A SURVEY OF RAINFALL PREDICTION

USING DEEP LEARNING

2021 3rd International Conference on Electrical, Control and Instrumentation Engineering (ICECIE) | 978-1-6654-4966-3/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICECIE52348.2021.9664730

Jamal Hussain Chawngthu Zoremsanga

Department of Mathematics and Computer Science Department of Mathematics and Computer Science
Mizoram Univerisity Mizoram Univerisity
Aizawl, India Aizawl, India
jamal.mzu@gmail.com zoremsanga@gmail.com

Abstract— Prediction of rainfall is a difficult task because of on how the neural networks can predict various weather
the high volatility and complicated nature of the atmospheric phenomena such as temperature, tidal level, rainfall, and
data. Recently, various deep learning methods were successfully flood. A survey of satellite-based rainfall prediction technique
applied to forecast rainfall. We survey papers that employ deep was presented by [18]. [19] presented a study of statistical
learning techniques to predict rainfall using meteorological methods and data mining techniques for the prediction of
data. The papers are examined in terms of the deep learning rainfall. [20] provided a critical study of papers published
methods applied, location of the study area, types of metrics and from 2013 to 2017 for rainfall prediction based on data mining
software used for implementing the model and, year-wise techniques. Recently, there is an increasing interest in the use
publication of the papers. From the surveyed papers, we found
of deep learning for rainfall prediction, however there is a lack
that deep learning methods can be applied successfully for
rainfall prediction and they are found to be superior than the
of survey papers that focused on rainfall prediction using deep
traditional machine learning methods and shallow neural learning methods. Thus, the purpose of this paper is to fill this
network models. We also provide future directions for research gap and study the papers that used deep learning methods for
in the area of rainfall prediction. the prediction of rainfall. We classify the papers based on the
nature of data used to train and test the deep learning models.
Keywords— Rainfall prediction, Artificial neural network, We study the types of metrics and software used, yearly trends
Machine Learning, Deep Learning of the papers published, the location of the case study, and the
type of deep learning methods applied for the prediction.
I. INTRODUCTION Furthermore, we also discuss the performance of deep
Accurate rainfall prediction is necessary because of its learning models compared to machine learning models and the
impact in the fields of agriculture, transportation, water potential future research directions.
supply, renewable energy management and, various activities This paper is organized as follows. Section II discusses the
of human beings. As rainfall prediction depends on multiple methodology for conducting the survey. In section III, we
environmental factors, it is a challenging task [1]. Rainfall summarized the deep learning methods applied for the
prediction methods can be categorized into physical methods, prediction of rainfall. Section IV presents the result and
statistical methods, and machine learning techniques. Physical discussion and in section V we mention the future directions.
methods are models that are implemented using numerical Finally, the conclusion is given in section VI.
weather prediction. Physical methods have a drawback
because they require large computational resources and large II. METHODOLOGY
data requirements for calibrating the model. On the other
The process of this survey consists of the following steps:
hand, the statistical model aims to uncover the mathematical
(a) Collection of papers with a focus on the use of deep
relationship between online time-series problems.
learning methods for rainfall prediction (b) detailed survey
Autoregressive Integrated Moving Average (ARIMA),
and analysis of the collected papers.
Multiple Regression and, Linear Regression (LR) are
commonly used in statistical modelling. For many years In the first step, well-known digital libraries such as
researchers have applied machine learning models for ScienceDirect, IEEE Xplore, Springer, and Google Scholar
prediction of rainfall [1]–[5]. Support Vector Machine (SVM) were searched for Journal articles and Conference paper using
and neural network model such as Artificial Neural Network the combination of keywords given below: -
(ANN) are commonly used machine learning models for
rainfall prediction. A type of Machine learning model called [“deep learning” OR “machine learning”] AND [“rainfall”
Deep learning model implement deep structure in its OR “precipitation] AND [“forecasting” OR “prediction”]
architecture. It is a composition of several processing layers to Using the search criteria given above, 246 papers were
learn the data representations using multiple level of collected. The collected papers were screened and 45 papers
abstraction [6]. Deep learning methods are employed were selected for the detailed study after applying the
successfully in the area of forecasting, classification problem, following inclusion criteria: -
image and natural language processing, speech recognition,
object detection, etc. [7]–[14]. 1) Papers published from January 2015 to June 2020
2) Papers from peer-reviewed journals and conferences.
Rainfall forecasting has been an active area in literature. 3) Papers that predict rainfall
Several survey papers that applied machine learning models
4) Papers that used deep learning methods
in forecasting rainfall and other weather parameters have been
published. [15] presented an overview of various
computational intelligence tools for weather prediction aiming

978-1-6654-4966-3/21/$31.00 ©2021 IEEE

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 In the second step, 45 papers selected from the above steps The proposed SAE-SVM and Stacked Autoencoder Neural
were studied in detail. The following research questions were Network (SAE-NN) model was compared with Stacked
considered while surveying the selected papers: Autoencoder - Anomaly Frequency Method (SAE-AFM),
Principal Component Analysis- Support Vector Machine
1) Which deep learning methods were used for rainfall (PCA-SVM), and Fisher Linear Discriminant Analysis
predictions? (LDA). Results showed that the proposed model is better for
2) Which metrics were used for evaluating the heavy rainfall prediction.
performance of the method?
[26] proposed Deep Belief Network for Precipitation
3) What are the sources of datasets used in the papers?
Forecast (DBNPF) for short-term rainfall forecasting. The
4) What parameters are used for training? authors compared the DBNPF with Radial Basis Function
5) Does deep learning provide better prediction (RBF) neural network, SVM, ARIMA, Extreme Learning
compared to other methods? Machine (ELM), and SAE models. [27] also implemented
DBNPF model for predicting rainfall in four areas of China
III. APPLICATIONS OF DEEP LEARNING IN RAINFALL such as Zunyi of Guizhou Province, Hezuo of Gansu
PREDICTION
Province, Jinan of Shandong Province, and Changchun of Jilin
In this paper, we categorized the rainfall forecasting task Province. In both the papers, the research results suggested
based on the type of data used for training and evaluation of that the DBNPF model is better than the other models.
the models. We classify the type of data as weather parameter
data, radar image data, and satellite image data. Out of 45 Echo State Networks (ESN) and Multi-Gene Genetic
papers surveyed, 26 papers used weather parameters, 13 Programming (MGGP) were proposed by (Ouyang and Lu,
papers used radar images, and 6 papers used satellite images 2017) to forecast monthly rainfall. The proposed models were
for training and testing the models. compared with the Support Vector Regression (SVR) method
for 1-, 3- and 6-months lead time. The authors also compared
A. Rainfall prediction using weather parameters the performance of Wavelet Transform (WT), Singular
In this section, we present the papers that used weather Spectrum Analysis (SSA), and Ensemble Empirical Mode
parameters collected using Meteorological Observation Decomposition (EEMD) for pre-processing of data. Among
Stations such as surface observation, upper air observation, the forecasting models, ESN outperformed SVR and MGGP,
and ocean observation. Table I summarized the papers that and SVR was better than MGGP. However, all three models
used weather parameters for the prediction of rainfall. The were found to be suitable for monthly rainfall forecasting.
deep learning methods used, compared methods, framework Among the data pre-processing methods, WT was a good
used, country, temporal resolution, and metrics for evaluating technique for forecasting short-term rainfall whereas, SSA
the models were also given in the table. performed better for forecasting of long-term rainfall. EEMD
showed the poorest performance compared to WT and SSA.
(Saha et al., 2016, 2017, 2021) identified the predictors for The best-performing model (SSA-ESN) can forecast rainfall
early-late Indian summer monsoon, monsoon in the up to 2 years lead time with acceptable accuracy.
homogeneous regions of India, and aggregate Indian summer
monsoon using a deep learning method called stacked [29] predicted short-term rainfall using the features
autoencoder. The identified predictors are used for the long- collected from a network of rain gauges. Deep Convolutional
term forecast of monsoon using machine learning models such Neural Network (CNN) was used for extracting features from
as Regression Tree with Bagging algorithm (RegTreeB) and the input data and a fully connected layer for the prediction
Decision Tree with Bagging algorithm (DecTreeB). The task. The authors compared the proposed multi-task CNN
proposed prediction models are compared with Indian method with Quantitative Precipitation Forecast (QPF), LR,
Meteorological Department (IMD) monsoon prediction MLP, AE-MLP, Long Short-Term Memory (LSTM), CNN,
models. Results of the study showed that the proposed Multi-task MLP, and Multi-task Recurrent Neural Network
prediction models with the identified monsoon predictors (RNN). Experimental results showed that the proposed model
outperformed the compared IMD models. significantly outperformed the compared models.

