Study of Rainfall Occurrence Process by Markov Chain Models and Decision Tree-based Ensemble and Boosting Techniques

Rainfall prediction is vital for water resource management, agricultural planning, and urban design. While extensive research exists on rainfall magnitude forecasting, less attention has been given to predicting rainfall occurrences. This study addresses this gap by examining rainfall state prediction in four cities in India: Bhubaneshwar, Pune, Bangalore, and Hyderabad. Traditionally, the Markov Chain model has been used to simulate rainfall occurrences, but higher-order Markov models remain unexplored. Additionally, not many studies have utilized Machine Learning (ML) models for rainfall state prediction. This research compares Decision Tree-based Random Forest and Extreme Gradient Boosting (XGBoost) techniques with 1st, 2nd, and 3rd order Markov Chains. Results indicate that Random Forest and XGBoost outperform Markov Chain models in predicting daily rainfall states. However, for rainfall statistics like wet days, dry spell duration, and rain event length, the stochastic Markov Chain model proves more effective. This study’s findings are crucial for enhancing model selection criteria, thereby improving the efficiency and applicability of rainfall occurrence models in various fields. Graphical Abstract