Speaker
Description
The spatial extension of active regions of the Sun and their associated images can strongly vary from one case to the next. This inhomogeneity is problematic when studying solar flares with convolutionnal neural networks (CNNs) due to their fixed input size. Several processes can be performed to produce a database with homogeneous-sized data, such as coarse resizing, cropping, or padding of raw images. Unfortunately, key features can be lost or distorted beyond recognition during these processes. This can lead to a deterioration of the ability of CNNs to predict flares of different soft X-ray classes, especially those from active regions with structures of great complexity.
This study aims to implement and test a CNN architecture that retains features of characteristic scales as fine as the original resolution of the input images. To do this, we compare the performance of two convolutional neural network models for solar flare prediction: the first one is a traditional CNN with convolution layers, batch normalization layers, max-pooling layers, and resized input whereas the other implements a spatial pyramid pooling (SPP) layer instead of a max pooling layer before the flatten-layer and without any input resizing. The models are trained on the SHARP Line-of-sight magnetogram database from 2010-05 to 2021-08 and using only images within 45◦ of the central meridian of the Sun. We also study two cases of binary classification: in the first case, our model has to distinguish active regions producing flares in less than 24h of class ≥C1.0 from active regions producing flares in more than 24h or never; in the second case, it has to distinguish active regions producing flares in less than 24h of class ≥M1.0 from active regions producing flares in more than 24h or never, or flares in less than 24h but of class lower than M1.0.
Our model implementing an SPP layer predicts flares ≥C1.0 within 24 hours more accurately than the traditional CNN model with a better TSS and PR AUC. However, its performances degrade sharply when the images of active regions producing a C-class flare are classified as negative. This may be attributed to its success in identifying features that appear in active regions a few hours before the flare, independently of their soft X-ray class. Furthermore, the results of this study may be the first lead to the importance of image size and ratio for flare forecasting using deep-learning methods. Further studies on the impact of image size and ratio may uncover important features for flare-triggering mechanisms.
| Do you plan to attend in-person or online? | In-person |
|---|