[24] used a deep learning technique called Denoising [30] applied deep learning model called Deep Belief
Autoencoder (DAE) and a Multilayer Perceptron (MLP) for Network (DBN) for precipitation forecasting. DBN was
predicting the next day's rainfall. The autoencoder extract compared with SVM and SVM based on different
non-linear features from the input meteorological data and the optimization algorithms such as Particle Swarm Optimization
MLP network is used for classification and prediction. The (PSO). The time consumed by DBN was lower than the SVM
authors compared the proposed model with MLP, naive methods, and it was founded that SVM methods can be used
approach, Back Propagation network (BP), Layer Recurrent for small datasets whereas the DBN method can be used for
Network (LRN), Cascaded Back-Propagation (CBP), large-scale datasets.
Ensemble Empirical Mode Decomposition (EEMD), and Very Short-Term precipitation forecasting was studied by
Feed-forward Neural Networks (FNN). The result showed that [31] and compared the performance of Deep Neural Network
the proposed method achieved lower Mean Squared Error (DNN) with SVM, XGBoost (XGB), Random Forest (RF),
(MSE) and Root Mean Squared Error (RMSE) than the and Random Prediction (RP) methods. Among the compared
compared methods. methods, DNN yields the highest accuracy for rain prediction.
[25] predicted heavy rainfalls during the monsoon season [32] investigated the capability of DNN to predict monthly
(June, July, August, and September). Due to a huge set of rainfall. After testing different hidden layers and hidden
features, Stacked Autoencoder (SAE) was used as a tool for nodes, the authors considered 5 hidden layers with 128 nodes
feature reduction. The reduced features are then used for the in each layer for the DNN model. Based on the results, the
classification of heavy rainfalls using a cost-sensitive SVM. DNN model was found to be appropriate for forecasting

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
monthly rainfall with a one-month lead time. The accuracy of (ARIMA), Autoencoder Multilayer Perceptron (AE-MLP),
the model was decreased when increasing the lead time. Multitasking Convolutional Neural Network (MT-CNN),
Multitasking Gated Recurrent Unit (MT-GRU). For a single
[33] compared the performance of LSTM with MLP and time-step, the proposed model provides a lower Root Mean
Seasonal Neural Networks (SNN) for predicting rainfall. The
Squared Error (RMSE) than the compared models. In the case
result of the experiment suggested that LSTM was better in of a multi-step model with a rolling mechanism to forecast the
terms of performance than the compared methods. The next 6-time steps, the proposed model is found to be useful but
authors stated that LSTM can be a promising model for
limit accuracy.
estimating precipitation.
[39] proposed Intensified LSTM for the prediction of
[34] applied LSTM for forecasting monthly rainfall and rainfall. To solve the issue of the LSTM vanishing gradient
explores the selection of optimal time lag for the model. The problem, the authors modified the LSTM by multiplying the
LSTM model was compared with RNN and it was tested on input value with the sigmoid function in the input gate and the
various homogeneous regions of India. The study observed tanh function in the candidate vector. This reduced the training
that the LSTM model outperformed the RNN model for time of the network which in turn caused a high learning rate
different fitness measures and 12 to 15 antecedent rainfall and reduced the losses. The proposed Intensified LSTM
events provide more valuable information. model was compared with Holt-Winters, Extreme Learning
[35] developed a deep neural network composed of Machine (ELM), ARIMA, Recurrent Neural Network (RNN
convolutional layers and an LSTM network for improving with Rectified Linear Unit (ReLU), RNN with Sigmoid Linear
Monsoon precipitation prediction. The convolutional layers Unit (SiLU), and LSTM models. Experimental results showed
were used to extract spatial features of the raw input data, that the accuracy of the proposed model outperformed the
which was then fed to the LSTM networks. The effectiveness compared models.
of the predictors was studied and the geopotential height was [40] developed two LSTM based models, Wavelet Long
found to be the most important predictor. The proposed
Short-Term Memory (WLSTM) and CLSTM to forecast
Convolutional Long Short-Term Memory (ConvLSTM) streamflow and rainfall. WLSTM is a hybrid model composed
model was compared with Quantile Mapping (QM) method,
of LSTM and wavelet transform and CLSTM is composed of
SVM, and CNN and the precipitation estimate given by the CNN and LSTM. WLSTM and CLSTM models were
ConvLSTM network was found to have the highest
compared with three layers MLP and LSTM. Results showed
performance. that WLSTM and CLSTM outperformed LSTM and MLP for
[36] proposed the identification of break and active both stream fall and rainfall forecasting. The forecasting
monsoon spells for the central region of the Indian accuracy of LSTM was improved using the wavelet transform
subcontinent using LSTM and Sequence-to- and convolutional layers.
Sequence (Seq2Seq) models. The Seq2Seq model consists of [41] adopted a deep learning model called LSTM and K-
two LSTM units, a dense soft-max layer, and an attention Means clustering method. The data samples were first divided
mechanism. The authors classified each day as dry, wet, or into four categories using the K-means clustering method
normal day. Then the collection of classification at daily scale followed by building the models using LSTM for the different
was used to detect the break or active monsoon spells. Daily
data types. The proposed model was compared with
rainfall from June to September (1948 - 2014) was considered Frequency matching, Linear regression, SVM, and DBN
for detection of monsoon spells. The proposed models were
using RMSE and Threat Score (TS). The proposed model was
compared with SVM and K-Nearest Neighbor (KNN) and it found to reduce the RMSE effectively and it improved the TS
was observed that both LSTM and Seq2Seq performed better of light and heavy rain.
than SVM and KNN. Additionally, the Seq2Seq model was
found to be superior to LSTM for detecting monsoon spells. [42] constructed and trained a deep CNN model for severe
convective weather such as heavy rain, hail, convective gusts,
[37] used Echo State Networks (ESN) and DeepESN to and thunderstorms. The authors constructed two databases
predict rainfall using meteorological data. The study was based on Severe Convective Weather (SCW) observations and
conducted in the area of Southern Taiwan. The authors NCEP final (FNL) analysis data. The proposed CNN model
compared the performance of the proposed models with Back
was compared with Logistic Regression (Logit Reg), Random
Propagation Network (BPN) and SVR. To find the most Forest (RF), Support Vector Machine (SVM), and Multilayer
important parameter to predict rainfall, the authors used the Perceptron (MLP). Results showed that the deep CNN model
Principal Component Analysis (PCA) method. Rainfall, outperformed the compared traditional machine learning
pressure, and humidity are found to be the most important
algorithms in SCW forecasting over China.
parameters. The experimental result showed that the
correlation coefficient for the ESN and DeepESN is greater ConvLSTM tuned using the Salp-Stochastic Gradient
than BPN and SVR model. It was also found that DeepESN Descent (S-SGD) algorithm was proposed by [43] for the
was more accurate and has the best performance than the prediction of rainfall in India. S-SGD algorithm is a hybrid of
compared models. Salp Swarm Algorithm (SSA) and Stochastic Gradient
Descent (SGD) algorithm and they are used to select the
[38] proposed a cascading deep learning method to
optimal weights of the ConvLSTM model. The authors
classify rain/no-rain and to predict the amount of rainfall. implemented the MapReduced framework to deal with a large
CNN was used as a classification model to classify rain or no- amount of data in parallel. The performance of the model was
rain and using the classified rain class Gated Recurrent Unit compared with ConvLSTM, Clusterwise linear regression
(GRU) was used to predict the amount of rainfall. Focal loss (CLR), MLP, and Dynamic Self-organizing Multilayer
with sigmoid activation was used by the authors to prevent Network Inspired by the Immune Algorithm (DSMIA). It is
bias to non-rain class. The proposed cascaded model was found that the proposed model S-SGD based ConvLSTM
compared with Autoregressive Integrated Moving Average

