How ocean thermal structure and atmospheric steering govern typhoon evolution: evidence from interpretable machine learning
TL;DR
This study develops a tree-based framework for simultaneous multi-step prediction of typhoon trajectory and intensity in Western North Pacific (WNP), and applies SHapley additive explanations (SHAP) to quantify the contribution of individual environmental drivers to each forecast.
Abstract
Reliable typhoon forecasting demands both predictive accuracy and physical transparency. This study develops a tree-based framework for simultaneous multi-step prediction of typhoon trajectory and intensity in Western North Pacific (WNP), and applies SHapley additive explanations (SHAP) to quantify the contribution of individual environmental drivers to each forecast. The framework integrates Japan Meteorological Agency Best Track data (1977–2024) with ERA5 and ORAS5 reanalyses, incorporating mean potential temperature over 0–100 m ( T100), a predictor largely overlooked in favour of sea surface temperature, together with upper ocean heat content (OHC) and thermocline depth, within a 780-dimensional 72 h lagged feature space. Among three tree-based architectures evaluated under identical conditions, XGBoost performs best across all targets and horizons: at +24 h on the independent 2020–2024 test set it attains RMSE of 1.90∘ and 2.92∘ for latitude and longitude and 11.9 hPa and 16.2 kt for central pressure (CP) and maximum wind speed (MW). This has effect of reducing RMSE by about 45% (track) and 24%–25% (intensity) relative to persistence, with bootstrap 95% confidence intervals excluding zero in every case, and by 9% and 13% relative to random forest for CP and MW. elbow-based feature selection shows trajectory converges on a compact predictor subset whereas intensity demands a richer feature space. SHAP analysis reveals a physically coherent horizon-dependent shift: at +6 h kinematic persistence governs both track and intensity, while at +24 h large-scale steering flow dominates track displacement and ocean–atmosphere thermodynamic forcing progressively takes over intensity. T100 exhibits a sharp nonlinear threshold near 26 ∘C that is robust across the test period and encodes upper OHC dynamics without explicit parameterisation. By coupling statistically significant forecast skill with transparent, physics-consistent attribution, the framework offers a physically interpretable complement to black-box systems for typhoon guidance in the WNP.