Bypassing the static input size of neural networks in flare forecasting by using spatial pyramid pooling

The spatial extension of active regions (ARs) of the Sun can vary from one case to the next. This is a problem when studying solar flares with Convolutional Neural Networks (CNNs) as they generally use input images of a fixed size. Different processes can be performed to retrieve a database with homogeneous-sized data, such as resizing. Unfortunately, key features can be lost or distorted during these processes. This can lead to a deterioration of the ability of CNNs to classify flares of different soft X-ray classes, especially those from ARs with complex structures. Our work aims to implement and test a CNN architecture that retains the full features of the original resolution of the input images. We compare the performance of two CNN architectures for solar flare prediction: the first is a traditional CNN with resized input whereas the other implements a spatial pyramid pooling (SPP) layer without any input resizing. Both are trained on the Spaceweather HMI Active Region Patch line-of-sight magnetogram database. We also study two cases of binary classification. In the first case, our model distinguishes ARs producing flares in less than 24h of class greater or equal to C1.0 from ARs producing flares in more than 24h or never; in the second case, it distinguishes ARs producing flares in less than 24h of class greater or equal to M1.0 from the other ARs. Our models implementing an SPP layer outperform the traditional CNN models when predicting flares greater or equal to C1.0 within 24h. However, their performances degrade sharply along the other models studied in this paper, when trained to classify images greater or equal to M1.0 flares. The degradation in SPP models when classifying only images greater or equal to M1.0 flares as positive may be attributed to its success in identifying features that appear in ARs a few hours before the flare, independently of their soft X-ray class.