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. SUMMARY OF PAPERS WHICH PREDICT RAINFALL USING WEATHER PARAMETERS

Author Methods Comparison Framework Country Resolution Metrics

[25] SAE-SVM, SAE-NN SAE-AFM, PCA-SVM, - India Daily, 6-hours -
Fisher LDA
[24] DAE, MLP MLP, Naïve, BPN, LRN, Theano Colombia Daily MSE, RMSE
CBP, EEMD, FFNN
[66] SAE, RegTreeB IMD Models - India Monthly -
[21] Sparse AE, RegTreeB, IMD Models - India Monthly MAE
DecTreeB
[28] ESN, MGGP SVR MATLAB China Monthly RMSE, MAE, NSE
[29] MT-CNN LR, MLP, AE-MLP, LSTM, PAI Colombia Daily MSE, CSI, CORR
CNN, China
MT-MLP, MT-RNN, ECMWF
[22] SAE, RegTreeB IMD Models MATLAB India Monthly MAE
[26] DBNPF RBF, SVM, ARIMA, ELM, MATLAB China Daily MAE, RMSE
Sparse AE
[33] LSTM MLP, SNN MATLAB, Vietnam Monthly CORR, R, RMSE, MAE
Python
[30] DBN SVM, PSO-SVM Theano China 3 hours -
[32] DNN - Tensorflow Thailand Monthly SEF
[31] DNN SVM, XGB, RF, RP Chainer Japan 10 minutes recall, F-score
[27] DBNPF SVM, RBF, ARIMA, ELM MATLAB China MAE, RMSE
[34] LSTM RNN Keras India Monthly RMSE, CORR, NSE, MAE
[67] CNN-GRU ARIMA, AE-MLP, CNN, Thailand Hourly F1 score,RMSE
MT-GRU,CNN-GRU
[43] S-SGD-based convLSTM, CLR, MLP, DSMIA MapReduce India Monthly, RMSE, NSE, R2
convLSTM framework Quarterly,
MATLAB Yearly
[35] ConvLSTM QM, SVM, CNN Tensorflow China Daily NSE, RB
[40] WLSTM, CLSTM MLP, LSTM China Monthly RMSE, Accuracy
[39] Intensified LSTM Holt–Winters,ELM, ARIMA Keras India RMSE, TS
RNN-Relu, RNN-Silu, LSTM
[36] LSTM, SVM, K-NN India Daily Precision, Recall,
Seq2Seq model Accuracy, F1-score
[37] ESN, DeepESN BPN, SVR, ECMWF MATLAB Taiwan Hourly RMSE, NRMSE, TS
CORR, POD, FAR,
[41] LSTM LR, SVM, DBN Python China MSE, PRD
[42] CNN Logit Reg,RF, SVM, MLP China 6 hours TS,ETS, POD,FAR

[45] MIMO‐LSTM, SR, SVR, RF, ARIMA, VAR, Keras 3 hours TS, ETS, POD, FAR
MISO‐LSTM, VECM, AFE, WRF-NWP
MIMO‐TCN,
MISO‐TCN,
[44] Stacked LSTM ANN, GRU - Australia 30 secs, RMSE, MAE, R2,CV,Bias
15 mins

[23] SAE, RegTreeB, IMD Models - India MAE

DecTree

outperformed the compared models in terms of Percentage GRU and much better than ANN in terms of RMSE, Mean
Root mean square Difference (PRD) and Mean Squared Error Absolute Error (MAE), Coefficient of Determination (R2),
(MSE). and Coefficient of Variation (VC).
[44] designed a deep learning model based on a two-layer [45] proposed LSTM and Temporal Convolutional
LSTM model and trained the model using a disdrometer- Networks (TCN) for short-term forecasting of rainfall using
derived dataset. The model was then applied to improve 10 surface weather parameters. They compared the
rainfall estimation using Commercial Microwave Link (CML) performance of the models with Standard Regression (SR),
data. Based on the attenuation of the transmitted SVR, RF, ARIMA, Vector Auto Regression (VAR), Vector
electromagnetic signal as it passed through the rain, rainfall Error Correction Model (VECM), and Arbitrage of
was estimated using CML networks. The authors compared Forecasting Expert (AFE) models. Two regression models
the LSTM model with GRU and ANN. The results showed namely Multi-Input Multi-Output (MIMO) and Multi-Input
that GRU performed better than LSTM in terms of relative Single-Output (MISO) were proposed to evaluate the
bias, whereas the LSTM model performed slightly better than proposed models. MIMO-LSTM and MISO-LSTM

