profile

Dataset Description This repository contains the open-source data, trained models, and visualization scripts accompanying the manuscript: Detection of Eye Occurrence in Sequential Satellite Infrared Imagery and Its Application to Improve Deep Learning-based Tropical Cyclone Intensity Estimation Yi Liu¹², Jing-Yi Zhuo³, Kekuan Chu¹²*, Zhe-Min Tan¹²¹ Key Laboratory of Mesoscale Severe Weather Meteorological Science and Technology, Nanjing University² School of Atmospheric Sciences, Nanjing University³ High Meadows Environmental Institute, Princeton University The dataset provides all materials needed to reproduce the analyses, figures, and case studies in the paper, including preprocessed satellite infrared (IR) imagery, model predictions, and inference-ready model files. 1. Data 1.1 CSV files (data/csv/) This directory contains the prediction results for all models discussed in the paper. Each file includes the matched IBTrACS labels and the model-predicted intensity and eye probability for the entire test set. CNTL_testset_labels_and_preds.csv — Baseline model with only IR as inputEPI6 / EPI12 / EPI18 / EPI24 — Models enhanced with EPI at different time windowsEPI24_all0 / EPI24_all1 — Sensitivity experiments in which the EPI24 is fixed to 0 or 1EPI24_TCcases_labels_and_preds.csv — Four super TC cases used in Fig. 6fig2_labels.csv & fig2_eyeprob_preds.csv — Supporting data for Fig. 2 Each file contains: * Metadata (time, location, storm ID, etc.)* Best-track intensity (USA_WIND)* Model-predicted intensity and/or eye probability 1.2 IR image patches (data/IR/) This folder contains the IR image sequences used for model inference and for figure generation. fig2_ir_patches.npy — 30 samples used to illustrate eye/no-eye patterns in Fig. 2IR_CHANTHU_lifecycle_raw/normalized — 301×301 and 64×64 IR sequences covering Typhoon Chanthu (2021), used in the case-study demonstration (Fig. 6) All IR images are extracted from the HURSAT-B1 dataset and preprocessed following the pipeline described in the manuscript. 2. Models (model/) Two inference-ready models are provided as serialized pickle files, enabling users to run predictions directly: DeepTCEye.pkl — Eye-probability detection networkDeepTCNet_addphyEPI24TCF.pkl — Intensity estimation model enhanced with eye-presence information (24-h context) Both models can be directly loaded with the accompanying script ‘scripts/model_infer.py’. 3. Visualization Scripts (scripts/) Reproducible scripts for generating all paper figures: fig2.py — code to generate Fig2 in the manuscriptfig3.py — code to generate Fig3 in the manuscriptfig4.py — code to generate Fig4 in the manuscriptfig5.py — code to generate Fig5 in the manuscriptfig6abcd.py, fig6e.py — code to generate Fig6 in the manuscriptmodel_infer.py — Example showing how to load the pkl models and run inference Each script corresponds precisely to the figures in the main manuscript. 4. Figures (pics/) The folder contains the high-resolution versions of all figures (Fig. 2–6) generated by the scripts above. These figures are included for reference and verification. 5. Citation If you use this dataset, models, or scripts, please cite cite the following references: Zhuo, J.-Y., & Tan, Z.-M. (2021). Physics-augmented deep learning to improve tropical cyclone intensity and size estimation from satellite imagery. Monthly Weather Review, 149(7), 2097-2113, https://doi.org/10.1175/MWR-D-20-0333.1 Liu, Y., Zhuo, J.-Y., Chu, K., & Tan, Z.-M. (2025). Detection of eye occurrence in sequential satellite infrared imagery and its application to improve deep learning-based tropical cyclone intensity estimation. Manuscript submitted for publication to Journal of Geophysical Research: Machine Learning and Computation.