[Paper Review] DeepRain: ConvLSTM Network for Precipitation Prediction using Multichannel Radar Data

Seongchan Kim, Seungkyun Hong+2 more|arXiv (Cornell University)|Nov 7, 2017

Precipitation Measurement and Analysis1 references124 citations

TL;DR

DeepRain applies ConvLSTM to three-dimensional, four-channel radar data to predict rainfall; a two-stack ConvLSTM achieves RMSE of 11.31, outperforming linear regression by 23%.

ABSTRACT

Accurate rainfall forecasting is critical because it has a great impact on people's social and economic activities. Recent trends on various literatures show that Deep Learning (Neural Network) is a promising methodology to tackle many challenging tasks. In this study, we introduce a brand-new data-driven precipitation prediction model called DeepRain. This model predicts the amount of rainfall from weather radar data, which is three-dimensional and four-channel data, using convolutional LSTM (ConvLSTM). ConvLSTM is a variant of LSTM (Long Short-Term Memory) containing a convolution operation inside the LSTM cell. For the experiment, we used radar reflectivity data for a two-year period whose input is in a time series format in units of 6 min divided into 15 records. The output is the predicted rainfall information for the input data. Experimental results show that two-stacked ConvLSTM reduced RMSE by 23.0% compared to linear regression.

Motivation & Objective

• Motivate accurate rainfall forecasting due to social and economic impacts.
• Develop a data-driven model using radar data to predict rainfall amounts.
• Leverage ConvLSTM to capture spatiotemporal features in 3D, multi-channel radar data.
• Evaluate performance against baseline linear regression and FC-LSTM.

Proposed method

• Use ConvLSTM to handle 3D, four-channel radar input (101x101x4) across 15 time steps.
• Structure DeepRain as a many-to-one predictor to estimate rainfall 1–2 hours ahead.
• Replace FC-LSTM with convolution operations inside LSTM gates to capture spatial features.
• Train with Adam optimizer (lr=0.001) on preprocessed radar data converted to TFRecord format.
• Compare one-stack and two-stack ConvLSTM configurations against linear regression and FC-LSTM baselines.

Experimental results

Research questions

• RQ1 Can ConvLSTM on 3D, multi-channel radar data outperform traditional linear models and FC-LSTM in rainfall prediction?
• RQ2 Does stacking ConvLSTM layers improve predictive accuracy for radar-based rainfall nowcasting?
• RQ3 What is the impact of input dimensionality (three dimensions, four channels) on rainfall prediction performance?

Key findings

• Two-stack ConvLSTM achieves RMSE of 11.31 on the test set.
• Two-stack ConvLSTM reduces RMSE by 23.0% compared to linear regression.
• ConvLSTM (one-stack) achieves RMSE of 11.51, also outperforming FC-LSTM (14.46) and linear regression (14.69).
• ConvLSTM training converges faster than FC-LSTM under Adam optimization (epochs and loss curves reported).
• ConvLSTM better preserves spatial information than FC-LSTM due to convolutional gates.
• The model processes 15 time-step inputs of size 101x101x4 to produce a single rainfall prediction.