DEEP LEARNING ENABLED MULTI-STEP NOWCASTING RAINFALL

USING GEODETIC AND METEOROLOGICAL OBSERVATIONS FOR

INDIAN COASTAL CITY
Sumit Pandeya, Devara Meghanadhb and Ramji Dwivedia
a
Geographic Information System (GIS) Cell
Motilal Nehru National Institute of Technology Allahabad
Prayagraj (U.P.) 211004, India
b
Mobis Technical Center of India (Hyundai Mobis R&D)
[email protected]
[email protected]
[email protected]

discharge rates, which can provide critical lead time for

severe weather warnings [4].
ABSTRACT Precipitable water vapor (PWV) plays a significant role in
the development of severe weather events, as even a slight
Accurate short-term forecasting, or nowcasting, of extreme increase in WV can trigger intense rainfall. Numerous
weather events like rainfall is vital for aviation, studies have explored the use of GNSS-derived PWV for
transportation, and public safety. This study develops a various meteorological applications, including analysing
Long Short-Term Memory (LSTM) model to forecast the spatio-temporal variation of PWV and predicting
precipitation in Thiruvananthapuram, Kerala, leveraging a extreme rainfall. It is evident that weather station data and
unique combination of hourly data from 2019 to 2021, from GNSS derived zenith wet delay (ZWD), representing high-
ERA5, GNSS-derived water vapor and doppler weather resolution information on atmospheric moisture, captures
radar from the Thumba Equatorial Rocket Launching the formation of extreme event due to its sensitivity to the
station. The developed model was optimized through pre-convection environment. On the other hand, the radar
and lightning data provide real-time insights into
hyperparameter tuning, achieving high performance with
immediate understanding of severe weather formation and
70 epochs, a batch size of 32, and 12-time step sequences further intensification. When weather station data is
for nowcasting one hour with two-time step. Results combined with GNSS derived ZTD, radar and lightning
indicate a Probability of Detection (POD) exceeding 90%, data, this resultant data information becomes crucial input
a False Alarm Rate (FAR) and a Critical Success Index for identifying the preconditions that lead to convective
(CSI) better than 9%, and of 90% respectively for each storms [2].
output step. The configuration combining weather data, Traditional weather forecasting methods often
GNSS, and radar, yielded the highest correlation better than lack the precision and timeliness necessary for applications
0.998, and RMSE of 0.439 and 0.465 for first and second in aviation, transportation, and public safety, particularly
hour respectively, demonstrating the model's robustness in when dealing with severe weather events, as they fail to
capturing complex weather dynamics. provide the accuracy needed for real-time decision-making
[5]. These limitations have prompted the initial
investigation into conventional machine learning
Index Terms— LSTM, nowcasting, autoencoder.
techniques, followed by more recent advancements in
machine learning, particularly LSTM networks, which
1. INTRODUCTION have demonstrated potential in managing the complexities
of atmospheric data [6]. Recent advances in machine
The escalating frequency and intensity of severe weather learning, particularly the application of Recurrent Neural
phenomena in India, such as cyclones and floods, have Networks (RNNs) and their variants, have introduced
raised significant concerns for disaster management effective tools for forecasting weather events. RNNs,
(NDMA). This trend correlates with changes in especially LSTM networks are highly effective for time-
atmospheric conditions, including heightened water vapor series data as they can capture long-term dependencies [7].
in the lower troposphere[1], [2]. Heightened water vapor in Additionally, while existing studies have explored
the lower troposphere influences RADAR by altering the integration of various data sources, a need for more
signal reflectivity and attenuation, increases lightning robust methodologies that effectively combine ground-