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
outperformed the compared models and they were selected as To enable researchers to develop, test and deploy new
the proposed model. Since MISO-LSTM and MIMO-LSTM models significantly faster [51] studied a distributed learning
do not produce much difference and due to easily handle and approach to train a precipitation nowcasting model. In this
less time and power consumption MIMO-LSTM was selected study, a data-parallel model was implemented in which a CNN
for further analysis. Then, the author compared the MIMO- model and the training batches were replicated across multiple
LSTM with Weather Research and Forecasting - Numerical compute nodes. The CNN model was Fully Convolutional
Weather Prediction (WRF-NWP) model and found that without dense layers. The authors implemented the model
MIMO-LSTM provides better prediction up to 12 hours. using TensorFlow/Keras and Horovod framework and the
model was trained using up to 128 GPUs. Using the proposed
B. Rainfall prediction using Radar image approach, the training time for a given nowcasting model
To increase the spatial coverage and resolution of the data, architecture was reduced from 59 hours to just about 1 hour.
radar images are widely used by researchers to predict rainfall. The result also showed that the validation loss reduced
In Table II we summarized the papers that used radar images smoothly up to 24 GPUs. A noisy behavior was detected in
for the prediction of rainfall. [46] proposed convolutional the validation loss after increasing the number of GPUs
LSTM (ConvLSTM) for short-term rainfall prediction. The beyond 24 which could be due to a significant reduction in the
authors extended the Fully Connected LSTM (FC-LSTM) by training images available for each device.
incorporating convolutional structures in the input-to-state
and state-to-state transitions. ConvLSTM is compared against [52] proposed a Generative Adversarial ConvGRU (GA-
FC-LSTM and the Real-time Optical flow by Variational ConvGRU) model which is a composition of two adversarial
methods for Echoes of Radar (ROVER). The FC-LSTM does learning systems, ConvGRU-based generator, and a
not perform very well due to the spatial correlation in the radar convolutional neural-network-based discriminator. The
data. ROVER is found to give a sharper prediction, but it authors utilized a sequence of five radar echo images and
triggers more False alarms and is found to be less precise than predict ten radar echo maps. Results of the experiments
ConvLSTM. The authors concluded that ConvLSTM is better showed that GA-ConvGRU outperformed ConvGRU and
in capturing the spatiotemporal correlations and it also optical flow methods.
provides better predictions than the ROVER algorithm. [53] compared the performance of U-Net CNN with the
[47] predicted short-term precipitation using Optical flow model, persistence model, and NOAA's
Convolutional Long Short-Term Memory (convLSTM) using numerical one-hour High-Resolution Rapid Refresh (HRRR)
the radar data. Hyper-parameter search was performed using for short-term precipitation prediction. The authors treated the
Spearmint to select the convolutional kernel size, the number forecasting problem as an image-to-image translation problem
of convolutional filters, learning rate, and momentum. The where n sequence of radar images was used as input to the
results show that the standard encoder-decoder method is model. The study showed that the proposed model
more successful in Probability of Detection (POD) and outperformed the compared models.
Critical Success Index (CSI) and only a slight increase in false [54] proposed sequence-to-sequence model called dec-
alarm over the attention model. seq2seq model. The dec-seq2seq model consists of dec-
[48] proposed a Trajectory Gated Recurrent Unit TrajGRU, dec-ConvGRU, and dec-ConvLSTM. The dec-
(TrajGRU) for short-term rainfall prediction and compared its seq2seq models were compared with TrajGRU, ConvGRU,
performance with Convolutional Gated Recurrent Unit and ConvLSTM and showed improvements over the
(ConvGRU), Dynamic Filter Network (DFN), 2D and 3D compared models. Among the dec-seq2seq models, dec-
Convolutional Neural Networks (CNNs), and two optical TrajGRU performed better than the other models. To resolve
flow-based models (ROVER and its nonlinear variant). Due the blurry image issue due to the impact of the loss functions
to high imbalance in the proportions of rainfall events at such as MAE or MSE, an image quality assessment metrics
different rain-rate, the authors also proposed Balanced Mean Structural Similarity (SSIM) and Multi-Scale Structural
Square Error (B-MSE) and Balanced Mean Absolute Error (B- Similarity (MS-SSIM) were proposed by the authors. The
MAE) for training and evaluation of the models. The experimental result showed that the best loss function is
experiments show that TrajGRU outperforms the compared combination of SSIM, MSE, and MAE and the dec-seq2seq
models and training of the models using the balanced loss models can tolerate high and increasing uncertainty.
function performs better than training without balanced loss. To improve the accuracy of Doppler radar detection of
[49] explored the prediction of rainfall using a radar echo short-term rainfall prediction, Tiny-RainNet was proposed by
dataset by incorporating convolution operations within the [55]. Tiny-RainNet consists of a combination of Bi-directional
vanilla recurrent neural network. The proposed Conv-RNN Long Short-Term Memory (BiLSTM) and CNN to extract the
model was compared against Conv-LSTM and Eulerian temporal and spatial information. The authors compared the
Persistence models. The authors states that Conv-RNN can be proposed model with ConvLSTM, LSTM, FC-LSTM, and
used for learning the features of the Doppler weather radar AlexNet and it was found that Tiny-RainNet had a better
phenomenon and lesser parameters are used compared to other performance than the compared models.
hybrid approaches. Convolutions in recurrence also encode [56] developed RainNet, a deep convolutional neural
the Spatio-temporal correlations. network for radar-based short-term precipitation forecasting.
DeepRain- a ConvLSTM model was proposed by [50] to RainNet model consists of a stacked of CNN following a
predict the amount of rainfall using radar observation. The standard encoder-decoder structure with skip connection
prediction accuracy of the proposed method is found to be between its branches. Initially, RainNet predicted the
better than Linear Regression and FC-LSTM models. The precipitation for a lead time of 5 minutes and to predict a larger
result also showed that the two-stacked ConvLSTM lead time up to 60 minutes. RainNet was applied recursively
performed more stable than the one-stacked ConvLSTM. by using the previous output as the next input. The

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. SUMMARY OF PAPERS WHICH PREDICT RAINFALL USING RADAR IMAGE

Author Methodology Comparison Framework Country Spatial Temporal Metrics

Resolution Resolution
[46] ConvLSTM FC-LSTM, ROVER Theano - - - MSE, CSI, FAR,
POD, CORR
[47] ConvLSTM Attention CNTK USA 2 x 2 km - POD, FAR, CSI
[50] ConvLSTM LR, FC-LSTM TensorFlow China 101 X 101 km 6 minutes RMSE
[49] ConvRNN, ConvLSTM, USA 100 × 100 km Precision, Recall,
Multi-layer Eulerian persistence F1 score
ConvRNN

[48] TrajGRU ConvGRU, 2D and 3D CNN, - China - - B-MSE, B-MAE,

ROVER (nonlinear variant) CSI, HSS
[53] U-Net CNN MRMS persistence, HRRR Tensorflow USA 1 × 1km 2 minutes Precision, Recall
Optical flow method,
[51] CNN Keras, Horovod USA 256 x 256 km MSE
[52] GA-ConvGRU Optical flow method, - China 900 × 900 km POD, FAR,
ConvGRU CSI, HSS
[54] dec-ConvLSTM, TrajGRU, ConvGRU, TensorFlow China 101 × 101 km 6 minutes CSI, FAR, POD,
dec-ConvGRU, ConvLSTM MSE, MAE, SSIM,
dec-TrajGRU MS-SSIM, PCC

[55] CNN-BiLSTM ConvLSTM, LSTM, - China 10×10 Km 6 minutes RMSE

FC-LSTM, AlexNet
[56] RainNet (deepCNN) CNN Keras Germany 1 km × 1 km 5 minutes MAE, CSI, FSS
[57] ConvLSTM COTREC,ConvLSTM (with PyTorch China 1 km x 1 km 6 minutes CSI
cross entropy loss)
[58] MAR- CNN dual-channel CNN attention, Tensorflow China 41 × 41 km 6 minutes RMSE, EVS
dual-channel CNN,
single-channel CNN,
GBDT, SVM

experimental result showed that RainNet significantly satellite observation provides coverage over a large area and
outperformed the benchmark models Rainymotion and at regular intervals [59]. There is multiple satellites launch for
persistence method at all lead times up to 60 minutes. observation of meteorological phenomena. For decades,
researchers have used the data provided by satellites for
ConvLSTM with a star-shaped bridge architecture was predicting rainfall and other meteorological phenomena. In
implemented by [57] for precipitation nowcasting and this section, we present the literature that implements deep
compared the performance of the model with Continuous learning methods for rainfall prediction using satellite images.
Tracking Radar Echo by Correlation (COTREC) and the
Table III summarizes the papers found in the literature.
ConvLSTM with cross-entropy loss. The authors used Group
Normalization to refine the convergence performance in [60] applied a Stacked Denoising Auto-Encoder (SDAE)
optimization for ConvLSTM and they employed a special for bias correction on satellite precipitation product. The
multisigmoid loss. Experimental results showed that the SDAE is used to improve the Precipitation Estimation from
proposed model achieved state-of-the-art performance. Remotely Sensed Imagery using an Artificial Neural Network
Cloud Classification System (PERSIANN-CCS). The model
[58] proposed Multihead Attention Residual was evaluated including the detection of Rain or No-Rain
Convolutional Neural Network (MAR-CNN) for short-term pixels and the detection of amount of rainfall for both warm
precipitation forecasting. The proposed method used two and cold seasons. The study shows that the proposed model
CNN architectures. The first CNN model extracts deep can detect false alarm pixels in the PERSIAN-CCS and it is
characteristics from radar images and the second CNN model also able to rectify the bias of the overall precipitation level in
acquired deep features from the non-image input. Multihead
the warm and cold seasons.
attention was also introduced by the authors to emphasize the
key areas corresponding to precipitation and a residual [61] forecasted short-term precipitation using Cloud-Top
connection to avoid global information loss which was caused Brightness Temperature (CTBT). To forecast the next value
by the attention layer. The authors compared the MAR-CNN of CTBT image, deep learning algorithm LSTM was proposed
model with dual-channel convolutional attention model, dual- and the Precipitation Estimation using Remotely Sensed
channel convolutional model, single-channel CNN model, Information using Artificial Neural Networks (PERSIANN)
Gradient Boosted Decision Tree (GBDT), and SVM. The algorithm was used for estimating precipitation from the
proposed model was found to have a better prediction forecasted CTBT image. The model was compared with RNN
performance than the compared models. with PERSIANN, the Persistency method with PERSIANN,
C. Rainfall prediction using Satellite image
The distribution of rain gauges and radar systems is
common but limited to their spatial coverage. In contrast,

