Replication Data and Code for: Typhoon Cloud Morphology as an Early Satellite Proxy for Storm-Surge Risk under ERA5 Physical Benchmarking

Overview This repository contains the core scripts, input datasets, and model output tables required to reproduce the data processing, feature extraction, predictive modeling, and robustness analyses presented in the manuscript. To comply with open science and file size constraints, this package focuses on the data-driven and machine-learning pipelines. It excludes raw high-resolution satellite imagery (Himawari-8/9) and local hydrodynamic simulation outputs (e.g., ADCIRC), which are either hosted on official meteorological platforms or discussed separately in the manuscript. Additional documentation: `DATA_SOURCES.md` documents the original data providers, derivative tables included in this archive, and data excluded from redistribution. `ENVIRONMENT.md` provides environment setup instructions and the recommended script execution order. `requirements.txt` lists the Python packages needed to run the analysis scripts. Directory Structure To ensure a clean and reproducible environment, the repository is organized into three primary directories: `01_Scripts/` Contains all Python scripts used for reproducing the study. Data processing and Himawari B13 cloud feature extraction. ERA5 physical variables extraction and alignment. Model training (Cloud-Sequence Transformer, Ridge Regression, HistGradientBoosting baselines). Robustness checks, external validation, and interpretability analyses. `02_Input_Data/` Contains the core analysis-ready tables. Hong Kong Observatory (HKO) storm surge event and station tables. Extracted objective cloud morphology metrics. Integrated model-ready tables joining surge targets, cloud features, and physical proxies. `03_Results_and_Tables/` Contains the comprehensive outputs generated by the scripts. Transformer and ERA5 baseline predictions. Performance metrics (RMSE, AUC, Precision-Recall). Interpretability permutation results and robustness check tables (leave-one-out, threshold sensitivity).