[Paper Review] Precipitaion Nowcasting using Deep Neural Network

Mohamed Chafik Bakkay, Mathieu Serrurier+9 more|arXiv (Cornell University)|Mar 24, 2022

Meteorological Phenomena and Simulations4 citations

TL;DR

This paper proposes a deep learning framework for high-resolution precipitation nowcasting using U-Net, ConvLSTM, and SVG-LP models trained on 2D radar precipitation maps. It introduces a patch extraction method preserving spatial context and a weighted, combined loss (WSSIM + WMSE) to reduce blurriness and minimize the impact of zero-precipitation pixels, achieving superior spatiotemporal forecasting accuracy, especially with U-Net and the hybrid loss function.

ABSTRACT

Precipitation nowcasting is of great importance for weather forecast users, for activities ranging from outdoor activities and sports competitions to airport traffic management. In contrast to long-term precipitation forecasts which are traditionally obtained from numerical models, precipitation nowcasting needs to be very fast. It is therefore more challenging to obtain because of this time constraint. Recently, many machine learning based methods had been proposed. We propose the use three popular deep learning models (U-net, ConvLSTM and SVG-LP) trained on two-dimensional precipitation maps for precipitation nowcasting. We proposed an algorithm for patch extraction to obtain high resolution precipitation maps. We proposed a loss function to solve the blurry image issue and to reduce the influence of zero value pixels in precipitation maps.

Motivation & Objective

• To address the challenge of fast, high-resolution precipitation nowcasting under strict time constraints, which traditional numerical models cannot meet due to high computational cost.
• To improve prediction accuracy by leveraging deep learning models that capture complex spatiotemporal patterns in radar-observed precipitation.
• To overcome resolution loss from standard resizing by introducing a patch extraction algorithm that preserves full-resolution spatial context with neighboring information.
• To mitigate the blurriness common in deep learning forecasts and the bias from zero-precipitation pixels by designing a weighted, hybrid loss function combining WSSIM and WMSE.

Proposed method

• The models are trained to perform regression on continuous precipitation values rather than classification, enabling output of real-valued precipitation rates.
• A patch extraction algorithm generates input patches of size 256×256 with a 200-pixel interpolated neighborhood around each target patch of size 128×128 to preserve local spatial context.
• A novel loss function combines structural similarity (SSIM) and weighted mean squared error (WMSE), with pixel-wise weighting to reduce the influence of zero-precipitation values.
• The loss function is designed to improve structural fidelity (via SSIM) and reduce blurriness, while the weighting mechanism prevents the model from being overly penalized by frequent zero values in radar maps.
• Three models—U-Net (CNN-based), ConvLSTM (RNN-based), and SVG-LP (stochastic RNN-based)—are evaluated for their ability to predict future precipitation sequences over 1–1.5 hours.

Experimental results

Research questions

• RQ1 Can a patch-based input strategy preserve high-resolution spatial information in precipitation nowcasting without GPU memory saturation?
• RQ2 Does a combined WSSIM and WMSE loss function improve forecast sharpness and reduce blurriness compared to standard MSE or SSIM alone?
• RQ3 How do different deep learning architectures—U-Net, ConvLSTM, and SVG-LP—compare in predicting spatiotemporal evolution of precipitation fields?
• RQ4 To what extent does weighting the loss function reduce the impact of zero-precipitation pixels on model training and prediction quality?

Key findings

• The U-Net model outperformed both ConvLSTM and SVG-LP in predicting high-intensity precipitation events and maintaining accurate rainfall contours.
• The WSSIM loss function generated more realistic high-precipitation regions (>5 mm/h) in correct spatial locations compared to WMSE, achieving the highest F1-score at the 1 mm/h threshold.
• The combination of WSSIM and WMSE losses (WSSIM+WMSE) yielded the best overall performance, balancing structural fidelity and pixel-level accuracy, as shown in both qualitative and quantitative evaluations.
• The patch extraction method with 200-pixel neighborhood information significantly improved MAE and F1-score over a baseline patch method with only 64-pixel margins, demonstrating the importance of contextual spatial input.
• ConvLSTM showed a tendency to blur later frames, while SVG-LP produced more false alarms and less precise predictions, indicating limitations in handling deterministic atmospheric dynamics.