Applicability of statistical and deep-learning models for rainfall disaggregation at metropolitan stations in India

Abstract

Study regionMetropolitan stations in India with distinct climatic conditions, namely, Delhi, Chennai, Kolkata, and Mumbai were selected for this study. Study focusRainfall disaggregation models were studied based on four models, namely Neyman-Scott Rectangular pulse (NSRP) process, Microcanonical Multiplicative Random Cascade (MMRC) process and its variant MMRC-K, and one Deep-Learning based process (ANN-K), using metrics like dry periods, event rainfall volumes, extreme rainfall characteristics, etc. New hydrological insights for the regionThe study successfully established individual rainfall volume, event rainfall volume, and event durations are all within 10% of each other for the four stations. Delhi due to its continental climate showed a higher percentage of dry periods and longer dry periods, which were most successfully modelled by MMRC and MMRC-K. Kolkata and Mumbai stations displayed a higher number of extremely intense and cloudburst types of rainfall, which were modeled effectively by the Deep-Learning Model. Chennai has a different rainfall pattern due to returning monsoon which was also captured by the models. Generally, for extreme rainfall parameters, the ANN-K model performs significantly better, successfully reproducing the characteristics at all quantiles, especially with rainfall above 100 mm/hour intensity or cloud bursts, while 50% of the models overestimated these. NSRP on the other hand performs reasonably well for most considered parameters, without being exceptional at any of them. MMRC and MMRC-K most accurately modeled the dry period parameters.