Overview

Problem Setting

We tackle the challenging task of predicting the class of the largest solar flare within a 24-hour horizon using full-disk multi-wavelength solar images. This is formulated as a multi-class classification problem with significant real-world implications for space weather forecasting.

Flare Class
Peak X-ray Flux (I) [W/m2]
X
I > 10-4
M
10-5 < I ≤ 10-4
C
10-6 < I ≤ 10-5
O
I ≤ 10-6

Correspondence between flare classes and peak X-ray flux intensities.

Model Architecture

We propose Deep SWM, a novel architecture extending deep state-space models for classifying the maximum solar flare class within a 24-hour horizon, utilizing HMI and multi-wavelength AIA images.

Deep SWM Architecture

The novelties of our proposed method are as follows:

Solar Spatial Encoder (SSE)

Comprising the Depth-wise Channel Selective Module (DCSM) and the Spatio-Temporal State-Space Module (ST-SSM). The DCSM selectively weights multi-wavelength image channels to emphasize features relevant to solar events, while the ST-SSM efficiently captures long-range spatio-temporal dependencies in the solar images.

Long-range Temporal SSM (LT-SSM)

Extends deep state-space models to effectively model temporal dependencies exceeding the solar rotation period within the intermediate features obtained from the pretraining stage. This allows the LT-SSM to efficiently capture long-range relationships that are crucial for solar flare prediction.

Sparse MAE

A pretraining strategy tailored for solar images that extends the Masked Autoencoder (MAE). Sparse MAE addresses the challenge of sparse, yet crucial, information regions in solar images (e.g., sunspots) using a novel two-phase masking approach. This ensures that these crucial regions are less likely to be completely masked during pretraining, leading to improved intermediate feature representations.

Sparse MAE Pretraining
(a) An example of an original image. (b) Visualization of patches with the top α% highest standard deviation highlighted, often corresponding to sunspot regions. (c) Illustration of the spatial-level masking. (d) Reconstructed image by our proposed Sparse MAE.

Quantitative Results

Our method outperforms all baseline approaches across all metrics and even surpasses human expert performance, demonstrating the effectiveness of our approach for solar flare prediction.

Table 1: Comparison of our method with state-of-the-art approaches and human experts. Higher values are better for all metrics.

Method
Test period
GMGS↑
BSS≥M↑
TSS≥M↑
Flare Transformer w/o PF [35]
2014-2017 (4 years)
0.220±0.116
-1.770±0.225
0.198±0.371
DeFN-R [54]
2014-2015 (2 years)
0.302±0.055
0.036±0.982
0.279±0.162
CNN-LSTM
2019-12-01 – 2022-11-30 (3 years)
0.315±0.166
0.272±0.259
0.330±0.306
DeFN [53]
2014-2015 (2 years)
0.375±0.141
0.022±0.782
0.413±0.150
Flare Transformer [35]
2014-2017 (4 years)
0.503±0.059
0.082±0.974
0.530±0.112
Ours
2019-12-01 – 2022-11-30 (3 years)
0.582±0.032
0.334±0.299
0.543±0.074
Human experts [38, 51]
2000-2015 (16 years)
0.48
0.16
0.50

Qualitative Results

Qualitative results for flare predictions
Qualitative results for flare predictions. (a) and (b) illustrate successful X-class predictions, (c) illustrates a successful M-class prediction, and (d) illustrates a failed X-class prediction. For each case, we present Extreme Ultraviolet images at 131 Å, 193 Å, and 304 Å from time t-k to t, the ground truth (GT) label, and the predictions from the baseline and our proposed method.
Reconstruction results
Reconstruction results obtained from the baseline MAE (ρ=0.5) and our proposed Sparse MAE. Rows (a), (b), (c), (d), (e), and (f) show 94 Å, 171 Å, 304 Å, 1600 Å, 4500 Å AIA, and HMI images, respectively, captured three hours before an upcoming X-class flare. Columns (i), (ii), (iii), (iv), and (v) present the original image, the baseline reconstruction, a visualization of patches with the top α% highest standard deviation highlighted, the spatial-level masking of the Sparse MAE, and the reconstruction of the Sparse MAE, respectively.