Machine learning enabled extreme rainfall prediction using GNSS, Radar, lightning, and weather data for Indian coastal city

Extreme weather events, such as heavy rainfall, is frequent nowadays and present significant challenges across various sectors, including aviation, transportation, and public safety. Predicting such weather events within a short time frame is critical for applications like aviation safety, urban planning, and disaster management. This study focuses on developing and implementing an advanced machine learning model for predicting extreme precipitation for Thiruvananthapuram, Kerala, India, a city which experiences a tropical monsoon climate, characterized by high humidity and significant rainfall throughout the year. The research leverages weather, GNSS, radar, and lightning data for this study. In India, since a few sites are available with both radar and lightning data apart from weather and GNSS station, thus, making this study more relevant to attempt the rainfall prediction work involving multi-source meteorological data. The Thumba Equatorial Rocket Launching (TERL) station site serves as the study area due to its C-band Radar station and collocated with GNSS receiver. Weather data is sourced from the ERA5 reanalysis dataset provided by the Copernicus Climate Change Service, ECMWF due to low availability of weather station data prior to 2021. The methodology involves integrating time-series hourly datasets to train a Long Short-Term Memory (LSTM) model, with a batch size of 32, 128 neuron each in 128 layers. The model is trained on a comprehensive dataset encompassing three years (2019-2021) of data from abovementioned sources, with 80% used for training and 20% for testing. The study demonstrates that an LSTM model with tuned hyper-parameters is well-suited for meteorological variables and predicting rainfall. Quantitative results show the model's high accuracy in predicting extreme precipitation events and the performance was assessed using the Probability of Detection (POD), False Alarm Rate (FAR), and Critical Success Index (CSI). The nowcasting results are promising, with a high POD, a FAR of 4.96%, and a CSI of 95.03%. Different configurations of input datasets are evaluated to determine the optimal combination for accurate predictions. It is observed that model performance with data combination involving radar and lightning in addition to weather, GNSS, windspeed data is better with rmse 1.12 as compared to weather, windspeed and GNSS data only.  Another data configuration using only weather, GNSS, and radar data also showed excellent performance, making it preferable due to its adaptability to various weather conditions in absence of lightning data. These metrics encourage to implement a similar framework with a high accuracy in detecting extreme precipitation events, a low rate of false alarms and strong performance in nowcasting extreme rain events within the first 60 minutes.