activity due to enhanced convection, and affects GNSS based observations, satellite data, and real-time
signals by causing delays that can be used to estimate measurements is reported to enhance predictive accuracy
atmospheric moisture [3]. Recent advancements, including [8]. Furthermore, the temporal resolution of current models
Ground-based Global Navigation Satellite System (GNSS), often falls short of the rapid changes associated with
have enhanced nowcasting capabilities by estimating extreme weather events, highlighting the necessity for
Integrated Water Vapor (IWV) and monitoring lightning models that can provide high-frequency updates [3].
 Machine learning techniques, particularly LSTM which can lead to severe flooding, particularly in low-lying
models, have demonstrated significant potential in areas. The city has witnessed significant flood events in
forecasting and nowcasting applications. Recent studies recent years, exacerbated by urbanization and inadequate
indicate that LSTM approaches can outperform traditional drainage systems. The combination of high rainfall
numerical weather prediction models, providing efficient intensity and the potential for extreme weather events poses
and accurate forecasts up to 12 hours [3], [9]. The recent considerable challenges for disaster management and
applications of LSTM for time series prediction have been necessitates the development of effective forecasting
greatly benefited from the availability of data source having models to mitigate the impacts of such occurrences on the
a high spatio-temporal resolution from a dense sensor local population and economy.
network facilitating the integration of multiple input
sources, including weather station data, GNSS
measurements, radar reflectivity, and lightning data [10].
Such integration is the need of the hour for severe weather
forecast as it is hard to detect majority of strong convection.
It has been recommended in various previous studies for
improving the nowcasting.
This study focuses on developing a deep learning
model, specifically an LSTM model, to predict
precipitation events in Thiruvananthapuram, Kerala,
India—a region characterized by a tropical monsoon
climate. The study leverages a unique combination of
weather, GNSS, and radar data particularly relevant due to
the limited availability of comprehensive meteorological
datasets in the country. In India, only a few sites are
available with both radar and lightning data apart from Fig. 1: Study Area – Thiruvananthapuram (Kerala)
weather and GNSS station, thus, making this study more 2.1. Data Collection
relevant to attempt the rainfall nowcasting work involving The study leverages data from various sources to enhance
multi-source meteorological data. The research utilizes data the accuracy of nowcasting. The primary data sources
from multiple sources: radar data from the Thumba include:
Equatorial Rocket Launching (TERL) station, GNSS data
for atmospheric water vapor content, and weather data from Weather Data: Collected from the ERA5 reanalysis
the ERA5 reanalysis dataset by the Copernicus Climate dataset by the Copernicus Climate Change Service,
Change Service. The limited availability of comprehensive providing information on temperature, humidity, wind
meteorological datasets in this area underscores the speed, and other meteorological variables.
relevance of this research.
GNSS Data: Obtained from the Indian Space Research
2. STUDY AREA AND DATASET Organisation (ISRO), offering high-resolution information
on atmospheric water vapor content.
Thiruvananthapuram, the capital of Kerala, India, as
Radar Data: Sourced from the Thumba Equatorial Rocket
depicted in Figure 1, is located in a tropical monsoon
Launching (TERL) station, providing real-time monitoring
climate zone, marked by distinct wet and dry seasons. The
of precipitation intensity and movement from MOSDAC
region experiences high humidity and significant rainfall,
(https://www.mosdac.gov.in/)
particularly during the southwest monsoon, which typically
lasts from June to September. In terms of extreme weather
events, Thiruvananthapuram is prone to heavy rainfall,

