Machine Learning Based Structural Characterization of Rainfall Event Typology and Regime Shifts in Pakistan (1981–2024)

This study develops an event-based morphological framework for analyzing 44 years of dekadal rainfall data across 30 districts of Pakistan (1981–2024). Moving beyond aggregate precipitation metrics, rainfall is modeled as discrete events and represented through eight structural features capturing intensity, duration, temporal gradients, and antecedent moisture conditions.

A total of 4,563 rainfall events are extracted and grouped into seven distinct typologies using Gaussian Mixture Models. Regime shift analysis via the PELT algorithm identifies a persistent structural transition in typology composition around 1996, consistent with reported changes in South Asian monsoon dynamics. Four machine learning models are evaluated, with Random Forest achieving the strongest performance (96.6% accuracy, 0.94 macro F1), while per-class analysis highlights reduced performance on minority typologies.

The findings show that temporal structure and pre-event conditions provide stronger discriminative power than aggregate intensity measures. The framework offers a scalable approach for structural rainfall analysis with applications in flood risk assessment, agricultural planning, and climate pattern monitoring.

Code repository:
https://github.com/AbdullahPatti/Pakistan-Rainfall-Morphology-Phonology-Event-Typology-Regime-Shifts-Explainable-ML

Interactive dashboard:
https://pakistan-rainfall-analysis.streamlit.app/