Rainfall Pattern in Kenya: Bayesian Non-parametric Model Based on the Normalized Generalized Gamma Process

Understanding the pattern of rainfall in Kenya is crucial for a range of sectors, including agriculture, water management, and disaster risk reduction. In this research, we propose a Bayesian non-parametric approach to model the rainfall patterns in Kenya. Specifically, we use a hierarchical Dirichlet process mixture model to cluster the rainfall stations and identify groups of stations with similar rainfall patterns. We then model the rainfall distribution within each group using a Bayesian non-parametric model based on the normalized generalized gamma process. We apply our method to a dataset of daily rainfall measurements from 150 stations across Kenya for the period 1980-2021. Our results reveal distinct regional patterns of rainfall, with some regions experiencing bimodal rainfall patterns while others have unimodal patterns. We also find that the rainfall distribution within each region exhibits heavy tails and skewedness, which cannot be accurately captured by parametric models. In conclusion, our approach provides a flexible and interpretable framework for modeling complex spatio-temporal data such as rainfall patterns, and can inform decision-making in various sectors.