Table 1: List of input variable

Variable Availability Unit Station Type Lat (0) Lon (0) Period

0
Temperature 100% C Weather 8.5 76.5 2019-2021
Surface Pressure 100% mb Weather 8.5 76.5 2019-2021
Precipitation 100% mm Weather 8.5 76.5 2019-2021
U & V Component of 100% m/s Weather 8.5 76.5 2019-2021
Wind Velocity
Relative Humidity 100% % Weather 8.5 76.5 2019-2021
ZTD 100% m GNSS 8.5 76.5 2019-2021
Reflectivity low dBZ Radar 8.5 76.5 2019-2021
 Fig. 2: Seasonal Variation Over a Three-Year Period: Trends and Patterns

3. THEORETICAL BACKGROUND: LSTM 1. Forget Gate: The forget gate determines which
information in the cell state should be discarded or
3.1. Background Theory preserved. It analyses the previous cell state and
determines which pieces of information are no longer
LSTM networks are a sophisticated form of RNNs relevant to the current context and should be
specifically designed to address the limitations of standard "forgotten." By doing so, the forget gate helps prevent
RNNs, particularly in retaining and modelling long-term the cell from becoming cluttered with outdated or
dependencies in sequential data. Conventional RNNs unnecessary data, ensuring that only useful
frequently encounter difficulties in learning such information is carried forward in the sequence [11]. It
dependencies due to problems like vanishing or exploding is computed as:
gradients. These problems arise during the backpropagation
process, where gradients either diminish to near zero or ft =  (W f [ht −1 , xt ] + b f ) (1)
grow exponentially, leading to unstable training and
difficulty in retaining information over extended sequences where ft is the output of the forget gate at any time step t,
[7]. W f is the weight matrix, ht −1 is the hidden state of the
LSTM networks address these challenges by
incorporating a sophisticated mechanism of gates within previous time step, xt is the current input, and b f is the
their architecture. Specifically, LSTMs utilize three
bias term. The sigmoid function σ ensures that the gate
primary gates—namely, the input gate, forget gate, and
output is between 0 and 1.
output gate. These gates work in concert to carefully
2. Input Gate: The input gate controls the amount of new
manage the flow of information within the network,
information from the current input that is incorporated
deciding which information to keep, update, or discard as it
into the cell state. It decides what portion of the
processes the sequence. This gating mechanism allows
incoming data is relevant and should be considered for
LSTM networks to maintain a stable gradient, effectively
updating the memory. This gate plays a crucial role in
capturing dependencies across long sequences and enabling
determining which parts of the new input should
the model to remember information over much longer
influence the cell state, allowing the network to
periods, making them particularly well-suited for tasks
selectively update its knowledge based on the most
involving time series data or sequences where long-term
important aspects of the new information [11]. It is
context is crucial. [7].
computed as:
Mathematically, an LSTM unit is composed of several
components: it =  (Wi [ht −1 , xt ] + bi ) (2)
The candidate cell state Ct is generated using a tanh learning to compress input data and subsequently
reconstruct it, enabling the identification of latent patterns
activation function:
within the data that may improve model performance when
Ct = tanh(WC  [ht −1 , xt ] + bC ) (3) integrated with LSTM for predictive tasks.
The cell state is then updated as: For example, [12], it has utilized autoencoders in
combination with LSTM to enhance weather forecasting by
Ct = ft  Ct −1 + it  Ct (4) reducing the noise in satellite and ground-based
3. Output Gate: The output gate controls what part of measurements, resulting in more accurate predictive
the cell state should be exposed as the output at each models.
time step. This gate determines how much of the The model architecture includes an input layer, 2
current cell state should be passed on to the next layer LSTM layers, one repeat vector layer and 2 dense output
or as the final output of the LSTM unit. It ensures that layers. The input layer accepts time-series data from
the output reflects the most relevant information while various meteorological sources, while the LSTM layers
keeping other parts of the cell state intact for future process this data to learn temporal patterns. The dense
time steps. This selective process enables the LSTM to output layer generates predictions for the target variable—
focus on producing meaningful outputs that are precipitation.
informed by both the current input and the long-term
context stored in the cell state [11]. It is calculated as: 3.2.2. Model Post Processing
ot =  (Wo  [ht −1 , xt ] + bo ) (5) To refine the model's predictions, several post-processing
techniques were applied. These included:
The ht , the hidden state of the current time step is then a) Normalization of outputs to align with the range
computed as: of observed values.
ht = ot  tanh(Ct ) (6) b) Application of correction models to adjust for
