An Alternative Similar Tropical Cyclone Identification Algorithm for Statistical TC Rainfall Prediction in the Western North Pacific

Improving tropical cyclone (TC) rainfall prediction is vital as climate change has led to an increase in TC rainfall rates. Enhanced reliability in predicting TC tracks has paved the way for statistical methodologies to make use of them in estimating current TC rainfall, achieved by identifying similar past TC tracks and obtaining their corresponding rainfall data. While the Fuzzy C Means (FCM) clustering algorithm is widely used, it has limitations stemming from its clustering-centric design, hindering its ability to pinpoint the most appropriate similar TCs. Our study introduces the Sinkhorn distance, a novel similarity metric that measures the cost of transforming one set of data to another, for assessing TC similarity in rainfall prediction. Our findings indicate that utilizing Sinkhorn distance enhances the accuracy of TC rainfall predictions across the Western North Pacific region. When compared to the conventional approach using FCM, our Sinkhorn distance-based methodology yields slightly better yet statistically significant results. The improvement is due to better identification of similar TCs, characterized by closer proximity of similar TC tracks to the target TC track, facilitated by Sinkhorn distance. This underscores how minor differences in TC track can alter rainfall distribution, emphasizing the critical importance of accurate track prediction in rainfall prediction and the need to reconsider how we categorize TCs together, which can have implications for climate and atmospheric sciences. Collectively, the inclusion of Sinkhorn distance stands as a valuable addition to our toolkit for discerning similar TC tracks, thus elevating the accuracy of TC rainfall predictions.