An Interpretable Deep Learning Model for Detecting Tropical Cyclone Genesis Through Incorporation of the Dynamic and Thermodynamic Environment
In one sentence. Interpretable 3D-residual GRU model detects tropical cloud clusters that will undergo genesis, reaching Heidke skill 0.80 and 94% detection; SHAP shows dynamics dominate.
Abstract
Abstract In recent decades, tropical cyclone (TC) track forecasting has improved significantly. However, prediction of TC genesis (TCG) remains challenging. Accurate prediction of TCG is critical for disaster prevention and mitigation. This study develops an interpretable deep learning model that integrates 3D‐residual blocks with gated recurrent units to detect the probability of tropical cloud clusters (TCCs) undergoing TCG using a global data set of tropical convective clusters spanning from 1982 to 2018. The model utilizes dynamic and thermodynamic predictors, including relative vorticity, the dynamic genesis potential index, zonal wind, meridional wind, vertical velocity, atmospheric temperature, mid‐level moisture and infrared brightness temperature. This model shows considerable skill (Heidke Skill Score = 0.8016, False Alarm Rate = 8%, Probability of Detection = 94%). SHapley Additive exPlanations analysis highlights that dynamical environmental factors are the primary predictors, with a secondary contribution from thermodynamical predictors and a lesser role from convective development. This study highlights a skillful tool for the detection of TCG from precursor TCCs and underscores the significant role of the dynamic environment in TCG.
Labels
- Topic
- Extremes & hazards, Explainability, uncertainty & physics
- Type
- New model or method
- Method
- CNN / U-Net, RNN / LSTM / ConvLSTM
- Time scale
- 6 h–3 days
- Domain
- Global
- Data
- ERA5 / reanalysis, Satellite
- Status
- Research
- Identifiers
- DOI:10.1029/2025jh001152 OpenAlex:W7164058218
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.