systematic biases observed during testing. Various
These equations allow LSTM networks to maintain long- regression models, such as Polynomial and Ridge
term memory over time, making them highly effective for Regression, were evaluated for this purpose and
tasks involving sequential data, such as weather Extra Trees Regression model was chosen.
nowcasting.
3.2.3. Model Training and Evaluation
The LSTM model was trained using a design of
experiments approach, which involved extensive
hyperparameter tuning. The hyperparameters adjusted
included the number of epochs, number of LSTM units,
batch size, and the input sequences length. The optimal
configuration was determined to be 70 epochs, a batch size
of 32, 128 dense neurons, and 128 LSTM units, with an
Fig. 3: LSTM Architecture input sequence length of 12-time steps.
The model's loss function, which guides the
3.2. Design of Experiment optimization process during training, was also carefully
selected. After comparing several options, including mean
3.2.1. Model Preparation squared error and mean squared logarithmic error, the
The foundation of the predictive model is an LSTM Huber loss function was chosen due to its robustness to
network, specifically designed to effectively capture long- outliers. The final model achieved a test loss of 3.31×10-5,
term dependencies present in weather-related data In this indicating its ability to generalize well to unseen data.
study, we utilize LSTM autoencoders—sophisticated
neural network structures that are employed to learn and 3.2.4. Nowcasting Performance Assessment
extract efficient representations of data in an unsupervised Model nowcasting performance was assessed using several
manner An autoencoder is typically divided into two metrics:
primary components: the encoder and the decoder. The a) Probability of Detection (POD): POD evaluates
encoder functions to reduce or transform the input data into the ability of the model to correctly predict the
a lower-dimensional latent space, capturing the essential occurrence of extreme weather events.
characteristics of the data. In contrast, the decoder is b) False Alarm Ratio (FAR): Assesses the
responsible for reconstructing the original input from this proportion of incorrect predictions relative to all
compressed latent representation. This architecture is positive predictions.
particularly advantageous when dealing with complex c) Critical Success Index (CSI): The Critical
datasets, such as weather data, as it allows for significant Success Index (CSI) is a statistical measure used
dimensionality reduction while maintaining the integrity of to evaluate the accuracy of forecasts, particularly
crucial data features [4]. Autoencoders are commonly used in meteorology. It is calculated by comparing the
for anomaly detection, noise reduction, and feature number of correct predictions of an event (hits)
extraction. They generally have the characteristics of with the total number of instances where the event
 was either predicted or actually occurred. Initially, the RMSE is observed more than one for the
Specifically, CSI is the ratio of the number of true nowcasted results which is further corrected using Extra
positive forecasts (hits) to the sum of hits, false trees Regressor corrector model [13]. It is evident from
alarms (events predicted but not occurred), and figure 4 that nowcasting has a similar trend to that of
misses (events occurred but not predicted). This observation, showing a bias due to intensity of the rain.
metric provides a balanced assessment of forecast Further, the trend ceases to follow the trend leading to
performance, taking into account both the higher RMSE in further output time step.
successes and the errors, making it particularly Incorporating multi-source data as well fine
useful in scenarios where both false alarms and tuning the hyperparameters was pivotal to the model's
missed events need to be minimized [9]. success. The RMSE corresponding to the 12 input steps
The model demonstrated a high POD and CSI, with a comes at the lowest and therefore, adopted for this study.
FAR of 9.19%, underscoring its effectiveness in The configuration that included weather data, GNSS-
nowcasting severe weather. derived ZWD, and radar achieved the highest correlation
(0.9988) post-correction, and the lowest Root Mean Square
4. RESULTS AND DISCUSSION
Error (RMSE) of 0.4337, and a Mean Absolute Error
The results of the study indicate that the LSTM-based (MAE) of 0.0343 for the first time step i.e. one hour. These
nowcasting model is highly effective in predicting rainfall results highlight the importance of data diversity in
events. The model's performance varied with lead time, improving nowcasting accuracy. For both one and two
with higher accuracy observed at shorter lead times. The output time-steps, results indicate a POD exceeding 90%, a
integration of weather data (WS), GNSS and radar data was FAR and a CSI better than 9%, and of 90% respectively.
particularly beneficial, enhancing the model's ability to
detect the onset and intensity of convective storms.

