Detection of Bow Echoes in Kilometer-Scale Forecasts Using a Convolutional Neural Network
Source: American Meteorological Society
In this study, researchers Arnaud Mounier, Laure Raynaud, Lucie Rottner, Matthieu Plu, Philippe Arbogast, Michaël Kreitz, Léo Mignan and Benoît Touzé describe how their convolutional neural network (CNN) model was used to predict thunderstorm intensity.
To train their model (CNN), the researchers proceeded in 6 steps:
Step 1:
Patches from the pseudoreflectivity fields are randomly selected, and Npatches are extracted from each field (values in Table 1). The associated ground truth patches are also extracted.
Step 2:
The pairs of patches (pseudoreflectivity/ground truth) are split into two groups: the patches with BEs in which at least one grid point is labeled as a BE and the patches without any BE.
Step 3:
The number of patches with BEs is very limited compared to patches without any BE (1 patch with a BE for every 1600 patches without a BE). Initially, BEs with moderate pseudoreflectivities were not correctly predicted because they were underrepresented in the dataset. A data augmentation technique is proposed to solve this problem: in a given patch, the pseudoreflectivities are multiplied by a coefficient of 0.75 if a BE is within this patch and the maximum pseudoreflectivity is above a given threshold (mentioned in the next section). This new patch is added in the ones with BEs (step 2). The pseudoreflectivity field must remain physically consistent as much as possible. That is why lower coefficients are not investigated and only patches with large magnitude in pseudoreflectivities are taken into account.
Step 4:
To reduce the number of patches without a BE, the patches without precipitation (pseudoreflectivity maximum < 0.1 mm h−1) are deleted.
Step 5:
Even after the filtering procedure of step 4, the number of patches without a BE remains high (1 patch with a BE for every 400 patches without a BE). To limit the number of patches without a BE, the ratio between the patches with and without a BE (Ratio_noBE/BE) is fixed to a lower value. The patches retained are randomly selected and the unnecessary patches without a BE are deleted. This ratio is discussed in the next section.
Step 6:
During the first tests, all patterns with strong pseudoreflectivities were detected as BEs and consequently the number of false alarms was very high in the validation database. Strong pseudoreflectivities are rare in space and time and the majority of patches without a BE contains no or weak precipitation, whereas BEs are frequently associated with heavy precipitation. Only 0.7% of patches without a BE are associated with heavy precipitation (i.e., above 60 mm h−1), whereas, after the data augmentation in amplitude (step 3), around 55% of patches with BEs are associated with heavy precipitation. In this case, the pseudoreflectivity magnitude is relied on too heavily to detect the BEs in the pseudoreflectivity fields. Patches with large magnitude pseudoreflectivities but without a BE are forced in the training database to solve this problem and the rate of large magnitude pseudoreflectivity patches without a BE in the total number of patches without a BE is defined (heavy_rate). A patch is considered with large magnitude pseudoreflectivities if the maximum is above 60 mm h−1. Another way to address this problem could be to add more input predictors.
Source:
Mounier, A., L. Raynaud, L. Rottner, M. Plu, P. Arbogast, M. Kreitz, L. Mignan, and B. Touzé, 2022: Detection of Bow Echoes in Kilometer-Scale Forecasts Using a Convolutional Neural Network. Artif. Intell. Earth Syst., 1, e210010, https://doi.org/10.1175/AIES-D-21-0010.1.
The researchers have made the source code of their application available.
Figure 1 - U-Net architecture for BE detection. Citation: Artificial Intelligence for the Earth Systems 1, 2;