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 TABLE III. SUMMARY OF PAPERS WHICH PREDICT RAINFALL USING SATELLITE IMAGE

Author Methods Comparison Country Spatial Temporal Lead time Metrics

Resolution Resolution
[60] SDAE PERSIANN-CCS, USA 0.08° Hourly - POD, FAR, HSS, Bias
Stage IV MSE, Variance
[61] LSTM- RNN-PERSIANN, USA GOES-IR: GOES-IR: 6 hours CORR, RMSE, POD,
PERSIANN Persistency- 0.04°×0.04°, 30 minutes FAR, CSI
PERSIANN, Q2 dataset: Q2 dataset:
Farneback optical flow- 0.01°×0.01°, 5 minutes
PERSIANN, CTBT data and Q2
Rapid Refresh dataset: 0.25°×0.25°
[63] PredNet TrajGRU Japan 0.01° 5 minutes 10 frames CSI, HSS
network, (50 minutes)
ConvGRU
[62] ConvLSTM BMA, MSMES Brazil 0.05° Daily 1 day MAE, RMSE
[65] ConvLSTM LSTM - 0.1° 30 minutes 150 minutes RMSE, Accuracy
Bias, ETS, FAR, HSS,
ORSS, PFD, TS, RMSE,
Success Ratio, CORR,
Multiplicative Bias,
[64] CNN-LSTM CNN, LSTM, MLP China 0.05° Daily 1 Day RMSE, RB, MAE,
CORR

Farneback optical flow with PERSIANN algorithm, and used to forecast precipitation up to 150 minutes. The proposed
Rapid Refresh (RAPv1.0). The proposed model shows model was compared with LSTM and four optical flow-based
superiority in short-term precipitation forecasting. methods- Sparse Single Delta (SparseSD), Sparse, Dense, and
Dense Rotation (DenseROT). Based on the authors
[62] proposed rainfall prediction using an ensemble experiment, LSTM was not suitable for data with Spatio-
approach based on a deep neural network. A Convolutional temporal information and Convcast outperformed the
LSTM was compared with Bayesian Model Averaging compared models.
(BMA) and Master Super Model Ensemble System (MSMES)
methods for predicting rainfall. The experimental results IV. RESULTS AND DISCUSSION
showed that the proposed model is 50% more precise than the A. Temporal distribution of studies
compared models.
In Fig. 1, we plot the temporal distribution of the 45 papers
[63] proposed precipitation nowcasting model based on selected for the study. The plotted graph shows an increasing
PredNet network architecture and compared with TrajGRU trend in the number of papers published in the area of rainfall
method. ConvGRU was used as the unit of PredNet instead of prediction using deep learning. In 2019, the number of papers
the ConvLSTM unit. The experiment showed that the grew significantly, which is 51% of the total papers. This
proposed model achieved state-of-the-art performance in the shows that there is an increasing interest in applying deep
MovingMNIST++ dataset and an acceptable result in the real learning methods for rainfall prediction. The decreased in the
precipitation data. The model also consumes less GPU number of papers from 2019 to 2020 is because we consider
memory compared to the TrajGRU model. up to the month of June only.
[64] proposed to merge the Tropical Rainfall Measuring B. Deep learning methods used for rainfall prediction
Mission (TRMM) 3B42 V7 satellite image, rain gauge output, In Fig. 2, the frequency of deep learning methods used by
and thermal infrared images to enhanced the accuracy of the authors for rainfall prediction is given. From this figure, it
quantitative precipitation estimation (QPE). The authors used can be seen that LSTM (10 papers) and ConvLSTM (9 papers)
a combination of CNN and LSTM models to extract the spatial are the most frequently used methods. AE and SAE are
characteristics and time dependence of the merged dataset and usually used by the authors to discover the predictor variables
compared the accuracy with CNN, LSTM, and MLP models. when the input variables are large. When the input is a satellite
The CNN-LSTM model considered the time and space image or a radar image, most of the authors used convolutional
dependence of precipitation, thus outperformed the compared layers for learning the features of the input. The convolution
models which considered either spatial information only or operators enable to learn the spatial information in addition to
temporal information only. The authors also showed that
under different precipitation intensities the CNN-LSTM 20
model can correct and improved the TRMM data.
Number of paper

15
A precipitation nowcasting model called Convcast was
published

proposed by [65] to predict short-term precipitation. The 10

authors used Convcast a stack of three ConvLSTM layers for 5
learning the spatial and temporal features and a 3D 0
convolutional layer to predict the precipitation. The eleventh
2015 2016 2017 2018 2019 2020
sequence of precipitation was predicted from an input of ten
consecutive precipitation sequences at an interval of 30 Fig. 1. Frequency of publication during 2015-2020
minutes. The predicted precipitation sequence was further

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 China 17
10 India 9
10 9
USA 6
Number of articles

8 Thailand 2
6 Japan 2
4 Colombia
3 3 2
4
2 2 2 2 Vietnam 1
2 1 1 1 1 Taiwan 1
0 Germany 1
Brazil 1
ConvLSTM
ConvRNN

DeepESN

ESN
DBNPF

LSTM
CNN-GRU

MGGP
CNN

DBN

DNN
ConvGRU
CNN-LSTM

Australia 1
0 5 10 15 20
No. of papers
Fig. 2. Frequency of deep learning methods in the surveyed papers

the temporal information. We found that when compared with Fig. 3. Global distribution of the papers understudy
traditional machine learning models, deep learning models are
more accurate for rainfall prediction and they can capture the Correlation Coefficient (PCC), Probability of False Detection
temporal or spatial information of the input data. When the (PFD), Percentage Root mean square Difference (PRD),
input data consists of temporal information, traditional Stochastic Efficiency of Forecast (SEF), Nash–Sutcliffe
machine learning models cannot retain the past information Efficiency coefficient (NSE), Relative Bias (RB), Structural
which is required for predicting future rainfall. Similarity (SSIM), Success ratio and Variance.