Fig. 4: Comparison of actual precipitation and predicted precipitation (before and after correction)
Table 2: Result metrics before and after correction for different output length
Version I/P O/P Loss Correlation Correlation RMSE RMSE
Sequence Before After Before After
WS+GNSS+RADAR 12,2 0.00010469 0.9901 0.998 1.293 0.465
WS+GNSS+RADAR 12,1 0.0000334 0.991 0.998 1.135 0.437
 Figure 5: Loss function curve

5. CONCLUSION vol. 48, no. 2, Apr. 2010, doi:

10.1029/2009RG000301.
This study demonstrates the potential of LSTM networks
for nowcasting rainfall for Thiruanantpuram, Kerala, India. [3] P. Hewage, M. Trovati, E. Pereira, and A. Behera,
The presented study can be further extended to other study “Deep learning-based effective fine-grained
sites, particularly in regions with complex weather patterns weather forecasting model,” Pattern Analysis and
and limited meteorological data infrastructure. The Applications, vol. 24, no. 1, pp. 343–366, Feb.
integration of multi-source data, including weather data, 2021, doi: 10.1007/s10044-020-00898-1.
GNSS-derived atmospheric moisture content, and radar
data was critical to the model's success. [4] E. Williams et al., “The behavior of total lightning
Multiple experiments have helped in identifying activity in severe Florida thunderstorms,” 1999.
the critical hyperparameters as well as the data input
configuration for the proposed model. The LSTM model is [5] A. R. Siems-Anderson, C. L. Walker, G. Wiener, W.
highly effective for weather prediction tasks due to its P. Mahoney, and S. E. Haupt, “An adaptive big data
capability to capture long-range dependencies in time- weather system for surface transportation,” Transp
series data. The study's findings suggest that further Res Interdiscip Perspect, vol. 3, Dec. 2019, doi:
improvements in nowcasting accuracy can be achieved by 10.1016/j.trip.2019.100071.
incorporating additional data sources and refining the
model architecture. The proposed LSTM model is [6] K. S. Rao, G. L. Devi, and N. Ramesh, “Air Quality
investigated on a single location and can be applied to the Prediction in Visakhapatnam with LSTM based
entire area in future subject to availability of the data. Recurrent Neural Networks,” International
Future work will explore these avenues, with the aim of Journal of Intelligent Systems and Applications,
enhancing early warning systems and mitigating the vol. 11, no. 2, pp. 18–24, Feb. 2019, doi:
impacts of extreme weather events. 10.5815/ijisa.2019.02.03.
The results of this study have significant
implications for improving weather nowcasting in regions [7] S. Hochreiter and J. ¨ Urgen Schmidhuber, “Long
with limited data availability. By leveraging advanced Short-Term Memory.”
machine learning techniques and multi-source data
integration, this research contributes to the development of [8] H. Ma et al., “Evaluation of six satellite- and
more accurate and reliable weather forecasting models, model-based surface soil temperature datasets
which are essential for safeguarding public safety and using global ground-based observations,” Remote
optimizing operations in weather-sensitive sectors. Sens Environ, vol. 264, Oct. 2021, doi:
10.1016/j.rse.2021.112605
REFERENCES
[1] K. E. Trenberth, J. Fasullo, and L. Smith, “Trends [9] S. Furkan Tekin, A. Fazla, and S. Serdar Kozat,
and variability in column-integrated atmospheric “Numerical Weather Forecasting Using
water vapor,” Clim Dyn, vol. 24, no. 7–8, pp. 741– Convolutional-LSTM with Attention and Context
758, May 2005, doi: 10.1007/s00382-005-0017-4. Matcher Mechanisms,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 62, 2024,
[2] S. C. Sherwood, R. Roca, T. M. Weckwerth, and N. doi: 10.1109/TGRS.2024.3409084.
G. Andronova, “Tropospheric water vapor,
convection, and climate,” Reviews of Geophysics,
[10] L. Wang, M. Zhang, X. Gao, and W. Shi, [12] X. Tong, J. Wang, C. Zhang, T. Wu, H. Wang, and
“Advances and Challenges in Deep Learning- Y. Wang, “LS-LSTM-AE: Power load forecasting
Based Change Detection for Remote Sensing via Long-Short series features and LSTM-
Images: A Review through Various Learning Autoencoder,” Energy Reports, vol. 8, pp. 596–
Paradigms,” Mar. 01, 2024, Multidisciplinary 603, Apr. 2022, doi: 10.1016/j.egyr.2021.11.172.
Digital Publishing Institute (MDPI). doi:
10.3390/rs16050804. [13] M. W. Ahmad, J. Reynolds, and Y. Rezgui,
“Predictive modelling for solar thermal energy
[11] F. Zhang et al., “PewLSTM: Periodic LSTM with systems: A comparison of support vector
Weather-Aware Gating Mechanism for Parking regression, random forest, extra trees and
Behavior Prediction,” 2020. [Online]. Available: regression trees,” J Clean Prod, vol. 203, pp. 810–
https://www.statista.com/statistics/183505/number 821, Dec. 2018, doi:
-of-vehicles- 10.1016/j.jclepro.2018.08.207.