Scholarlink

<p>We developed a Long Short-Term Memory (LSTM) ensemble modeling framework to examine the influence of the Indian Summer Monsoon (ISM) on post–monsoon TC genesis over the Bay of Bengal (BoB). To incorporate large-scale climate modes, the effects of the El Niño–Southern Oscillation (ENSO) and Madden–Julian Oscillation (MJO) are explicitly included together with key ocean–atmosphere predictors. To uncover the robustness of the model, two experiments were conducted across different periods, the first experiment utilised the data for the entire period (1977–2022) and the second experiment used recent epochs (1993–2022). The validation metrics indicate good accuracy and predictive skill, with the top models explaining up to 82% of the variance during the recent epoch. In experiment 1, the model–agnostic interpretability tests revealed that the dominant features were low-level vorticity, relative humidity and the Niño–4 index. In experiment 2, vorticity, the ISM rainfall index and vertical wind shear emerged as the largest contributors to model predictions. Low–level vorticity is the dominant feature in both experiments, while the lagged effect of ISM on TC genesis prediction increased during the recent period (1993–2022). The role of relative humidity is reduced in the second experiment and MJO shows a limited impact on the model predictions in both experiments. A contrasting influence of ENSO on TC genesis prediction is observed in SHAP Analysis, highlighting the importance of incorporating nonlinear relationships. This understanding of the non-linear characteristics of significant predictors of TC genesis is valuable for improving TC forecast systems.</p> Graphical Abstract <p></p> <p>Post–monsoon tropical cyclone (TC) genesis over the Bay of Bengal (BoB) is intimately linked to the environmental conditions established during the monsoon–driven circulation. This study developed a Long Short–Term Memory (LSTM) ensemble framework to model the nonlinear interactions among the key predictors during the Indian summer monsoon (ISM) and the subsequent post-monsoon seasons over BoB. Two experiments were conducted to evaluate the robustness and temporal consistency of the model for two time periods 1977–2022 and the recent period, 1993–2022. Explainable artificial-intelligence diagnostics revealed that low–level vorticity is the dominant controlling factor in post-monsoon TC genesis. SHapley Additive exPlanations showed a 12.4% and 7.4% increase in the relative contributions of monsoon rainfall variability and vertical wind shear to TC genesis in recent epochs (1993–2022), compared with the entire period (1977–2022), indicating an increasing influence of monsoon–driven circulations on BoB TC formation. The nonlinear influence of relative humidity and El Niño–Southern Oscillation on TC genesis is also revealed in this study. The findings from the LSTM ensemble modeling framework improve the understanding of how the ISM influences subsequent TC genesis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Decoding the Non-linear Impact of Monsoon on Cyclogenesis in the Bay of Bengal by Applying Deep Learning

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • R. Rahul,
  • Arun Chakraborty,
  • J. Kuttippurath,
  • Mansour Almazroui,
  • Khlood Ghalib Alrasheedi,
  • Gargi V. Pillai

摘要

We developed a Long Short-Term Memory (LSTM) ensemble modeling framework to examine the influence of the Indian Summer Monsoon (ISM) on post–monsoon TC genesis over the Bay of Bengal (BoB). To incorporate large-scale climate modes, the effects of the El Niño–Southern Oscillation (ENSO) and Madden–Julian Oscillation (MJO) are explicitly included together with key ocean–atmosphere predictors. To uncover the robustness of the model, two experiments were conducted across different periods, the first experiment utilised the data for the entire period (1977–2022) and the second experiment used recent epochs (1993–2022). The validation metrics indicate good accuracy and predictive skill, with the top models explaining up to 82% of the variance during the recent epoch. In experiment 1, the model–agnostic interpretability tests revealed that the dominant features were low-level vorticity, relative humidity and the Niño–4 index. In experiment 2, vorticity, the ISM rainfall index and vertical wind shear emerged as the largest contributors to model predictions. Low–level vorticity is the dominant feature in both experiments, while the lagged effect of ISM on TC genesis prediction increased during the recent period (1993–2022). The role of relative humidity is reduced in the second experiment and MJO shows a limited impact on the model predictions in both experiments. A contrasting influence of ENSO on TC genesis prediction is observed in SHAP Analysis, highlighting the importance of incorporating nonlinear relationships. This understanding of the non-linear characteristics of significant predictors of TC genesis is valuable for improving TC forecast systems.

Graphical Abstract

Post–monsoon tropical cyclone (TC) genesis over the Bay of Bengal (BoB) is intimately linked to the environmental conditions established during the monsoon–driven circulation. This study developed a Long Short–Term Memory (LSTM) ensemble framework to model the nonlinear interactions among the key predictors during the Indian summer monsoon (ISM) and the subsequent post-monsoon seasons over BoB. Two experiments were conducted to evaluate the robustness and temporal consistency of the model for two time periods 1977–2022 and the recent period, 1993–2022. Explainable artificial-intelligence diagnostics revealed that low–level vorticity is the dominant controlling factor in post-monsoon TC genesis. SHapley Additive exPlanations showed a 12.4% and 7.4% increase in the relative contributions of monsoon rainfall variability and vertical wind shear to TC genesis in recent epochs (1993–2022), compared with the entire period (1977–2022), indicating an increasing influence of monsoon–driven circulations on BoB TC formation. The nonlinear influence of relative humidity and El Niño–Southern Oscillation on TC genesis is also revealed in this study. The findings from the LSTM ensemble modeling framework improve the understanding of how the ISM influences subsequent TC genesis.