Applying Weighted Salinity Stratification to Rapid Intensification Prediction of Tropical Cyclone With Machine Learning

Abstract Tropical cyclone (TC) intensification is influenced by environmental conditions, inner‐core dynamics, and interactions with upper‐ocean layers. Rapid intensification (RI) is a significant threat that is difficult to predict, prompting multiple institutions to collaborate. However, the accuracy still needs further improvements. It is well‐known that a warm upper ocean is conducive to RI, but the role of salinity stratification in this process is not well understood, particularly under different TC translation speeds. This study reveals that rapidly intensifying TCs are related to large salinity stratification, especially when TC moves slowly. To develop a predictive model, several machine learning (ML) algorithms are used, with the most appropriate parameters and weights for each algorithm being determined. Our final ML model, which incorporates salinity stratification as a predictor and TC translation speed as a weight parameter, demonstrates superior performance across various predictive metrics, including the probability of detection (POD), false alarm ratio (FAR), and Peirce Skill Score (PSS) over the Western North Pacific during 2004–2022 compared to the model without these two factors. The most significant enhancement is observed for intense RI episodes. The improvements are up to 14% for both in POD; 7% and 13% in FAR; and 19% and 16% in PSS for 12.75 and 15.3 m s−1 RI thresholds, respectively. These results highlight the importance of including salinity stratification as a new predictor and TC translation speed as a weighted parameter using ML techniques in RI prediction models.