[Paper Review] Precipitation nowcasting using a stochastic variational frame predictor with learned prior distribution

Alexander Bihlo|arXiv (Cornell University)|May 13, 2019

Precipitation Measurement and Analysis28 citations

TL;DR

This paper proposes a stochastic variational frame predictor with a learned prior distribution for precipitation nowcasting using radar reflectivity maps, improving long-term forecast sharpness over standard convolutional LSTMs. The method learns dynamic prior distributions from past frames to model uncertainty and physical evolution, yielding significantly sharper and more accurate 2.5-hour forecasts with lower structural similarity index (SSIM) degradation than baseline models.

ABSTRACT

We propose the use of a stochastic variational frame prediction deep neural network with a learned prior distribution trained on two-dimensional rain radar reflectivity maps for precipitation nowcasting with lead times of up to 2 1/2 hours. We present a comparison to a standard convolutional LSTM network and assess the evolution of the structural similarity index for both methods. Case studies are presented that illustrate that the novel methodology can yield meaningful forecasts without excessive blur for the time horizons of interest.

Motivation & Objective

• To address the ill-posed nature of short-term precipitation forecasting, where multiple future states are physically plausible.
• To improve long-term precipitation nowcasting beyond the limitations of deterministic convolutional LSTM models, particularly their tendency to blur future frames.
• To incorporate uncertainty in precipitation evolution by learning a frame-dependent prior distribution from historical radar data.
• To develop a deep learning model that is end-to-end trainable with minimal pre-processing, suitable for real-time operational use.
• To enable ensemble-like forecasting by generating multiple plausible future precipitation scenarios through stochastic sampling.

Proposed method

• The model uses an encoder–prediction–decoder architecture resembling a conditional variational autoencoder, with latent space inference via a recurrent inference network.
• A stochastic prediction head samples from a multivariate Gaussian prior distribution $ p_{\text{prior}}(\mathbf{z}_t|\mathbf{x}_{1:i-1}) $, which is learned from previous radar frames and encodes evolving physical dynamics.
• The prior distribution is trained jointly with the inference network using a variational lower bound objective that includes a Kullback–Leibler divergence term to align $ q_{\text{inf}}(\mathbf{z}_i|\mathbf{x}_{1:i}) $ with $ p_{\text{prior}}(\mathbf{z}_i|\mathbf{x}_{1:i-1}) $.
• The model is trained on sequences of 2D radar reflectivity maps using a reconstruction loss and a KL divergence regularizer to ensure meaningful latent representations.
• Prediction involves sampling from the learned prior and decoding to generate future radar frames, with multiple realizations used to assess uncertainty.
• The framework enables end-to-end training without extensive pre-processing, and supports probabilistic forecasting by sampling from the stochastic prior.

Experimental results

Research questions

• RQ1 Can a stochastic variational frame predictor with a learned prior improve long-term precipitation nowcasting accuracy compared to deterministic convolutional LSTMs?
• RQ2 Does the learned prior distribution effectively encode physical evolution rules of precipitation cells across time?
• RQ3 How does the model perform in terms of structural similarity (SSIM) over extended lead times (up to 2.5 hours)?
• RQ4 To what extent does the model reduce the blurring artifact common in deterministic video prediction models?
• RQ5 Can the model generate multiple plausible future precipitation scenarios that reflect uncertainty in storm development?

Key findings

• The stochastic variational frame predictor achieved significantly better structural similarity index (SSIM) scores than the standard convolutional LSTM across all forecast horizons beyond frame 6, with a much slower degradation rate.
• While the convolutional LSTM outperformed the proposed model on the first two forecast frames, its SSIM dropped sharply thereafter, indicating severe blurring in later frames.
• The proposed method produced sharper and more realistic precipitation cell evolutions, particularly for both advective and convective precipitation systems, as shown in case studies.
• The model’s ability to generate multiple realizations (averaged over 10 samples) demonstrated meaningful uncertainty quantification, unlike the deterministic LSTM which produces only one fixed forecast.
• The learned prior distribution effectively captured temporal dynamics of precipitation cells, enabling more physically plausible long-term forecasts.
• The model was found to be easier to train in practice than the standard convolutional LSTM, which required longer training cycles.