0.6453 and FSS increasing from 0.0059 to 0.6669, while RHD decreases from 1.082 to 0.3023. The corresponding forecasts also exhibit visibly sharper and more spatially structured precipitation features than those obtained with MSE. These results indicate that FACL can substantially reduce the smoothing observed in MSE-trained forecasts for Spatiotemporal UNet, while its effect on ConvLSTM is comparatively limited. The findings highlight the importance of evaluating loss functions jointly with the architecture used for precipitation nowcasting.