C. Spatial distribution of studies V. FUTURE DIRECTIONS

In Fig. 3, we plotted the global distribution of the papers Despite all the research efforts and advancements in the
under study. It is based on the country in which the weather literature, there is still a room for further improvements. In the
data are collected and rainfall is predicted. The Majority of the following, we provide future directions in the area of rainfall
research was conducted in China (17 papers), followed by predictions.
India (9 papers) and the USA (6 papers).
i) In most of the papers, hyperparameters of the networks
D. Software used are searched using the trial and error method, and few papers
The type and frequency of software used for implementing used optimization algorithms to find the optimal
the machine learning models are given in Error! Reference hyperparameter. Research can be focused in this direction to
source not found.. More than half of the papers (57.8%) implement a hybrid model of optimization algorithms and
mention the software used, whereas 42.2% of the papers did deep learning methods.
not mention the software used for implementing the machine ii) Datasets used in most of the papers are weather
learning models. We found that MATLAB is the most parameters from a meteorological station or radar image or
frequently used software (7 papers), followed by Tensorflow satellite image. Research can be made to combine the data
(6 papers) and Keras (5 papers). It seems that due to the easy from the three sources to improve the prediction accuracy.
interface and support for machine learning provided by
MATLAB, Tensorflow and Keras are the most used software iii) Due lack of paper that consider all the popular deep
by the authors. Theano, Python, PyTorch, Platform for learning models with similar dataset and metrics, it is difficult
Artificial Intelligence (PAI), MapReduce, Hovorod, to see which model has the best performance. Thus, efforts can
Microsoft Cognitive Toolkit (CNTK), and Chainer were also be made in this direction to find the best model.
used in the paper under study.
TABLE I. FREQUENCY OF METRICS USED
E. Performance metrics
In Table IV we show the frequency of different No. Metrics Papers No. Metrics Papers
1 RMSE 20 18 RB 2
performance metrics used in the studied papers. In some 2 MAE 13 19 B-MAE 1
papers, the authors did not mention the metrics used, these are 3 CSI 10 20 B-MSE 1
not included in the figure. While some authors used a single 4 FAR 10 21 CV 1
metric, many authors used multiple metrics for comparing the 5 POD 9 22 EVS 1
performance of the models. We found that RMSE was the 6 CORR 8 23 FSS 1
most commonly used metrics (20 papers), followed by MAE 7 MSE 7 24 MB 1
8 HSS 5 25 MS-SSIM 1
(13 papers), CSI (10 papers), and False Alarm Rate (FAR) (10
9 TS 5 26 NRMSE 1
papers). Other metrics include Correlation (CORR), Balanced 10 F1-score 4 27 ORSS 1
Mean Absolute Error (B-MAE), Balanced Mean Squared 11 NSE 4 28 PCC 1
Error (B-MSE), Coefficient of Variation (CV), Equitable 12 Precision 4 29 PFD 1
threat score (ETS), Explained Variance Score (EVS), 13 Recall 4 30 PRD 1
Fractions Skill Score (FSS), Heidke Skill Score (HSS), 14 Accuracy 3 31 SEF 1
15 Bias 3 32 SSIM 1
Multiplicative Bias (MB), Multi-Scale Structural 16 ETS 3 33 Success Ratio 1
Similarity (MS-SSIM), Normalized root mean squared error 17 R-Squared 2 34 Variance 1
(NRMSE), Odds Ratio Skill Score (ORSS), Pearson

