Automatic fronts detection using a convolutional neural ...
Transcript of Automatic fronts detection using a convolutional neural ...
Training and tuning of hyper-parameters
3 classes : no front, warm front and cold front60 epochs, 1340 images of 200x420 pixels splitted in 107 200 images of 48x48 pixelsLoss function : weigthed categorical cross-entropy, Weigths : 1 for "no front" class and 10 for "warm front" and "cold front" classes
Ground truth : surface charts from Météo-France produced 4 times a day
Data : 1 year of analyses from Arpège (French global model) with a resolution of 0.1°
Testing dataset : from 01/02/2020 to 29/02/2020, 116 images
Training and validation dataset :
1340 images (2/3 for training, 1/3 for validation) from 01/07/2019 to 31/01/2020 and from 01/03/2020 to 01/07/2020
Datasets
Ground truth
Results
Surface chart 24/02/2020 00 UTC
extracted from surface chart + θ' w at 850hPa
Prediction CNN + θ' w at 850hPa
Issues Surface charts are drawn using information different from Arpège analyses
(other model analyses, observations, etc...).Data formats changed in 2019. Other preprocessing routines must be developed
to use the previous data.Imbalanced dataset : 80 "no front" pixels for 1 "cold front" or "warm front" pixel Quality of the prediction very sensitive to the activation function.
32 32
64 64
128 128
64 64
32 32
16 161 16 16
48x4
8
24x2
4
12x1
2
6x6
N_cl N_cl
conv 3x3, ReLu, drop out 20%copy and concatenatemax pool 2x2upsample 2x2conv 1x1, ReLusoftmax
U-Net architecture
Adapted from Ronneberger et al, 2015
Automatic fronts detection using a convolutional neural network
Lucie ROTTNER, Laure RAYNAUD, Arnaud MOUNIER, Matthieu PLU
have been used to detect weather fronts (Biard and Kunkel, 2019, Lagerquist et al, 2019, Matsukoa et al, 2019). Here we present a front dectection First attempts at automatic drawing of fronts and other synoptic-scale features were based on physical rules. Recently, convolutional neural networks
method for the French operationnal global model, Arpège. A U-Net is used to detect cold and warm fronts over the western Atlantic.
ECMWF-ESA Workshop on Machine Learning for Earth System Observation and Prediction, 5-8 octobre 2020
References
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net : Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention. Springer, 2015
James C Biard and Kenneth E Kunkel. Automated detection of weather fronts using a deep learning neural network. Advances in Statistical Climatology, Meteorology and Oceanography, 5(2) :147–160, 2019.Ryan Lagerquist, Amy McGovern, and David John Gagne II. Deep learning for spatially explicit prediction of synoptic-scale fronts. Weather and Forecasting, 34(4) :1137–1160, 2019.Daisuke Matsuoka, Shiori Sugimoto, Yujin Nakagawa, Shintaro Kawahara, Fumiaki Araki, Yosuke Onoue, Masaaki Iiyama, and Koji Koyamada. Automatic detection of stationary fronts around japan using a deep convolutional neural network. SOLA, 2019.
Centre National de Recherches Météorologiques, UMR 3589 CNRS/Météo-France, Toulouse, [email protected]
Data pre-processing
Front extraction from surface charts :
Geographical domain : 32°W-10°E and 40°N-60°N
extraction of front coordinates from json file interpolation on the ground truth grid
Data normalisation : normalisation min-max
Meteorological parameters
Mean sea level pressure
Pseudo-adiabatic potential temperature at 850hPa (θ' w)
10m wind (u and v components)
Absolute vorticity at 850hPa and 700 hPa
Relative humidity at 700 hPa
To be done
Evaluation by forecasters
Improving visualisation :
Adding front classes : occluded fronts,
Using longer datasets
quasi-stationnary fronts etc...
N_cl = 3 classes