This experiment consists in estimating the class of a Tropical Cyclone. Classes are defined based on intensity level, which is provided by meteorological agencies based on typhoon wind speed and centre pressure measurements.
Our model achieved nearly 57.83% accuracy on the test set. For more details refer to the examples in notebooks.
We also tested the performance of our model using a random split distribution between training and test sets. We observe that using completely random distribution lead to an accuracy of nearly 82.82% on the test set. However, we claim this is not a good practice. More details may be found in Lucas Rodés-Guirao thesis and in the notebook Random_Data_Split.
- Images must be in range of [0, 255], where 0 and 255 correspond to 160 Kelvin and 310 Kelvin, correspondingly.
- The model accepts 128x128 images with resolution 1 pixel ≈ 10 Km. To this end we crop 128x128 regions from resized Digital Typhoon 256x256 images (original images come as 512x512).
- Images are assumed to have the typhoon eye in the image centre (i.e . at position [63, 63]).
You can also use the model in your code.
from pyphoon.models.tc_multiclass import tcNet
model = tcNet('weights.hdf5')from pyphoon.models.tc_multiclass import tcPreprocessor
X = ... # Load (1, 256, 256) image or (N, 256, 256) array of images
X = tcPreprocessor().apply(X)Make sure to crop the images so as to take a centred square of dimension 128x128.
X = X[:, 64:64+128, 64:64+128, :]
y_pred = model.predict(X)