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 VI. CONCLUSIONS Data,” Front. Artif. Intell., vol. 3, no. February, pp. 1–23, 2020, doi:
10.3389/frai.2020.00004.
Rainfall prediction is still a challenging task due to the [15] N. O. Bushara and A. Abraham, “Computational intelligence in
complex and non-linear nature of the weather variables. But weather forecasting: A review,” J. Netw. Innov. Comput., vol. 1, no. 1,
due to the high impact of rainfall in our daily lives, it is still a pp. 320–331, 2013, [Online]. Available:
high research area. In this paper, we have surveyed 45 papers www.mirlabs.net/jnic/index.html.
published by well-known publishers. We classify rainfall [16] D. R. Nayak and P. Mahapatra, Amitav\Mishra, “A Survey on Rainfall
prediction based on the type of data used by the authors. Prediction using Artificial Neural Network,” Int. J. Comput. Appl., vol.
72, no. 16, pp. 32–40, 2013.
Rainfall and other weather phenomenon are usually collected
[17] D. Chauhan and J. Thakur, “Data Mining Techniques for Weather
as the weather parameter value, radar image, and satellite Prediction: A Review,” Int. J. Recent Innov. Trends Comput. Commun.,
image. We study the deep learning methods applied, types of vol. 2, no. 8, pp. 2184–2189, 2014, [Online]. Available:
input data used for the predictor, the type of metrics applied http://ijritcc.org/IJRITCC Vol_2 Issue_8/Data Mining Techniques for
for testing the performance of the models, and the software Weather Prediction A Review.pdf.
used for implementing the models. We also study the temporal [18] S. Sarumathi, N. Shanthi, and S. Vidhya, “Satellite Rainfall Prediction
and spatial distribution of the study. In conclusion, we find Techniques - A State of the Art Review,” Int. J. Environ. Chem. Ecol.
Geol. Geophys. Eng., vol. 9, no. 2, pp. 144–151, 2015.
that deep learning methods performed better and they are more
[19] N. Mishra, H. K. Soni, S. Sharma, and A. K. Upadhyay, “A
preferable compared to traditional machine learning models or comprehensive survey of data mining techniques on time series data
shallow neural network architecture for the task of rainfall for rainfall prediction,” J. ICT Res. Appl., vol. 11, no. 2, pp. 167–183,
prediction. 2017, doi: 10.5614/itbj.ict.res.appl.2017.11.2.4.
[20] S. Aftab, M. Ahmad, N. Hameed, M. S. Bashir, I. Ali, and Z. Nawaz,
ACKNOWLEDGMENT “Rainfall prediction using data mining techniques: A systematic
literature review,” Int. J. Adv. Comput. Sci. Appl., vol. 9, no. 5, pp. 143–
The authors would like to acknowledge the reviewers for their
150, 2018, doi: 10.14569/IJACSA.2018.090518.
valuable time and feedback.
[21] M. Saha, P. Mitra, and R. S. Nanjundiah, “Autoencoder-based
identification of predictors of Indian monsoon,” Meteorol. Atmos.
REFERENCES Phys., vol. 128, no. 5, pp. 613–628, 2016, doi: 10.1007/s00703-016-
[1] Y. Dash, S. K. Mishra, and B. K. Panigrahi, “Rainfall prediction for the 0431-7.
Kerala state of India using artificial intelligence approaches,” Comput. [22] M. Saha, P. Mitra, and R. S. Nanjundiah, “Deep learning for predicting
Electr. Eng., vol. 70, no. June, pp. 66–73, 2018, doi: the monsoon over the homogeneous regions of India,” J. Earth Syst.
10.1016/j.compeleceng.2018.06.004. Sci., vol. 126, no. 4, pp. 1–18, 2017, doi: 10.1007/s12040-017-0838-7.
[2] M. N. French, W. F. Krajewski, and R. R. Cuykendall, “Rainfall [23] M. Saha, A. Santara, P. Mitra, A. Chakraborty, and R. S. Nanjundiah,
forecasting in space and time using a neural network,” J. Hydrol., vol. “Prediction of the Indian summer monsoon using a stacked
137, no. 1–4, pp. 1–31, 1992, doi: 10.1016/0022-1694(92)90046-X. autoencoder and ensemble regression model,” Int. J. Forecast., vol. 37,
[3] A. Manzato, “Sounding-derived indices for neural network based no. 1, pp. 58–71, 2020, doi: 10.1016/j.ijforecast.2020.03.001.
short-term thunderstorm and rainfall forecasts,” Atmos. Res., vol. 83, [24] E. Hernández, V. Sanchez-Anguix, V. Julian, J. Palanca, and N. Duque,
no. 2-4 SPEC. ISS., pp. 349–365, 2007, doi: “Rainfall prediction: A deep learning approach,” in Lecture Notes in
10.1016/j.atmosres.2005.10.021. Computer Science (including subseries Lecture Notes in Artificial
[4] H. Meyer, M. Kühnlein, T. Appelhans, and T. Nauss, “Comparison of Intelligence and Lecture Notes in Bioinformatics), 2016, vol. 9648, doi:
four machine learning algorithms for their applicability in satellite- 10.1007/978-3-319-32034-2_13.
based optical rainfall retrievals,” Atmos. Res., vol. 169, pp. 424–433, [25] S. Gope, S. Sarkar, P. Mitra, and S. Ghosh, “Early prediction of
2016, doi: 10.1016/j.atmosres.2015.09.021. extreme rainfall events: A deep learning approach,” in Advances in
[5] S. Cramer, M. Kampouridis, A. A. Freitas, and A. K. Alexandridis, “An Data Mining. Applications and Theoretical Aspects. ICDM 2016,
extensive evaluation of seven machine learning methods for rainfall 2016, vol. 9728, pp. 154–167, doi: 10.1007/978-3-319-41561-1_12.
prediction in weather derivatives,” Expert Syst. Appl., vol. 85, pp. 169– [26] P. Zhang, L. Zhang, H. Leung, and J. Wang, “A Deep-Learning Based
181, 2017, doi: 10.1016/j.eswa.2017.05.029. Precipitation Forecasting Approach Using Multiple Environmental
[6] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. Factors,” in Proceedings - 2017 IEEE 6th International Congress on
521, no. 7553, pp. 436–444, 2015, doi: 10.1038/nature14539. Big Data, BigData Congress 2017, 2017, pp. 193–200, doi:
[7] C. Shen, “A Transdisciplinary Review of Deep Learning Research and 10.1109/BigDataCongress.2017.34.
Its Relevance for Water Resources Scientists,” Water Resour. Res., vol. [27] P. Zhang, Y. Jia, L. Zhang, J. Gao, and H. Leung, “A deep belief
54, no. 11, pp. 8558–8593, 2018, doi: 10.1029/2018WR022643. network based precipitation forecast approach using multiple
[8] S. Khan and T. Yairi, “A review on the application of deep learning in environmental factors,” Intell. Data Anal., vol. 22, no. 4, pp. 843–866,
system health management,” Mech. Syst. Signal Process., vol. 107, pp. 2018, doi: 10.3233/IDA-173427.
241–265, 2018, doi: 10.1016/j.ymssp.2017.11.024. [28] Q. Ouyang and W. Lu, “Monthly Rainfall Forecasting Using Echo
[9] W. Nash, T. Drummond, and N. Birbilis, “A review of deep learning State Networks Coupled with Data Preprocessing Methods,” Water
in the study of materials degradation,” npj Mater. Degrad., vol. 2, no. Resour. Manag., vol. 32, no. 2, pp. 659–674, 2017, doi:
1, pp. 1–12, 2018, doi: 10.1038/s41529-018-0058-x. 10.1007/s11269-017-1832-1.
[10] S. Pouyanfar et al., “A Survey on Deep Learning,” ACM Comput. [29] M. Qiu et al., “A short-term rainfall prediction model using multi-task
Surv., vol. 51, no. 5, pp. 1–36, 2019, doi: 10.1145/3234150. convolutional neural networks,” Proc. - IEEE Int. Conf. Data Mining,
ICDM, vol. 2017-Novem, pp. 395–404, 2017, doi:
[11] S. Dargan, M. Kumar, M. R. Ayyagari, and G. Kumar, “A Survey of 10.1109/ICDM.2017.49.
Deep Learning and Its Applications: A New Paradigm to Machine
Learning,” Arch. Comput. Methods Eng., vol. 27, no. 4, pp. 1071–1092, [30] J. Du, Y. Liu, and Z. Liu, “Study of precipitation forecast based on deep
2020, doi: 10.1007/s11831-019-09344-w. belief networks,” Algorithms, vol. 11, no. 9, pp. 1–11, 2018, doi:
10.3390/a11090132.
[12] H. Ismail Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P. A.
Muller, “Deep learning for time series classification: a review,” Data [31] K. Yonekura, H. Hattori, and T. Suzuki, “Short-term local weather
Min. Knowl. Discov., vol. 33, no. 4, pp. 917–963, 2019, doi: forecast using dense weather station by deep neural network,” in
10.1007/s10618-019-00619-1. Proceedings - 2018 IEEE International Conference on Big Data, Big
Data 2018, 2018, pp. 1683–1690, doi:
[13] L. Zhou, C. Zhang, F. Liu, Z. Qiu, and Y. He, “Application of Deep 10.1109/BigData.2018.8622195.
Learning in Food: A Review,” Compr. Rev. Food Sci. Food Saf., vol.
18, no. 6, pp. 1793–1811, 2019, doi: 10.1111/1541-4337.12492. [32] U. Weesakul, P. Kaewprapha, K. Boonyuen, and O. Mark, “Deep
learning neural network: A machine learning approach for monthly
[14] F. Emmert-Streib, Z. Yang, H. Feng, S. Tripathi, and M. Dehmer, “An rainfall forecast, case study in eastern region of thailand,” Eng. Appl.
Introductory Review of Deep Learning for Prediction Models With Big

