deep_typhoon's Introduction

deep_typhoon

Analysis satellite images of typhoons by deep-learning (CNN), based on PyTorch.

This CNN learns to map the satellite images of typhoons to their max wind speed from. The labeled train set is obtained from agora/JMA.

Online analysis: http://www.predictor.xin/typhoon/

Requirements

BeautifulSoup
PIL
Pytorch

Usage

Run download_agora.py, download the satellite images of typhoons to folder tys_raw.
Run create_samples.py, convert raw data into the legal samples for our CNN, create two new forlder train_set and test_set.
Train CNN using train_net.py, the trained CNN will be saved as a disk file net_relu.pth.
Run test_net.py, analysis the test set.

After 10 epoches training the CNN regressor reached mean loss about 8 (knots) in train set and about 10 (knots) in test set.

Here is what this CNN thinks of the top 20 typhoons sorted by max wind.

1 ('197920', 130.27679443359375)  
2 ('200914', 127.7662582397461)  
3 ('199019', 122.92172241210938)  
4 ('200918', 122.84004211425781)  
5 ('201614', 122.66597747802734)  
6 ('201601', 122.03250885009766)  
7 ('201513', 121.75947570800781)  
8 ('200922', 121.35771942138672)  
9 ('201013', 120.0194091796875)  
10 ('201330', 118.92587280273438)  
11 ('201419', 117.6025390625)  
12 ('198305', 117.10270690917969)  
13 ('201422', 116.77259063720703)  
14 ('198522', 116.46116638183594)  
15 ('201327', 116.42304992675781)  
16 ('201216', 116.36921691894531)  
17 ('198221', 116.18096923828125)  
18 ('199230', 115.96656799316406)  
19 ('198210', 115.96611022949219)  
20 ('201328', 115.57132720947266)

Tips

Memory should be at least 1.5G .
This project is written without cuda(), while you can use cuda() to transfer the CNN onto GPU and speedup the training.
The images and labels are crawled from agora.ax.nii.ac.jp/digital-typhoon , and the labels are refered to JMA(Japan Meteorological agency).