High accuracy Philippine landfall prediction of Pacific cyclones at their genesis using neural networks

Abstract

More than twenty (28±6) tropical cyclones (TC) are formed in the West Pacific basin annually with 18±4 of them entering the Philippine Area of Responsibility (PAR), and 10±4 eventually making landfall in the archipelago. Predicting accurately at the earliest possible time if a TC will strike land is vital in disaster preparedness and mitigation planning. We develop a supervised backpropagation neural network (NN) that predicts if a TC will eventually make landfall right after its genesis. The three-layer NN has six inputs (two TC eye coordinates, maximum sustained wind speed at genesis, and distances from three PAGASA weather stations), seven hidden nodes and two outputs (yes or no). It is able to predict if a TC enters PAR twelve hours after genesis, and hits land within fifty-four hours at a success rate of 94.82 ± 1.89% and 91.08 ± 0.89%, respectively. Network performance is evaluated at different lead times for different numbers of input and hidden nodes and pairs of training and test sets. The data sets are taken from the Joint Typhoon Warning Center Best Track archives (1945-2017).