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.
 Sci. Res., vol. 45, no. 3, pp. 203–211, 2018, doi: [51] S. Samsi, C. J. Mattioli, and M. S. Veillette, “Distributed deep learning
10.14456/easr.2018.24. for precipitation nowcasting,” arXiv, 2019.
[33] D. T. Anh, M. D. Bui, and P. Rutschmann, “A comparative study of [52] L. Tian, X. Li, Y. Ye, P. Xie, and Y. Li, “A Generative Adversarial
three different models to predict monthly rainfall in Ca Mau, Vietnam,” Gated Recurrent Unit Model for Precipitation Nowcasting,” IEEE
in Hydraulic engineering symposium Graz 2018. Water management - Geosci. Remote Sens. Lett., vol. 17, no. 4, pp. 601–605, 2019, doi:
innovation based on tradition, 2018, no. July, pp. 18–20. 10.1109/LGRS.2019.2926776.
[34] D. Kumar, A. Singh, P. Samui, and R. K. Jha, “Forecasting monthly [53] S. Agrawal, L. Barrington, C. Bromberg, J. Burge, C. Gazen, and J.
precipitation using sequential modelling,” Hydrol. Sci. J., vol. 64, no. Hickey, “Machine learning for precipitation nowcasting from radar
6, pp. 690–700, 2019, doi: 10.1080/02626667.2019.1595624. images,” arXiv, no. NeurIPS, pp. 1–6, 2019.
[35] Q. Miao, B. Pan, H. Wang, K. Hsu, and S. Sorooshian, “Improving [54] Q. K. Tran and S. K. Song, “Computer vision in precipitation
monsoon precipitation prediction using combined convolutional and nowcasting: Applying image quality assessment metrics for training
long short term memory neural network,” Water (Switzerland), vol. 11, deep neural networks,” Atmosphere (Basel)., vol. 10, no. 5, pp. 1–20,
no. 5, 2019, doi: 10.3390/w11050977. 2019, doi: 10.3390/atmos10050244.
[36] S. Viswanath, M. Saha, P. Mitra, and R. S. Nanjundiah, Deep Learning [55] changjiang zhang, H. Wang, J. Zeng, L. Ma, and L. Guan, “Tiny-
Based LSTM and SeqToSeq Models to Detect Monsoon Spells of India, RainNet: A Deep CNN-BiLSTM Model for Short-Term Rainfall
vol. 11537 LNCS. Springer International Publishing, 2019. Prediction,” Earth Sp. Sci. Open Arch., 2019, doi:
[37] M. H. Yen, D. W. Liu, Y. C. Hsin, C. E. Lin, and C. C. Chen, 10.1002/essoar.10501344.1.
“Application of the deep learning for the prediction of rainfall in [56] G. Ayzel, T. Scheffer, and M. Heistermann, “RainNet v1.0: A
Southern Taiwan,” Sci. Rep., vol. 9, no. 1, pp. 1–9, 2019, doi: convolutional neural network for radar-based precipitation
10.1038/s41598-019-49242-6. nowcasting,” Geosci. Model Dev., vol. 13, no. 6, pp. 2631–2644, 2020,
[38] F. Manokij, K. Sarinnapakorn, and P. Vateekul, “Forecasting doi: 10.5194/gmd-13-2631-2020.
Thailand’s Precipitation with Cascading Model of CNN and GRU,” [57] L. Chen, Y. Cao, L. Ma, and J. Zhang, “A Deep Learning-Based
Oct. 2019, doi: 10.1109/ICITEED.2019.8929975. Methodology for Precipitation Nowcasting With Radar,” Earth Sp.
[39] S. Poornima and M. Pushpalatha, “Prediction of rainfall using Sci., vol. 7, no. 2, 2020, doi: 10.1029/2019EA000812.
intensified LSTM based recurrent Neural Network with Weighted [58] Q. Yan, F. Ji, K. Miao, Q. Wu, Y. Xia, and T. Li, “Convolutional
Linear Units,” Atmosphere (Basel)., vol. 10, no. 11, 2019, doi: Residual-Attention: A Deep Learning Approach for Precipitation
10.3390/atmos10110668. Nowcasting,” Adv. Meteorol., vol. 2020, 2020, doi:
[40] L. Ni et al., “Streamflow and rainfall forecasting by two long short- 10.1155/2020/6484812.
term memory-based models,” J. Hydrol., vol. 583, p. 124296, 2019, [59] C. Kidd, V. Levizzani, J. Turk, and R. Ferraro, “Satellite precipitation
doi: 10.1016/j.jhydrol.2019.124296. measurements for water resource monitoring,” J. Am. Water Resour.
[41] C. J. Zhang, J. Zeng, H. Y. Wang, L. M. Ma, and H. Chu, “Correction Assoc., vol. 45, no. 3, pp. 567–579, 2009, doi: 10.1111/j.1752-
model for rainfall forecasts using the LSTM with multiple 1688.2009.00326.x.
meteorological factors,” Meteorol. Appl., vol. 27, no. 1, pp. 1–15, 2019, [60] Y. Tao, X. Gao, K. Hsu, S. Sorooshian, and A. Ihler, “A deep neural
doi: 10.1002/met.1852. network modeling framework to reduce bias in satellite precipitation
[42] K. Zhou, Y. Zheng, B. Li, W. Dong, and X. Zhang, “Forecasting products,” J. Hydrometeorol., vol. 17, no. 3, pp. 931–945, 2016, doi:
Different Types of Convective Weather: A Deep Learning Approach,” 10.1175/JHM-D-15-0075.1.
J. Meteorol. Res., vol. 33, no. 5, pp. 797–809, 2019, doi: [61] A. Akbari Asanjan, T. Yang, K. Hsu, S. Sorooshian, J. Lin, and Q.
10.1007/s13351-019-8162-6. Peng, “Short-Term Precipitation Forecast Based on the PERSIANN
[43] O. Manoj and J. P. Ananth, “MapReduce and Optimized Deep Network System and LSTM Recurrent Neural Networks,” J. Geophys. Res.
for Rainfall Prediction in Agriculture,” Comput. J., vol. 63, no. 6, pp. Atmos., vol. 123, no. 22, pp. 12,543-12,563, 2018, doi:
900–912, 2019, doi: 10.1093/comjnl/bxz164. 10.1029/2018JD028375.
[44] J. Pudashine et al., “Deep Learning for an Improved Prediction of [62] Y. M. Souto, F. Porto, A. M. Moura, and E. Bezerra, “A Spatiotemporal
Rainfall Retrievals From Commercial Microwave Links,” Water Ensemble Approach to Rainfall Forecasting,” in Proceedings of the
Resour. Res., vol. 56, no. 7, pp. 0–3, 2020, doi: International Joint Conference on Neural Networks, 2018, vol. 2018-
10.1029/2019WR026255. July, pp. 1–8, doi: 10.1109/IJCNN.2018.8489693.
[45] P. Hewage, M. Trovati, E. Pereira, and A. Behera, “Deep learning- [63] R. Sato, H. Kashima, and T. Yamamoto, Short-term precipitation
based effective fine-grained weather forecasting model,” Pattern Anal. prediction with skip-connected PredNet, vol. 11141 LNCS. Springer
Appl., no. 0123456789, 2020, doi: 10.1007/s10044-020-00898-1. International Publishing, 2018.
[46] X. Shi, Z. Chen, H. Wang, D. Y. Yeung, W. K. Wong, and W. C. Woo, [64] H. Wu, Q. Yang, J. Liu, and G. Wang, “A spatiotemporal deep fusion
“Convolutional LSTM network: A machine learning approach for model for merging satellite and gauge precipitation in China,” J.
precipitation nowcasting,” Adv. Neural Inf. Process. Syst., vol. 2015- Hydrol., vol. 584, no. 2006, p. 124664, 2020, doi:
Janua, no. June, pp. 802–810, 2015. 10.1016/j.jhydrol.2020.124664.
[47] A. Heye, K. Venkatesan, and J. Cain, “Precipitation Nowcasting: [65] A. Kumar, T. Islam, Y. Sekimoto, C. Mattmann, and B. Wilson,
Leveraging Deep Recurrent Convolutional Neural Networks,” 2017, “ConvCast: An embedded convolutional LSTM based architecture for
[Online]. Available: precipitation nowcasting using satellite data,” PLoS One, vol. 15, no.
https://cug.org/proceedings/cug2017_proceedings/includes/files/pap1 3, pp. 1–18, 2020, doi: 10.1371/journal.pone.0230114.
55s2-file1.pdf. [66] M. Saha, P. Mitra, and R. S. Nanjundiah, “Predictor discovery for
[48] X. Shi et al., “Deep learning for precipitation nowcasting: A early-late Indian Summer Monsoon using stacked autoencoder,”
benchmark and a new model,” arXiv, no. Nips, pp. 1–11, 2017. Procedia Comput. Sci., vol. 80, pp. 565–576, 2016, doi:
10.1016/j.procs.2016.05.337.
[49] S. Singh, S. Sarkar, and P. Mitra, “Leveraging convolutions in
recurrent neural networks for doppler weather radar echo prediction,” [67] F. Manokij, K. Sarinnapakorn, and P. Vateekul, “Forecasting
Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Thailand’s Precipitation with Cascading Model of CNN and GRU,” in
Lect. Notes Bioinformatics), vol. 10262 LNCS, pp. 310–317, 2017, doi: 2019 11th International Conference on Information Technology and
10.1007/978-3-319-59081-3_37. Electrical Engineering (ICITEE), Pattaya, Thailand, 2019, 2019, pp.
1–6, doi: 10.1109/ICITEED.2019.8929975.
[50] S. Kim, S. Hong, M. Joh, and S. K. Song, “DeepRain: ConvLSTM
network for precipitation prediction using multichannel radar data,”
arXiv, pp. 3–6, 2017.

Authorized licensed use limited to: NIELIT Aizawl. Downloaded on September 27,2023 at 07:15:47 UTC from IEEE Xplore. Restrictions apply.