Rainfall variability over Indian cities: Analysis and future prediction through AI/ML techniques

Abstract

Applications of rainfall analysis and forecasting, range from disaster management to agriculture. Due to the increasing impacts of climate change, occurrence of frequent and extreme rainfall events are expected to trigger severe floods, landslides, etc. Therefore, it is important to make a accurate prediction so that the intensity of the impacts on life and property could be reduced. In recent times, the advancement of AI/ML techniques has enabled researchers reasonably to apply them in weather and climate science. The present work is an effort in this direction that uses two climatological rainfall datasets for the analysis and prediction of temporal rainfall patterns across India, by considering city-specific information. Specifically, Deep learning (DL) approaches like Long Short Term Memory (LSTM), Bi-directional LSTM (BiLSTM), Gated Recurrent Unit (GRU), and Convolution LSTM (ConvLSTM) are considered for long-term rainfall prediction over hundred selected cities of India based on their location. Both univariate and multivariate analysis and forecasting is carried out in this study. Performance indicators like root means square error (RMSE) are computed to test the model training accuracy. The intercomparison of results indicate that ConvLSTM can be preferred for the analysis and forecasting of temporal rainfall variability irrespective of whether one follows univariate or multivariate approach. However, the performances of BiLSTM, and GRU is reasonably better in specific instances. Notably, the study is accompanied with city-based trend analysis of rainfall using the Mann-Kendall test. The current study also demonstrates the forecasting skills of the considered DL models till 2031 by adopting a city-based approach.