A statistical model of North Indian Ocean tropical cyclone genesis, tracks and landfall

Extensive damage and loss of life can be caused by landfalling tropical cyclones (TCs). Seasonal forecasting of TC landfall probabilities is potentially beneficial to insurance/re-insurance companies, decision makers, government policy and planning departments, and residents in coastal areas. In this study, climatological and statistical seasonal forecast models are developed for TC genesis, tracks and landfall for North Indian Ocean (NIO) rim countries based on kernel density estimation, a generalised additive model (GAM) including an Euler integration step, and landfall detection using a country mask approach. To forecast TC activity in the NIO region, the relative roles of climate modes (stratospheric Quasi-Biennial Oscillation (QBO), El Ni√±o ‚- Southern Oscillation (ENSO), Indian Ocean Dipole (IOD)) as predictor variables in the modelling schemes have been investigated. Using a 35-year record (1979-2013) of tropical cyclone track observations from the Joint Typhoon Warning Centre (part of the International Best Track Archive Climate Stewardship Version 6), the distribution of cyclone genesis points is approximated by kernel density estimation and the observed cyclone tracks are fitted using the GAM as smooth functions of location in each season. The model simulated TCs are randomly selected from the fitted kernel (TC genesis) and cyclonic paths, with random innovations, are simulated (TC tracks) to generate a suite of landfall statistics. Lead-lag analysis is undertaken to assess the utility of various climate mode predictor timescales for TC forecast potential. Three hindcast validation methods are applied to evaluate the integrity of the models. First, leaveone-out cross-validation is applied whereby the country of landfall is determined by the majority vote (considering the location by only the highest percentage of landfall) from the simulated tracks. Second, the probability distribution of simulated landfall is evaluated against the observed landfall. Third, the distances between the point of observed landfall and simulated landfall are compared and quantified. Overall, the models show very good cross-validated hindcast skill of modelled landfalling cyclones against observations for most NIO rim countries, with only a relatively small difference in the percentage of predicted landfall locations compared with observations. Finally, the developed models demonstrate that including information on the phase of the QBO (using the stratospheric QBO index) and ENSO (using the Southern Oscillation index (SOI)) can improve the skill of seasonal forecasts of TCs in the NIO region. It is shown that the most skilful model (the model skill is assessed based on predictor leads of from 1-6 months) for ENSO as predictor is found using a three-month-averaged SOI with twomonth lead ahead of each of the four NIO region designated TC seasons. Analogously, the threemonth averaged QBO is found to be potentially most skilful at the three-month lead. Both models demonstrate clear improvements over climatology. The hindcast probabilities and distribution of TC landfall occurrences using QBO and ENSO as climate predictor modes, match remarkably well against observations over most of the study domain. Finally, in a separate analysis, a Poisson regression model using the Markov Chain Monte Carlo method was also developed to forecast TC landfall probabilities in the NIO region. A combined three-predictor (sea surface temperature, ocean heat content, and SOI) model was found to perform best in this model formulation against climatology.