Official source code for paper 《TFG-Net: Tropical Cyclone Intensity Estimation from a Fine-grained Perspective with the Graph Convolution Neural Network》
The implementation of the FTFE module is referred to this repository. Many thanks to the contributor @jeong-tae
conda env create -f TI_Estimation.yaml
- Download the required GridSat dataset from NOAA official site through here and the required tropical cyclone best track dataset from NOAA official site through here.
Or you can download the preprocessing GridSat data from my google drive through here. Note that the ibtracs tropical cyclone best track dataset is provided in folder data.- Or you can download the preprocessing GridSat data from my baidu drive through here. Note that the ibtracs tropical cyclone best track dataset is provided in folder data.
We provide one of the five runs best-validated models in here. You can reproduce the result reported in the paper using this best-validated model.
-- data # dataset folder
-- GridSat_B1_new_npy # the GridSat data folder. You need to download it from google driver
-- gridsat.img.min.max.npy # the min and max value of the training GridSat dataset
-- gridsat.path.ibtr.windspeed.csv # the label GridSat file with satellite images save path
-- GridSat_B1_processor.ipynb # the orignal nc GridSat file processor
-- GridSAT_invalid_img.ipynb # the invalid GridSat preprocessor
-- ibtracs.ALL.list.v04r00.rar # the rar compression file of the IBTracs tropical cyclone best track dataset
-- figure # figure provider
-- network.png # architecture of TFG-Net model
-- TFG-Net_training@2x.gif # the training stage fine-grained tropical feature captures with 20 epochs interval
-- layers # necessary layer
-- AttentionCrop.py # the Attention Cropper
-- GraphConvolution.py # the Graph Convolution
-- MultiHeadGAT.py # Multi Head GAT
-- model_saver # model save path
-- best_validate_model.pth # best model (one of five runs). You need to download it from google driver
-- TFG-Net.log # the training loss of the TFG-Net
TI_Estimation.yaml # conda environment for the project
TFG-Net.ipynb # jupyter visualized code for the TFG-Net
When the conda environment and datasets are ready, you can train or reproduce our result by running the file TFG-Net.ipynb.
If you find this code or idea useful, please cite our work:
@inproceedings{xu2023TFG,
title={TFG-Net: Tropical Cyclone Intensity Estimation from a Fine-grained Perspective with the Graph Convolution Neural Network},
author = {Guangning Xu and Yan Li and Chi Ma and Xutao Li and Yunming Ye and Qingquan Lin and Zhichao Huang and Shidong Chen},
journal = {Engineering Applications of Artificial Intelligence},
volume = {118},
pages = {105673},
year = {2023},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2022.105673},
}