/ConvLSTM-CNN-for-tropical-cyclone

Images timeseries sequence with ConvLSTM for windspeed prediction & CNN cyclone intensity

Primary LanguageJupyter NotebookMIT LicenseMIT

ConvLSTM-CNN-for-tropical-cyclone-prediction

Table of contents

Quick start

This project was made for 2022 NTU Remote Sensing & Geospatial Information Analysis And Application.
There are two sections for in this project, ConvLSTM for windspeed time-series prediction and CNN for cyclone intensity prediction.
Give a ⭐ if you think this project is helpful😄
❗ You need GPU for this project, especially for ConvLSTM ❗

  • Section 1: ConvLSTM ---> See .ipynb & colab(link)   in Section 1 folder
  • Section 1: CNN ---> See .ipynb   in Section 2 folder
  • 👉 V100 32G & RTX 2080ti were used for ConvLSTM ---> Reduce batch size first if OOM occurs, also try simplifying the network structure
  • 👉 GTX 950M were used for CNN
  • Project Flow Chart as below 👇 👇
    flow

Status

Status Status Status Status Status Status

What's includeds

Project excecute order

Section 1 : Download_gfs.ipynb -> Generate_images_sequence.ipynb -> colab example / train.py -> make_gif.py
Section 2 : Inspect_track_data.ipynb -> Download_HURSAT.py -> Process.py -> train.py -> view_pred_images.py

Project Notice

Section 1 : images.npy (2022/01 - 2022/05) --> Smaller dataset --> prepocess contains in those files
train.py -1 images_all.npy (2021/05 - 2022/05) --> modified train.py at line 17 & 18 --> change data[:-3] to data[:-4]
train.py -2 images_all.npy (2021/05 - 2022/05) --> also change the batchsize to 619 in np.reshape() in line 23 & 24

Project File

Section 1/
└── windspeed_timeseries/
    ├── code/
    │   ├── train.py
    |   ├── make_gif.py
    |   ├── colab_train_link.txt
    |   ├── Generate_images_sequence.ipynb (provide link since > 30 Mb)
    |   └── Download_gfs.ipynb
    └── dataset/
        └──link for images.npy (2022/01 - 2022/05) & images_all.npy (2021/05 - 2022/05)
              
Section 2/
└── cyclone_intensity/
    ├── code/
    │   ├── Download_HURSAT.py
    │   |── Process.py
    |   |── train.py
    |   |── view_pred_images.py
    |   └── Inspect_track_data.ipynb
    └── dataset/
        └── link for images.npy & labels.npy &  5 fold prediction result

Results

Up: ConvLSTM predicts 2 in 5 frame // Down: CNN predition examples ConvLSTM_Result intensity

Bugs and feature requests

Have a bug or a feature request? Please first search for existing and closed issues.
If your problem or idea is not addressed yet, please open a new issue.

Creator

GMfatcat

Thanks

1.https://www.kaggle.com/code/kcostya/convlstm-convolutional-lstm-network-tutorial
2.https://www.kaggle.com/code/concyclics/analysis-typhoon-size/notebook
3.https://github.com/23ccozad/hurricane-wind-speed-cnn
4.https://www.ncree.narl.org.tw/home for High Performance Computing System (ConvLSTM)

Copyright and license

Code released under the MIT License.

Enjoy 🤘