Extreme rainfall hotspots in India based on spatio-temporal variability of rainfall using unsupervised clustering techniques

In past few decades there has been a noticeable increase in the frequency and intensity of extreme rainfall events (EREs) globally, including India. The Clausius-Clapeyron relationship explains how the warmer air can significantly hold more moisture. Hence, in present climate change scenario increasing temperature along with other factors can lead to further increase in EREs. Effective management strategeis in various sectors like disaster preparedness, smart-city planning, water quality, public health, agriculture planning, etc. can get improved, through proper understanding on the distribution and frequency of EREs. Keeping in mind the socio-economic impacts of EREs; this study aimed to identify the hotspot regions for EREs in India.

India is a country with vast spatio-temporal variability in rainfall pattern. Hence, this study implemented objective criteria to identify the spatio-temporal rainfall variability of EREs over four rainfall homogeneous regions for pre-monsoon, monsoon and post-monsoon seasons. Based on frequency distribution of daily accumulated rainfall, suitable rainfall threshold values for defining EREs are identified for each homogeneous region and each season. These threshold values vary region-wise as well as season-wise. Distribution of EREs show interannual as well as seasonal variability.

Clustering algorithms, popular unsupervised Machine Learning (ML) techniques, are handy tools to identify hotspots of extreme rainfall regions with similar spatial variability. To understand the ERE distribution and to identify rainfall hotspots based on long term daily gridded rainfall data, this study implemented K-means clustering and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithms. Comparative area distribution study between K-means and DBSCAN clustering help to identify the EREs hotspots in India. Overall, the K-means method shows more scattered hotspots compared to DBSCAN method, which are further validated using Davies-Boulding Index (DBI), Silhouette score, Calinski-Harabasz (CH) score and Dunn's Index. These score analysis methods serve as potential tools to support the clustering validation method. In addition to the area distribution, this study has addressed the temporal variability of the EREs hotspots. ST-OPTICS ( Spatio-Temporal Ordering Points to Identify the Clustering Structure) algorithm results clustering of hotspots based on their spatial and temporal similarity. This study shows that ML algorithms prove to be promising techniques for detecting and analyzing spatial as well as temporal variability of EREs hotspots which is effective for future management practice in various sectors.

Keywords: Extreme Rainfall Events; DBSCAN Clustering; K-Means Clustering; ST-OPTICS.