PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepRain: ConvLSTM Network for Precipitation Prediction using Multichannel Radar Data

Seongchan Kim, Seungkyun Hong, Minsu Joh, Sa-kwang Song

arXiv:1711.02316v1cs.LG

TL;DR

Rainfall forecasting matters for daily life and businesses, while prior ConvLSTM precipitation work used only one radar channel. DeepRain applies ConvLSTM to three-dimensional, four-channel radar data, and its two-stacked model achieved a lower reported result than linear regression and FC-LSTM.

  • Problem

    Prior ConvLSTM precipitation experiments used three-dimensional, one-channel data, while this study addresses rainfall prediction from four-channel radar data.

    Source: §I · p. 1

  • Method

    DeepRain applies ConvLSTM to 15 time-ordered radar inputs shaped as 101 × 101 grids with four altitude channels to predict rainfall.

    Source: §I · p. 1, §IV · p. 3

  • Results

    The two-stacked ConvLSTM result was 11.31, 23.0% lower than linear regression and 21.8% lower than FC-LSTM.

    Source: §V · p. 4

  • Takeaways & Limitations

    The study reports better rainfall prediction accuracy than linear regression and FC-LSTM for its radar-data setting.

    Source: §VI · p. 4

Abstract

from arXiv · show

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.

I. INTRODUCTION

DeepRain applies ConvLSTM to three-dimensional, four-channel radar data to estimate rainfall. The introduction positions this design against prior one-channel ConvLSTM work and reports better prediction than linear regression and fully connected LSTM.

  • Deep learning studies have applied CNNs, RNNs, or combinations of the two to precipitation and weather-related prediction.Source: §I · p. 1
  • Earlier ConvLSTM precipitation experiments used three-dimensional, one-channel data, whereas DeepRain uses three-dimensional, four-channel radar data.The four channels correspond to radar observations from four altitudes.Source: §I · p. 1
  • The paper identifies ConvLSTM as an LSTM variant that embeds convolution inside its cell to model spatial data.Source: §I · p. 1
  • Weather radar images encode rainfall-rate information using colors that vary with the degree of radar-signal reflection.Source: §I · p. 1, Figure 1 · p. 2
  • DeepRain was reported to predict rainfall more effectively than linear regression and fully connected LSTM (FC-LSTM).Source: §I · p. 2

II. RELATED WORK

Prior work applied CNNs, RNNs, and their combinations to rainfall prediction, but differed in input data and forecasting task. DeepRain uses radar data with ConvLSTM for one-step rainfall forecasting.

  • Zhuang and Ding used a spatiotemporal CNN with 62 years of meteorological data to predict heavy-precipitation clusters, while Zhang et al. used multichannel radar data to classify heavy rain.Source: §II · p. 2
  • Gope et al. combined CNN and LSTM as separate steps using atmospheric variables, but did not use radar data.Source: §II · p. 2
  • Shi et al. used ConvLSTM to generate future radar maps from past maps in a many-to-many setup, whereas this study forecasts rainfall with a many-to-one approach.Source: §II · p. 2

III. DATA

The experiments use Shenzhen radar reflectivity observations paired with measured rainfall. The dataset contains 10,000 records sampled across two years and is divided into training, validation, and test sets.

  • The dataset contains 10,000 records randomly selected over two years, split into 90% training, 5% validation, and 5% test data.Source: §III · p. 2
  • Each record contains 101 × 101 reflectivity values at four altitude groups and 15 six-minute intervals, paired with measured rainfall for a central 50 × 50 km² target area.The rainfall measurement covers the period from 1 to 2 hours.Source: §III · p. 2
  • Radar observations were collected in Shenzhen and supplied by the Shenzhen Meteorological Administration; the reflectivity values are integer dBZ readings.Source: §III · p. 2

IV. METHOD

DeepRain uses ConvLSTM to process spatial radar sequences, replacing gate-level matrix multiplication with convolution. Fifteen time-ordered radar inputs produce one expected rainfall amount.

IV. METHOD

  • ConvLSTM replaces matrix multiplication with convolution at each LSTM gate, capturing spatial features in multidimensional radar data.This contrasts with FC-LSTM, whose one-dimensional input is described as unsuitable for spatial sequence data.Source: §IV · pp. 2–3
  • DeepRain takes 15 radar records, each reshaped as 101 × 101 × 4, and outputs the final cell’s expected rainfall amount.The four channels correspond to radar observations from four altitudes.Source: §IV · p. 3, Figure 2 · p. 3
  • The model uses a many-to-one configuration and does not predict the next sequence of labels.The rainfall label is a single amount associated with radar data, rather than a sequence of future labels.Source: §IV · p. 3

V. EXPERIMENT

The experiment evaluates DeepRain on test radar data using RMSE and compares it with linear regression and FC-LSTM. The two-stacked ConvLSTM achieves the reported lowest error among these comparisons, while learning-curve and validation experiments inform training choices.

V. EXPERIMENT

  • Training used Adam at learning rate 0.001, shuffled mini-batches of 30, and 50 epochs; one-stack ConvLSTM training took about 12 hours.RMSE was used to measure prediction accuracy.Source: §V · p. 3
  • Five epochs brought ConvLSTM to loss 10.0, while FC-LSTM with GDO needed 300 epochs to lower loss from 15.5 to 14.4.The learning curves compare FC-LSTM and ConvLSTM under different optimizer conditions.Source: Figure 3 · p. 3, §V · p. 3
  • Two-stacked ConvLSTM was more stable than one-stacked; validation loss increased from epoch 5, and testing used the epoch-5 model.The training and validation curves for the two-stacked model are shown in Figure 4.Source: §V · p. 3, Figure 4 · p. 4
  • RMSE 11.31 for two-stacked ConvLSTM was 23.0% lower than linear regression and 21.8% lower than FC-LSTM on the test set.Table I reports test-set RMSE; the reported comparison is also stated in the results prose.Source: §V · pp. 3–4, Table I · p. 4

VI. CONCLUSION

DeepRain applies ConvLSTM to three-dimensional, four-channel radar data to predict rainfall 1–2 h ahead. The authors report better accuracy than linear regression and FC-LSTM and outline several future research directions.

VI. CONCLUSION

  • DeepRain predicts rainfall 1–2 h ahead from three-dimensional, four-channel radar data, with better accuracy than linear regression and FC-LSTM.The authors describe this as their first application of ConvLSTM to this data setting.Source: §VI · p. 4
  • Future work will compare ConvLSTM with ConvGRU and expand the dataset through augmentation, including crops of a central 50 * 50km2 area.Source: §VI · p. 4
  • The authors also plan to develop convolution methods for spatial three-dimensional data with multiple variables and channels, and consider El Nino’s effect on precipitation in the studied area.Source: §VI · p. 4
Loading 1711.02316v1…